diff --git a/croco/LICENSE b/croco/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..d9b84b1a65f9db6d8920a9048d162f52ba3ea56d --- /dev/null +++ b/croco/LICENSE @@ -0,0 +1,52 @@ +CroCo, Copyright (c) 2022-present Naver Corporation, is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license. + +A summary of the CC BY-NC-SA 4.0 license is located here: + https://creativecommons.org/licenses/by-nc-sa/4.0/ + +The CC BY-NC-SA 4.0 license is located here: + https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode + + +SEE NOTICE BELOW WITH RESPECT TO THE FILE: models/pos_embed.py, models/blocks.py + +*************************** + +NOTICE WITH RESPECT TO THE FILE: models/pos_embed.py + +This software is being redistributed in a modifiled form. The original form is available here: + +https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py + +This software in this file incorporates parts of the following software available here: + +Transformer: https://github.com/tensorflow/models/blob/master/official/legacy/transformer/model_utils.py +available under the following license: https://github.com/tensorflow/models/blob/master/LICENSE + +MoCo v3: https://github.com/facebookresearch/moco-v3 +available under the following license: https://github.com/facebookresearch/moco-v3/blob/main/LICENSE + +DeiT: https://github.com/facebookresearch/deit +available under the following license: https://github.com/facebookresearch/deit/blob/main/LICENSE + + +ORIGINAL COPYRIGHT NOTICE AND PERMISSION NOTICE AVAILABLE HERE IS REPRODUCE BELOW: + +https://github.com/facebookresearch/mae/blob/main/LICENSE + +Attribution-NonCommercial 4.0 International + +*************************** + +NOTICE WITH RESPECT TO THE FILE: models/blocks.py + +This software is being redistributed in a modifiled form. The original form is available here: + +https://github.com/rwightman/pytorch-image-models + +ORIGINAL COPYRIGHT NOTICE AND PERMISSION NOTICE AVAILABLE HERE IS REPRODUCE BELOW: + +https://github.com/rwightman/pytorch-image-models/blob/master/LICENSE + +Apache License +Version 2.0, January 2004 +http://www.apache.org/licenses/ \ No newline at end of file diff --git a/croco/NOTICE b/croco/NOTICE new file mode 100644 index 0000000000000000000000000000000000000000..d51bb365036c12d428d6e3a4fd00885756d5261c --- /dev/null +++ b/croco/NOTICE @@ -0,0 +1,21 @@ +CroCo +Copyright 2022-present NAVER Corp. + +This project contains subcomponents with separate copyright notices and license terms. +Your use of the source code for these subcomponents is subject to the terms and conditions of the following licenses. + +==== + +facebookresearch/mae +https://github.com/facebookresearch/mae + +Attribution-NonCommercial 4.0 International + +==== + +rwightman/pytorch-image-models +https://github.com/rwightman/pytorch-image-models + +Apache License +Version 2.0, January 2004 +http://www.apache.org/licenses/ \ No newline at end of file diff --git a/croco/README.MD b/croco/README.MD new file mode 100644 index 0000000000000000000000000000000000000000..38e33b001a60bd16749317fb297acd60f28a6f1b --- /dev/null +++ b/croco/README.MD @@ -0,0 +1,124 @@ +# CroCo + CroCo v2 / CroCo-Stereo / CroCo-Flow + +[[`CroCo arXiv`](https://arxiv.org/abs/2210.10716)] [[`CroCo v2 arXiv`](https://arxiv.org/abs/2211.10408)] [[`project page and demo`](https://croco.europe.naverlabs.com/)] + +This repository contains the code for our CroCo model presented in our NeurIPS'22 paper [CroCo: Self-Supervised Pre-training for 3D Vision Tasks by Cross-View Completion](https://openreview.net/pdf?id=wZEfHUM5ri) and its follow-up extension published at ICCV'23 [Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow](https://openaccess.thecvf.com/content/ICCV2023/html/Weinzaepfel_CroCo_v2_Improved_Cross-view_Completion_Pre-training_for_Stereo_Matching_and_ICCV_2023_paper.html), refered to as CroCo v2: + +![image](assets/arch.jpg) + +```bibtex +@inproceedings{croco, + title={{CroCo: Self-Supervised Pre-training for 3D Vision Tasks by Cross-View Completion}}, + author={{Weinzaepfel, Philippe and Leroy, Vincent and Lucas, Thomas and Br\'egier, Romain and Cabon, Yohann and Arora, Vaibhav and Antsfeld, Leonid and Chidlovskii, Boris and Csurka, Gabriela and Revaud J\'er\^ome}}, + booktitle={{NeurIPS}}, + year={2022} +} + +@inproceedings{croco_v2, + title={{CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow}}, + author={Weinzaepfel, Philippe and Lucas, Thomas and Leroy, Vincent and Cabon, Yohann and Arora, Vaibhav and Br{\'e}gier, Romain and Csurka, Gabriela and Antsfeld, Leonid and Chidlovskii, Boris and Revaud, J{\'e}r{\^o}me}, + booktitle={ICCV}, + year={2023} +} +``` + +## License + +The code is distributed under the CC BY-NC-SA 4.0 License. See [LICENSE](LICENSE) for more information. +Some components are based on code from [MAE](https://github.com/facebookresearch/mae) released under the CC BY-NC-SA 4.0 License and [timm](https://github.com/rwightman/pytorch-image-models) released under the Apache 2.0 License. +Some components for stereo matching and optical flow are based on code from [unimatch](https://github.com/autonomousvision/unimatch) released under the MIT license. + +## Preparation + +1. Install dependencies on a machine with a NVidia GPU using e.g. conda. Note that `habitat-sim` is required only for the interactive demo and the synthetic pre-training data generation. If you don't plan to use it, you can ignore the line installing it and use a more recent python version. + +```bash +conda create -n croco python=3.7 cmake=3.14.0 +conda activate croco +conda install habitat-sim headless -c conda-forge -c aihabitat +conda install pytorch torchvision -c pytorch +conda install notebook ipykernel matplotlib +conda install ipywidgets widgetsnbextension +conda install scikit-learn tqdm quaternion opencv # only for pretraining / habitat data generation + +``` + +2. Compile cuda kernels for RoPE + +CroCo v2 relies on RoPE positional embeddings for which you need to compile some cuda kernels. +```bash +cd models/curope/ +python setup.py build_ext --inplace +cd ../../ +``` + +This can be a bit long as we compile for all cuda architectures, feel free to update L9 of `models/curope/setup.py` to compile for specific architectures only. +You might also need to set the environment `CUDA_HOME` in case you use a custom cuda installation. + +In case you cannot provide, we also provide a slow pytorch version, which will be automatically loaded. + +3. Download pre-trained model + +We provide several pre-trained models: + +| modelname | pre-training data | pos. embed. | Encoder | Decoder | +|------------------------------------------------------------------------------------------------------------------------------------|-------------------|-------------|---------|---------| +| [`CroCo.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo.pth) | Habitat | cosine | ViT-B | Small | +| [`CroCo_V2_ViTBase_SmallDecoder.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo_V2_ViTBase_SmallDecoder.pth) | Habitat + real | RoPE | ViT-B | Small | +| [`CroCo_V2_ViTBase_BaseDecoder.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo_V2_ViTBase_BaseDecoder.pth) | Habitat + real | RoPE | ViT-B | Base | +| [`CroCo_V2_ViTLarge_BaseDecoder.pth`](https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo_V2_ViTLarge_BaseDecoder.pth) | Habitat + real | RoPE | ViT-L | Base | + +To download a specific model, i.e., the first one (`CroCo.pth`) +```bash +mkdir -p pretrained_models/ +wget https://download.europe.naverlabs.com/ComputerVision/CroCo/CroCo.pth -P pretrained_models/ +``` + +## Reconstruction example + +Simply run after downloading the `CroCo_V2_ViTLarge_BaseDecoder` pretrained model (or update the corresponding line in `demo.py`) +```bash +python demo.py +``` + +## Interactive demonstration of cross-view completion reconstruction on the Habitat simulator + +First download the test scene from Habitat: +```bash +python -m habitat_sim.utils.datasets_download --uids habitat_test_scenes --data-path habitat-sim-data/ +``` + +Then, run the Notebook demo `interactive_demo.ipynb`. + +In this demo, you should be able to sample a random reference viewpoint from an [Habitat](https://github.com/facebookresearch/habitat-sim) test scene. Use the sliders to change viewpoint and select a masked target view to reconstruct using CroCo. +![croco_interactive_demo](https://user-images.githubusercontent.com/1822210/200516576-7937bc6a-55f8-49ed-8618-3ddf89433ea4.jpg) + +## Pre-training + +### CroCo + +To pre-train CroCo, please first generate the pre-training data from the Habitat simulator, following the instructions in [datasets/habitat_sim/README.MD](datasets/habitat_sim/README.MD) and then run the following command: +``` +torchrun --nproc_per_node=4 pretrain.py --output_dir ./output/pretraining/ +``` + +Our CroCo pre-training was launched on a single server with 4 GPUs. +It should take around 10 days with A100 or 15 days with V100 to do the 400 pre-training epochs, but decent performances are obtained earlier in training. +Note that, while the code contains the same scaling rule of the learning rate as MAE when changing the effective batch size, we did not experimented if it is valid in our case. +The first run can take a few minutes to start, to parse all available pre-training pairs. + +### CroCo v2 + +For CroCo v2 pre-training, in addition to the generation of the pre-training data from the Habitat simulator above, please pre-extract the crops from the real datasets following the instructions in [datasets/crops/README.MD](datasets/crops/README.MD). +Then, run the following command for the largest model (ViT-L encoder, Base decoder): +``` +torchrun --nproc_per_node=8 pretrain.py --model "CroCoNet(enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_num_heads=12, dec_depth=12, pos_embed='RoPE100')" --dataset "habitat_release+ARKitScenes+MegaDepth+3DStreetView+IndoorVL" --warmup_epochs 12 --max_epoch 125 --epochs 250 --amp 0 --keep_freq 5 --output_dir ./output/pretraining_crocov2/ +``` + +Our CroCo v2 pre-training was launched on a single server with 8 GPUs for the largest model, and on a single server with 4 GPUs for the smaller ones, keeping a batch size of 64 per gpu in all cases. +The largest model should take around 12 days on A100. +Note that, while the code contains the same scaling rule of the learning rate as MAE when changing the effective batch size, we did not experimented if it is valid in our case. + +## Stereo matching and Optical flow downstream tasks + +For CroCo-Stereo and CroCo-Flow, please refer to [stereoflow/README.MD](stereoflow/README.MD). diff --git a/croco/assets/arch.jpg b/croco/assets/arch.jpg new file mode 100644 index 0000000000000000000000000000000000000000..3f5b032729ddc58c06d890a0ebda1749276070c4 Binary files /dev/null and b/croco/assets/arch.jpg differ diff --git a/croco/croco-stereo-flow-demo.ipynb b/croco/croco-stereo-flow-demo.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..2b00a7607ab5f82d1857041969bfec977e56b3e0 --- /dev/null +++ b/croco/croco-stereo-flow-demo.ipynb @@ -0,0 +1,191 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "id": "9bca0f41", + "metadata": {}, + "source": [ + "# Simple inference example with CroCo-Stereo or CroCo-Flow" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "80653ef7", + "metadata": {}, + "outputs": [], + "source": [ + "# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n", + "# Licensed under CC BY-NC-SA 4.0 (non-commercial use only)." + ] + }, + { + "cell_type": "markdown", + "id": "4f033862", + "metadata": {}, + "source": [ + "First download the model(s) of your choice by running\n", + "```\n", + "bash stereoflow/download_model.sh crocostereo.pth\n", + "bash stereoflow/download_model.sh crocoflow.pth\n", + "```" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "1fb2e392", + "metadata": {}, + "outputs": [], + "source": [ + "import torch\n", + "use_gpu = torch.cuda.is_available() and torch.cuda.device_count()>0\n", + "device = torch.device('cuda:0' if use_gpu else 'cpu')\n", + "import matplotlib.pylab as plt" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "e0e25d77", + "metadata": {}, + "outputs": [], + "source": [ + "from stereoflow.test import _load_model_and_criterion\n", + "from stereoflow.engine import tiled_pred\n", + "from stereoflow.datasets_stereo import img_to_tensor, vis_disparity\n", + "from stereoflow.datasets_flow import flowToColor\n", + "tile_overlap=0.7 # recommended value, higher value can be slightly better but slower" + ] + }, + { + "cell_type": "markdown", + "id": "86a921f5", + "metadata": {}, + "source": [ + "### CroCo-Stereo example" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "64e483cb", + "metadata": {}, + "outputs": [], + "source": [ + "image1 = np.asarray(Image.open(''))\n", + "image2 = np.asarray(Image.open(''))" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "f0d04303", + "metadata": {}, + "outputs": [], + "source": [ + "model, _, cropsize, with_conf, task, tile_conf_mode = _load_model_and_criterion('stereoflow_models/crocostereo.pth', None, device)\n" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "47dc14b5", + "metadata": {}, + "outputs": [], + "source": [ + "im1 = img_to_tensor(image1).to(device).unsqueeze(0)\n", + "im2 = img_to_tensor(image2).to(device).unsqueeze(0)\n", + "with torch.inference_mode():\n", + " pred, _, _ = tiled_pred(model, None, im1, im2, None, conf_mode=tile_conf_mode, overlap=tile_overlap, crop=cropsize, with_conf=with_conf, return_time=False)\n", + "pred = pred.squeeze(0).squeeze(0).cpu().numpy()" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "583b9f16", + "metadata": {}, + "outputs": [], + "source": [ + "plt.imshow(vis_disparity(pred))\n", + "plt.axis('off')" + ] + }, + { + "cell_type": "markdown", + "id": "d2df5d70", + "metadata": {}, + "source": [ + "### CroCo-Flow example" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "9ee257a7", + "metadata": {}, + "outputs": [], + "source": [ + "image1 = np.asarray(Image.open(''))\n", + "image2 = np.asarray(Image.open(''))" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "d5edccf0", + "metadata": {}, + "outputs": [], + "source": [ + "model, _, cropsize, with_conf, task, tile_conf_mode = _load_model_and_criterion('stereoflow_models/crocoflow.pth', None, device)\n" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "b19692c3", + "metadata": {}, + "outputs": [], + "source": [ + "im1 = img_to_tensor(image1).to(device).unsqueeze(0)\n", + "im2 = img_to_tensor(image2).to(device).unsqueeze(0)\n", + "with torch.inference_mode():\n", + " pred, _, _ = tiled_pred(model, None, im1, im2, None, conf_mode=tile_conf_mode, overlap=tile_overlap, crop=cropsize, with_conf=with_conf, return_time=False)\n", + "pred = pred.squeeze(0).permute(1,2,0).cpu().numpy()" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "26f79db3", + "metadata": {}, + "outputs": [], + "source": [ + "plt.imshow(flowToColor(pred))\n", + "plt.axis('off')" + ] + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 3 (ipykernel)", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.9.7" + } + }, + "nbformat": 4, + "nbformat_minor": 5 +} diff --git a/croco/datasets/__init__.py b/croco/datasets/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/croco/datasets/crops/README.MD b/croco/datasets/crops/README.MD new file mode 100644 index 0000000000000000000000000000000000000000..47ddabebb177644694ee247ae878173a3a16644f --- /dev/null +++ b/croco/datasets/crops/README.MD @@ -0,0 +1,104 @@ +## Generation of crops from the real datasets + +The instructions below allow to generate the crops used for pre-training CroCo v2 from the following real-world datasets: ARKitScenes, MegaDepth, 3DStreetView and IndoorVL. + +### Download the metadata of the crops to generate + +First, download the metadata and put them in `./data/`: +``` +mkdir -p data +cd data/ +wget https://download.europe.naverlabs.com/ComputerVision/CroCo/data/crop_metadata.zip +unzip crop_metadata.zip +rm crop_metadata.zip +cd .. +``` + +### Prepare the original datasets + +Second, download the original datasets in `./data/original_datasets/`. +``` +mkdir -p data/original_datasets +``` + +##### ARKitScenes + +Download the `raw` dataset from https://github.com/apple/ARKitScenes/blob/main/DATA.md and put it in `./data/original_datasets/ARKitScenes/`. +The resulting file structure should be like: +``` +./data/original_datasets/ARKitScenes/ +└───Training + └───40753679 + │ │ ultrawide + │ │ ... + └───40753686 + │ + ... +``` + +##### MegaDepth + +Download `MegaDepth v1 Dataset` from https://www.cs.cornell.edu/projects/megadepth/ and put it in `./data/original_datasets/MegaDepth/`. +The resulting file structure should be like: + +``` +./data/original_datasets/MegaDepth/ +└───0000 +│ └───images +│ │ │ 1000557903_87fa96b8a4_o.jpg +│ │ └ ... +│ └─── ... +└───0001 +│ │ +│ └ ... +└─── ... +``` + +##### 3DStreetView + +Download `3D_Street_View` dataset from https://github.com/amir32002/3D_Street_View and put it in `./data/original_datasets/3DStreetView/`. +The resulting file structure should be like: + +``` +./data/original_datasets/3DStreetView/ +└───dataset_aligned +│ └───0002 +│ │ │ 0000002_0000001_0000002_0000001.jpg +│ │ └ ... +│ └─── ... +└───dataset_unaligned +│ └───0003 +│ │ │ 0000003_0000001_0000002_0000001.jpg +│ │ └ ... +│ └─── ... +``` + +##### IndoorVL + +Download the `IndoorVL` datasets using [Kapture](https://github.com/naver/kapture). + +``` +pip install kapture +mkdir -p ./data/original_datasets/IndoorVL +cd ./data/original_datasets/IndoorVL +kapture_download_dataset.py update +kapture_download_dataset.py install "HyundaiDepartmentStore_*" +kapture_download_dataset.py install "GangnamStation_*" +cd - +``` + +### Extract the crops + +Now, extract the crops for each of the dataset: +``` +for dataset in ARKitScenes MegaDepth 3DStreetView IndoorVL; +do + python3 datasets/crops/extract_crops_from_images.py --crops ./data/crop_metadata/${dataset}/crops_release.txt --root-dir ./data/original_datasets/${dataset}/ --output-dir ./data/${dataset}_crops/ --imsize 256 --nthread 8 --max-subdir-levels 5 --ideal-number-pairs-in-dir 500; +done +``` + +##### Note for IndoorVL + +Due to some legal issues, we can only release 144,228 pairs out of the 1,593,689 pairs used in the paper. +To account for it in terms of number of pre-training iterations, the pre-training command in this repository uses 125 training epochs including 12 warm-up epochs and learning rate cosine schedule of 250, instead of 100, 10 and 200 respectively. +The impact on the performance is negligible. diff --git a/croco/datasets/crops/extract_crops_from_images.py b/croco/datasets/crops/extract_crops_from_images.py new file mode 100644 index 0000000000000000000000000000000000000000..870cf9f9690bfc53f10a59293aabc16da127b02e --- /dev/null +++ b/croco/datasets/crops/extract_crops_from_images.py @@ -0,0 +1,183 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# Extracting crops for pre-training +# -------------------------------------------------------- + +import os +import argparse +from tqdm import tqdm +from PIL import Image +import functools +from multiprocessing import Pool +import math + + +def arg_parser(): + parser = argparse.ArgumentParser( + "Generate cropped image pairs from image crop list" + ) + + parser.add_argument("--crops", type=str, required=True, help="crop file") + parser.add_argument("--root-dir", type=str, required=True, help="root directory") + parser.add_argument( + "--output-dir", type=str, required=True, help="output directory" + ) + parser.add_argument("--imsize", type=int, default=256, help="size of the crops") + parser.add_argument( + "--nthread", type=int, required=True, help="number of simultaneous threads" + ) + parser.add_argument( + "--max-subdir-levels", + type=int, + default=5, + help="maximum number of subdirectories", + ) + parser.add_argument( + "--ideal-number-pairs-in-dir", + type=int, + default=500, + help="number of pairs stored in a dir", + ) + return parser + + +def main(args): + listing_path = os.path.join(args.output_dir, "listing.txt") + + print(f"Loading list of crops ... ({args.nthread} threads)") + crops, num_crops_to_generate = load_crop_file(args.crops) + + print(f"Preparing jobs ({len(crops)} candidate image pairs)...") + num_levels = min( + math.ceil(math.log(num_crops_to_generate, args.ideal_number_pairs_in_dir)), + args.max_subdir_levels, + ) + num_pairs_in_dir = math.ceil(num_crops_to_generate ** (1 / num_levels)) + + jobs = prepare_jobs(crops, num_levels, num_pairs_in_dir) + del crops + + os.makedirs(args.output_dir, exist_ok=True) + mmap = Pool(args.nthread).imap_unordered if args.nthread > 1 else map + call = functools.partial(save_image_crops, args) + + print(f"Generating cropped images to {args.output_dir} ...") + with open(listing_path, "w") as listing: + listing.write("# pair_path\n") + for results in tqdm(mmap(call, jobs), total=len(jobs)): + for path in results: + listing.write(f"{path}\n") + print("Finished writing listing to", listing_path) + + +def load_crop_file(path): + data = open(path).read().splitlines() + pairs = [] + num_crops_to_generate = 0 + for line in tqdm(data): + if line.startswith("#"): + continue + line = line.split(", ") + if len(line) < 8: + img1, img2, rotation = line + pairs.append((img1, img2, int(rotation), [])) + else: + l1, r1, t1, b1, l2, r2, t2, b2 = map(int, line) + rect1, rect2 = (l1, t1, r1, b1), (l2, t2, r2, b2) + pairs[-1][-1].append((rect1, rect2)) + num_crops_to_generate += 1 + return pairs, num_crops_to_generate + + +def prepare_jobs(pairs, num_levels, num_pairs_in_dir): + jobs = [] + powers = [num_pairs_in_dir**level for level in reversed(range(num_levels))] + + def get_path(idx): + idx_array = [] + d = idx + for level in range(num_levels - 1): + idx_array.append(idx // powers[level]) + idx = idx % powers[level] + idx_array.append(d) + return "/".join(map(lambda x: hex(x)[2:], idx_array)) + + idx = 0 + for pair_data in tqdm(pairs): + img1, img2, rotation, crops = pair_data + if -60 <= rotation and rotation <= 60: + rotation = 0 # most likely not a true rotation + paths = [get_path(idx + k) for k in range(len(crops))] + idx += len(crops) + jobs.append(((img1, img2), rotation, crops, paths)) + return jobs + + +def load_image(path): + try: + return Image.open(path).convert("RGB") + except Exception as e: + print("skipping", path, e) + raise OSError() + + +def save_image_crops(args, data): + # load images + img_pair, rot, crops, paths = data + try: + img1, img2 = [ + load_image(os.path.join(args.root_dir, impath)) for impath in img_pair + ] + except OSError as e: + return [] + + def area(sz): + return sz[0] * sz[1] + + tgt_size = (args.imsize, args.imsize) + + def prepare_crop(img, rect, rot=0): + # actual crop + img = img.crop(rect) + + # resize to desired size + interp = ( + Image.Resampling.LANCZOS + if area(img.size) > 4 * area(tgt_size) + else Image.Resampling.BICUBIC + ) + img = img.resize(tgt_size, resample=interp) + + # rotate the image + rot90 = (round(rot / 90) % 4) * 90 + if rot90 == 90: + img = img.transpose(Image.Transpose.ROTATE_90) + elif rot90 == 180: + img = img.transpose(Image.Transpose.ROTATE_180) + elif rot90 == 270: + img = img.transpose(Image.Transpose.ROTATE_270) + return img + + results = [] + for (rect1, rect2), path in zip(crops, paths): + crop1 = prepare_crop(img1, rect1) + crop2 = prepare_crop(img2, rect2, rot) + + fullpath1 = os.path.join(args.output_dir, path + "_1.jpg") + fullpath2 = os.path.join(args.output_dir, path + "_2.jpg") + os.makedirs(os.path.dirname(fullpath1), exist_ok=True) + + assert not os.path.isfile(fullpath1), fullpath1 + assert not os.path.isfile(fullpath2), fullpath2 + crop1.save(fullpath1) + crop2.save(fullpath2) + results.append(path) + + return results + + +if __name__ == "__main__": + args = arg_parser().parse_args() + main(args) diff --git a/croco/datasets/habitat_sim/README.MD b/croco/datasets/habitat_sim/README.MD new file mode 100644 index 0000000000000000000000000000000000000000..a505781ff9eb91bce7f1d189e848f8ba1c560940 --- /dev/null +++ b/croco/datasets/habitat_sim/README.MD @@ -0,0 +1,76 @@ +## Generation of synthetic image pairs using Habitat-Sim + +These instructions allow to generate pre-training pairs from the Habitat simulator. +As we did not save metadata of the pairs used in the original paper, they are not strictly the same, but these data use the same setting and are equivalent. + +### Download Habitat-Sim scenes +Download Habitat-Sim scenes: +- Download links can be found here: https://github.com/facebookresearch/habitat-sim/blob/main/DATASETS.md +- We used scenes from the HM3D, habitat-test-scenes, Replica, ReplicaCad and ScanNet datasets. +- Please put the scenes under `./data/habitat-sim-data/scene_datasets/` following the structure below, or update manually paths in `paths.py`. +``` +./data/ +└──habitat-sim-data/ + └──scene_datasets/ + ├──hm3d/ + ├──gibson/ + ├──habitat-test-scenes/ + ├──replica_cad_baked_lighting/ + ├──replica_cad/ + ├──ReplicaDataset/ + └──scannet/ +``` + +### Image pairs generation +We provide metadata to generate reproducible images pairs for pretraining and validation. +Experiments described in the paper used similar data, but whose generation was not reproducible at the time. + +Specifications: +- 256x256 resolution images, with 60 degrees field of view . +- Up to 1000 image pairs per scene. +- Number of scenes considered/number of images pairs per dataset: + - Scannet: 1097 scenes / 985 209 pairs + - HM3D: + - hm3d/train: 800 / 800k pairs + - hm3d/val: 100 scenes / 100k pairs + - hm3d/minival: 10 scenes / 10k pairs + - habitat-test-scenes: 3 scenes / 3k pairs + - replica_cad_baked_lighting: 13 scenes / 13k pairs + +- Scenes from hm3d/val and hm3d/minival pairs were not used for the pre-training but kept for validation purposes. + +Download metadata and extract it: +```bash +mkdir -p data/habitat_release_metadata/ +cd data/habitat_release_metadata/ +wget https://download.europe.naverlabs.com/ComputerVision/CroCo/data/habitat_release_metadata/multiview_habitat_metadata.tar.gz +tar -xvf multiview_habitat_metadata.tar.gz +cd ../.. +# Location of the metadata +METADATA_DIR="./data/habitat_release_metadata/multiview_habitat_metadata" +``` + +Generate image pairs from metadata: +- The following command will print a list of commandlines to generate image pairs for each scene: +```bash +# Target output directory +PAIRS_DATASET_DIR="./data/habitat_release/" +python datasets/habitat_sim/generate_from_metadata_files.py --input_dir=$METADATA_DIR --output_dir=$PAIRS_DATASET_DIR +``` +- One can launch multiple of such commands in parallel e.g. using GNU Parallel: +```bash +python datasets/habitat_sim/generate_from_metadata_files.py --input_dir=$METADATA_DIR --output_dir=$PAIRS_DATASET_DIR | parallel -j 16 +``` + +## Metadata generation + +Image pairs were randomly sampled using the following commands, whose outputs contain randomness and are thus not exactly reproducible: +```bash +# Print commandlines to generate image pairs from the different scenes available. +PAIRS_DATASET_DIR=MY_CUSTOM_PATH +python datasets/habitat_sim/generate_multiview_images.py --list_commands --output_dir=$PAIRS_DATASET_DIR + +# Once a dataset is generated, pack metadata files for reproducibility. +METADATA_DIR=MY_CUSTON_PATH +python datasets/habitat_sim/pack_metadata_files.py $PAIRS_DATASET_DIR $METADATA_DIR +``` diff --git a/croco/datasets/habitat_sim/__init__.py b/croco/datasets/habitat_sim/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/croco/datasets/habitat_sim/generate_from_metadata.py b/croco/datasets/habitat_sim/generate_from_metadata.py new file mode 100644 index 0000000000000000000000000000000000000000..6bbfbc6bec23e182baed2c4eedf0535fbc6aaa97 --- /dev/null +++ b/croco/datasets/habitat_sim/generate_from_metadata.py @@ -0,0 +1,125 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +""" +Script to generate image pairs for a given scene reproducing poses provided in a metadata file. +""" +import os +from datasets.habitat_sim.multiview_habitat_sim_generator import ( + MultiviewHabitatSimGenerator, +) +from datasets.habitat_sim.paths import SCENES_DATASET +import argparse +import quaternion +import PIL.Image +import cv2 +import json +from tqdm import tqdm + + +def generate_multiview_images_from_metadata( + metadata_filename, + output_dir, + overload_params=dict(), + scene_datasets_paths=None, + exist_ok=False, +): + """ + Generate images from a metadata file for reproducibility purposes. + """ + # Reorder paths by decreasing label length, to avoid collisions when testing if a string by such label + if scene_datasets_paths is not None: + scene_datasets_paths = dict( + sorted(scene_datasets_paths.items(), key=lambda x: len(x[0]), reverse=True) + ) + + with open(metadata_filename, "r") as f: + input_metadata = json.load(f) + metadata = dict() + for key, value in input_metadata.items(): + # Optionally replace some paths + if key in ("scene_dataset_config_file", "scene", "navmesh") and value != "": + if scene_datasets_paths is not None: + for dataset_label, dataset_path in scene_datasets_paths.items(): + if value.startswith(dataset_label): + value = os.path.normpath( + os.path.join( + dataset_path, os.path.relpath(value, dataset_label) + ) + ) + break + metadata[key] = value + + # Overload some parameters + for key, value in overload_params.items(): + metadata[key] = value + + generation_entries = dict( + [ + (key, value) + for key, value in metadata.items() + if not (key in ("multiviews", "output_dir", "generate_depth")) + ] + ) + generate_depth = metadata["generate_depth"] + + os.makedirs(output_dir, exist_ok=exist_ok) + + generator = MultiviewHabitatSimGenerator(**generation_entries) + + # Generate views + for idx_label, data in tqdm(metadata["multiviews"].items()): + positions = data["positions"] + orientations = data["orientations"] + n = len(positions) + for oidx in range(n): + observation = generator.render_viewpoint( + positions[oidx], quaternion.from_float_array(orientations[oidx]) + ) + observation_label = f"{oidx + 1}" # Leonid is indexing starting from 1 + # Color image saved using PIL + img = PIL.Image.fromarray(observation["color"][:, :, :3]) + filename = os.path.join(output_dir, f"{idx_label}_{observation_label}.jpeg") + img.save(filename) + if generate_depth: + # Depth image as EXR file + filename = os.path.join( + output_dir, f"{idx_label}_{observation_label}_depth.exr" + ) + cv2.imwrite( + filename, + observation["depth"], + [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF], + ) + # Camera parameters + camera_params = dict( + [ + (key, observation[key].tolist()) + for key in ("camera_intrinsics", "R_cam2world", "t_cam2world") + ] + ) + filename = os.path.join( + output_dir, f"{idx_label}_{observation_label}_camera_params.json" + ) + with open(filename, "w") as f: + json.dump(camera_params, f) + # Save metadata + with open(os.path.join(output_dir, "metadata.json"), "w") as f: + json.dump(metadata, f) + + generator.close() + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--metadata_filename", required=True) + parser.add_argument("--output_dir", required=True) + args = parser.parse_args() + + generate_multiview_images_from_metadata( + metadata_filename=args.metadata_filename, + output_dir=args.output_dir, + scene_datasets_paths=SCENES_DATASET, + overload_params=dict(), + exist_ok=True, + ) diff --git a/croco/datasets/habitat_sim/generate_from_metadata_files.py b/croco/datasets/habitat_sim/generate_from_metadata_files.py new file mode 100644 index 0000000000000000000000000000000000000000..2376957e0578726a98515220167e86fbecc2d72d --- /dev/null +++ b/croco/datasets/habitat_sim/generate_from_metadata_files.py @@ -0,0 +1,36 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +""" +Script generating commandlines to generate image pairs from metadata files. +""" +import os +import glob +from tqdm import tqdm +import argparse + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--input_dir", required=True) + parser.add_argument("--output_dir", required=True) + parser.add_argument( + "--prefix", + default="", + help="Commanline prefix, useful e.g. to setup environment.", + ) + args = parser.parse_args() + + input_metadata_filenames = glob.iglob( + f"{args.input_dir}/**/metadata.json", recursive=True + ) + + for metadata_filename in tqdm(input_metadata_filenames): + output_dir = os.path.join( + args.output_dir, + os.path.relpath(os.path.dirname(metadata_filename), args.input_dir), + ) + # Do not process the scene if the metadata file already exists + if os.path.exists(os.path.join(output_dir, "metadata.json")): + continue + commandline = f"{args.prefix}python datasets/habitat_sim/generate_from_metadata.py --metadata_filename={metadata_filename} --output_dir={output_dir}" + print(commandline) diff --git a/croco/datasets/habitat_sim/generate_multiview_images.py b/croco/datasets/habitat_sim/generate_multiview_images.py new file mode 100644 index 0000000000000000000000000000000000000000..cf16062135dfbaeb38ff2ad91c33bcab50cb98aa --- /dev/null +++ b/croco/datasets/habitat_sim/generate_multiview_images.py @@ -0,0 +1,231 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +import os +from tqdm import tqdm +import argparse +import PIL.Image +import numpy as np +import json +from datasets.habitat_sim.multiview_habitat_sim_generator import ( + MultiviewHabitatSimGenerator, + NoNaviguableSpaceError, +) +from datasets.habitat_sim.paths import list_scenes_available +import cv2 +import quaternion +import shutil + + +def generate_multiview_images_for_scene( + scene_dataset_config_file, + scene, + navmesh, + output_dir, + views_count, + size, + exist_ok=False, + generate_depth=False, + **kwargs, +): + """ + Generate tuples of overlapping views for a given scene. + generate_depth: generate depth images and camera parameters. + """ + if os.path.exists(output_dir) and not exist_ok: + print(f"Scene {scene}: data already generated. Ignoring generation.") + return + try: + print(f"Scene {scene}: {size} multiview acquisitions to generate...") + os.makedirs(output_dir, exist_ok=exist_ok) + + metadata_filename = os.path.join(output_dir, "metadata.json") + + metadata_template = dict( + scene_dataset_config_file=scene_dataset_config_file, + scene=scene, + navmesh=navmesh, + views_count=views_count, + size=size, + generate_depth=generate_depth, + **kwargs, + ) + metadata_template["multiviews"] = dict() + + if os.path.exists(metadata_filename): + print("Metadata file already exists:", metadata_filename) + print("Loading already generated metadata file...") + with open(metadata_filename, "r") as f: + metadata = json.load(f) + + for key in metadata_template.keys(): + if key != "multiviews": + assert ( + metadata_template[key] == metadata[key] + ), f"existing file is inconsistent with the input parameters:\nKey: {key}\nmetadata: {metadata[key]}\ntemplate: {metadata_template[key]}." + else: + print("No temporary file found. Starting generation from scratch...") + metadata = metadata_template + + starting_id = len(metadata["multiviews"]) + print(f"Starting generation from index {starting_id}/{size}...") + if starting_id >= size: + print("Generation already done.") + return + + generator = MultiviewHabitatSimGenerator( + scene_dataset_config_file=scene_dataset_config_file, + scene=scene, + navmesh=navmesh, + views_count=views_count, + size=size, + **kwargs, + ) + + for idx in tqdm(range(starting_id, size)): + # Generate / re-generate the observations + try: + data = generator[idx] + observations = data["observations"] + positions = data["positions"] + orientations = data["orientations"] + + idx_label = f"{idx:08}" + for oidx, observation in enumerate(observations): + observation_label = ( + f"{oidx + 1}" # Leonid is indexing starting from 1 + ) + # Color image saved using PIL + img = PIL.Image.fromarray(observation["color"][:, :, :3]) + filename = os.path.join( + output_dir, f"{idx_label}_{observation_label}.jpeg" + ) + img.save(filename) + if generate_depth: + # Depth image as EXR file + filename = os.path.join( + output_dir, f"{idx_label}_{observation_label}_depth.exr" + ) + cv2.imwrite( + filename, + observation["depth"], + [cv2.IMWRITE_EXR_TYPE, cv2.IMWRITE_EXR_TYPE_HALF], + ) + # Camera parameters + camera_params = dict( + [ + (key, observation[key].tolist()) + for key in ( + "camera_intrinsics", + "R_cam2world", + "t_cam2world", + ) + ] + ) + filename = os.path.join( + output_dir, + f"{idx_label}_{observation_label}_camera_params.json", + ) + with open(filename, "w") as f: + json.dump(camera_params, f) + metadata["multiviews"][idx_label] = { + "positions": positions.tolist(), + "orientations": orientations.tolist(), + "covisibility_ratios": data["covisibility_ratios"].tolist(), + "valid_fractions": data["valid_fractions"].tolist(), + "pairwise_visibility_ratios": data[ + "pairwise_visibility_ratios" + ].tolist(), + } + except RecursionError: + print( + "Recursion error: unable to sample observations for this scene. We will stop there." + ) + break + + # Regularly save a temporary metadata file, in case we need to restart the generation + if idx % 10 == 0: + with open(metadata_filename, "w") as f: + json.dump(metadata, f) + + # Save metadata + with open(metadata_filename, "w") as f: + json.dump(metadata, f) + + generator.close() + except NoNaviguableSpaceError: + pass + + +def create_commandline(scene_data, generate_depth, exist_ok=False): + """ + Create a commandline string to generate a scene. + """ + + def my_formatting(val): + if val is None or val == "": + return '""' + else: + return val + + commandline = f"""python {__file__} --scene {my_formatting(scene_data.scene)} + --scene_dataset_config_file {my_formatting(scene_data.scene_dataset_config_file)} + --navmesh {my_formatting(scene_data.navmesh)} + --output_dir {my_formatting(scene_data.output_dir)} + --generate_depth {int(generate_depth)} + --exist_ok {int(exist_ok)} + """ + commandline = " ".join(commandline.split()) + return commandline + + +if __name__ == "__main__": + os.umask(2) + + parser = argparse.ArgumentParser( + description="""Example of use -- listing commands to generate data for scenes available: + > python datasets/habitat_sim/generate_multiview_habitat_images.py --list_commands + """ + ) + + parser.add_argument("--output_dir", type=str, required=True) + parser.add_argument( + "--list_commands", action="store_true", help="list commandlines to run if true" + ) + parser.add_argument("--scene", type=str, default="") + parser.add_argument("--scene_dataset_config_file", type=str, default="") + parser.add_argument("--navmesh", type=str, default="") + + parser.add_argument("--generate_depth", type=int, default=1) + parser.add_argument("--exist_ok", type=int, default=0) + + kwargs = dict(resolution=(256, 256), hfov=60, views_count=2, size=1000) + + args = parser.parse_args() + generate_depth = bool(args.generate_depth) + exist_ok = bool(args.exist_ok) + + if args.list_commands: + # Listing scenes available... + scenes_data = list_scenes_available(base_output_dir=args.output_dir) + + for scene_data in scenes_data: + print( + create_commandline( + scene_data, generate_depth=generate_depth, exist_ok=exist_ok + ) + ) + else: + if args.scene == "" or args.output_dir == "": + print("Missing scene or output dir argument!") + print(parser.format_help()) + else: + generate_multiview_images_for_scene( + scene=args.scene, + scene_dataset_config_file=args.scene_dataset_config_file, + navmesh=args.navmesh, + output_dir=args.output_dir, + exist_ok=exist_ok, + generate_depth=generate_depth, + **kwargs, + ) diff --git a/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py b/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..b073407ec169be0674cbd33a1197731ec0dd3be3 --- /dev/null +++ b/croco/datasets/habitat_sim/multiview_habitat_sim_generator.py @@ -0,0 +1,501 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +import os +import numpy as np +import quaternion +import habitat_sim +import json +from sklearn.neighbors import NearestNeighbors +import cv2 + +# OpenCV to habitat camera convention transformation +R_OPENCV2HABITAT = np.stack( + (habitat_sim.geo.RIGHT, -habitat_sim.geo.UP, habitat_sim.geo.FRONT), axis=0 +) +R_HABITAT2OPENCV = R_OPENCV2HABITAT.T +DEG2RAD = np.pi / 180 + + +def compute_camera_intrinsics(height, width, hfov): + f = width / 2 / np.tan(hfov / 2 * np.pi / 180) + cu, cv = width / 2, height / 2 + return f, cu, cv + + +def compute_camera_pose_opencv_convention(camera_position, camera_orientation): + R_cam2world = quaternion.as_rotation_matrix(camera_orientation) @ R_OPENCV2HABITAT + t_cam2world = np.asarray(camera_position) + return R_cam2world, t_cam2world + + +def compute_pointmap(depthmap, hfov): + """Compute a HxWx3 pointmap in camera frame from a HxW depth map.""" + height, width = depthmap.shape + f, cu, cv = compute_camera_intrinsics(height, width, hfov) + # Cast depth map to point + z_cam = depthmap + u, v = np.meshgrid(range(width), range(height)) + x_cam = (u - cu) / f * z_cam + y_cam = (v - cv) / f * z_cam + X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1) + return X_cam + + +def compute_pointcloud(depthmap, hfov, camera_position, camera_rotation): + """Return a 3D point cloud corresponding to valid pixels of the depth map""" + R_cam2world, t_cam2world = compute_camera_pose_opencv_convention( + camera_position, camera_rotation + ) + + X_cam = compute_pointmap(depthmap=depthmap, hfov=hfov) + valid_mask = X_cam[:, :, 2] != 0.0 + + X_cam = X_cam.reshape(-1, 3)[valid_mask.flatten()] + X_world = X_cam @ R_cam2world.T + t_cam2world.reshape(1, 3) + return X_world + + +def compute_pointcloud_overlaps_scikit( + pointcloud1, pointcloud2, distance_threshold, compute_symmetric=False +): + """ + Compute 'overlapping' metrics based on a distance threshold between two point clouds. + """ + nbrs = NearestNeighbors(n_neighbors=1, algorithm="kd_tree").fit(pointcloud2) + distances, indices = nbrs.kneighbors(pointcloud1) + intersection1 = np.count_nonzero(distances.flatten() < distance_threshold) + + data = {"intersection1": intersection1, "size1": len(pointcloud1)} + if compute_symmetric: + nbrs = NearestNeighbors(n_neighbors=1, algorithm="kd_tree").fit(pointcloud1) + distances, indices = nbrs.kneighbors(pointcloud2) + intersection2 = np.count_nonzero(distances.flatten() < distance_threshold) + data["intersection2"] = intersection2 + data["size2"] = len(pointcloud2) + + return data + + +def _append_camera_parameters(observation, hfov, camera_location, camera_rotation): + """ + Add camera parameters to the observation dictionnary produced by Habitat-Sim + In-place modifications. + """ + R_cam2world, t_cam2world = compute_camera_pose_opencv_convention( + camera_location, camera_rotation + ) + height, width = observation["depth"].shape + f, cu, cv = compute_camera_intrinsics(height, width, hfov) + K = np.asarray([[f, 0, cu], [0, f, cv], [0, 0, 1.0]]) + observation["camera_intrinsics"] = K + observation["t_cam2world"] = t_cam2world + observation["R_cam2world"] = R_cam2world + + +def look_at(eye, center, up, return_cam2world=True): + """ + Return camera pose looking at a given center point. + Analogous of gluLookAt function, using OpenCV camera convention. + """ + z = center - eye + z /= np.linalg.norm(z, axis=-1, keepdims=True) + y = -up + y = y - np.sum(y * z, axis=-1, keepdims=True) * z + y /= np.linalg.norm(y, axis=-1, keepdims=True) + x = np.cross(y, z, axis=-1) + + if return_cam2world: + R = np.stack((x, y, z), axis=-1) + t = eye + else: + # World to camera transformation + # Transposed matrix + R = np.stack((x, y, z), axis=-2) + t = -np.einsum("...ij, ...j", R, eye) + return R, t + + +def look_at_for_habitat(eye, center, up, return_cam2world=True): + R, t = look_at(eye, center, up) + orientation = quaternion.from_rotation_matrix(R @ R_OPENCV2HABITAT.T) + return orientation, t + + +def generate_orientation_noise(pan_range, tilt_range, roll_range): + return ( + quaternion.from_rotation_vector( + np.random.uniform(*pan_range) * DEG2RAD * habitat_sim.geo.UP + ) + * quaternion.from_rotation_vector( + np.random.uniform(*tilt_range) * DEG2RAD * habitat_sim.geo.RIGHT + ) + * quaternion.from_rotation_vector( + np.random.uniform(*roll_range) * DEG2RAD * habitat_sim.geo.FRONT + ) + ) + + +class NoNaviguableSpaceError(RuntimeError): + def __init__(self, *args): + super().__init__(*args) + + +class MultiviewHabitatSimGenerator: + def __init__( + self, + scene, + navmesh, + scene_dataset_config_file, + resolution=(240, 320), + views_count=2, + hfov=60, + gpu_id=0, + size=10000, + minimum_covisibility=0.5, + transform=None, + ): + self.scene = scene + self.navmesh = navmesh + self.scene_dataset_config_file = scene_dataset_config_file + self.resolution = resolution + self.views_count = views_count + assert self.views_count >= 1 + self.hfov = hfov + self.gpu_id = gpu_id + self.size = size + self.transform = transform + + # Noise added to camera orientation + self.pan_range = (-3, 3) + self.tilt_range = (-10, 10) + self.roll_range = (-5, 5) + + # Height range to sample cameras + self.height_range = (1.2, 1.8) + + # Random steps between the camera views + self.random_steps_count = 5 + self.random_step_variance = 2.0 + + # Minimum fraction of the scene which should be valid (well defined depth) + self.minimum_valid_fraction = 0.7 + + # Distance threshold to see to select pairs + self.distance_threshold = 0.05 + # Minimum IoU of a view point cloud with respect to the reference view to be kept. + self.minimum_covisibility = minimum_covisibility + + # Maximum number of retries. + self.max_attempts_count = 100 + + self.seed = None + self._lazy_initialization() + + def _lazy_initialization(self): + # Lazy random seeding and instantiation of the simulator to deal with multiprocessing properly + if self.seed == None: + # Re-seed numpy generator + np.random.seed() + self.seed = np.random.randint(2**32 - 1) + sim_cfg = habitat_sim.SimulatorConfiguration() + sim_cfg.scene_id = self.scene + if ( + self.scene_dataset_config_file is not None + and self.scene_dataset_config_file != "" + ): + sim_cfg.scene_dataset_config_file = self.scene_dataset_config_file + sim_cfg.random_seed = self.seed + sim_cfg.load_semantic_mesh = False + sim_cfg.gpu_device_id = self.gpu_id + + depth_sensor_spec = habitat_sim.CameraSensorSpec() + depth_sensor_spec.uuid = "depth" + depth_sensor_spec.sensor_type = habitat_sim.SensorType.DEPTH + depth_sensor_spec.resolution = self.resolution + depth_sensor_spec.hfov = self.hfov + depth_sensor_spec.position = [0.0, 0.0, 0] + depth_sensor_spec.orientation + + rgb_sensor_spec = habitat_sim.CameraSensorSpec() + rgb_sensor_spec.uuid = "color" + rgb_sensor_spec.sensor_type = habitat_sim.SensorType.COLOR + rgb_sensor_spec.resolution = self.resolution + rgb_sensor_spec.hfov = self.hfov + rgb_sensor_spec.position = [0.0, 0.0, 0] + agent_cfg = habitat_sim.agent.AgentConfiguration( + sensor_specifications=[rgb_sensor_spec, depth_sensor_spec] + ) + + cfg = habitat_sim.Configuration(sim_cfg, [agent_cfg]) + self.sim = habitat_sim.Simulator(cfg) + if self.navmesh is not None and self.navmesh != "": + # Use pre-computed navmesh when available (usually better than those generated automatically) + self.sim.pathfinder.load_nav_mesh(self.navmesh) + + if not self.sim.pathfinder.is_loaded: + # Try to compute a navmesh + navmesh_settings = habitat_sim.NavMeshSettings() + navmesh_settings.set_defaults() + self.sim.recompute_navmesh(self.sim.pathfinder, navmesh_settings, True) + + # Ensure that the navmesh is not empty + if not self.sim.pathfinder.is_loaded: + raise NoNaviguableSpaceError( + f"No naviguable location (scene: {self.scene} -- navmesh: {self.navmesh})" + ) + + self.agent = self.sim.initialize_agent(agent_id=0) + + def close(self): + self.sim.close() + + def __del__(self): + self.sim.close() + + def __len__(self): + return self.size + + def sample_random_viewpoint(self): + """Sample a random viewpoint using the navmesh""" + nav_point = self.sim.pathfinder.get_random_navigable_point() + + # Sample a random viewpoint height + viewpoint_height = np.random.uniform(*self.height_range) + viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP + viewpoint_orientation = quaternion.from_rotation_vector( + np.random.uniform(0, 2 * np.pi) * habitat_sim.geo.UP + ) * generate_orientation_noise(self.pan_range, self.tilt_range, self.roll_range) + return viewpoint_position, viewpoint_orientation, nav_point + + def sample_other_random_viewpoint(self, observed_point, nav_point): + """Sample a random viewpoint close to an existing one, using the navmesh and a reference observed point.""" + other_nav_point = nav_point + + walk_directions = self.random_step_variance * np.asarray([1, 0, 1]) + for i in range(self.random_steps_count): + temp = self.sim.pathfinder.snap_point( + other_nav_point + walk_directions * np.random.normal(size=3) + ) + # Snapping may return nan when it fails + if not np.isnan(temp[0]): + other_nav_point = temp + + other_viewpoint_height = np.random.uniform(*self.height_range) + other_viewpoint_position = ( + other_nav_point + other_viewpoint_height * habitat_sim.geo.UP + ) + + # Set viewing direction towards the central point + rotation, position = look_at_for_habitat( + eye=other_viewpoint_position, + center=observed_point, + up=habitat_sim.geo.UP, + return_cam2world=True, + ) + rotation = rotation * generate_orientation_noise( + self.pan_range, self.tilt_range, self.roll_range + ) + return position, rotation, other_nav_point + + def is_other_pointcloud_overlapping(self, ref_pointcloud, other_pointcloud): + """Check if a viewpoint is valid and overlaps significantly with a reference one.""" + # Observation + pixels_count = self.resolution[0] * self.resolution[1] + valid_fraction = len(other_pointcloud) / pixels_count + assert valid_fraction <= 1.0 and valid_fraction >= 0.0 + overlap = compute_pointcloud_overlaps_scikit( + ref_pointcloud, + other_pointcloud, + self.distance_threshold, + compute_symmetric=True, + ) + covisibility = min( + overlap["intersection1"] / pixels_count, + overlap["intersection2"] / pixels_count, + ) + is_valid = (valid_fraction >= self.minimum_valid_fraction) and ( + covisibility >= self.minimum_covisibility + ) + return is_valid, valid_fraction, covisibility + + def is_other_viewpoint_overlapping( + self, ref_pointcloud, observation, position, rotation + ): + """Check if a viewpoint is valid and overlaps significantly with a reference one.""" + # Observation + other_pointcloud = compute_pointcloud( + observation["depth"], self.hfov, position, rotation + ) + return self.is_other_pointcloud_overlapping(ref_pointcloud, other_pointcloud) + + def render_viewpoint(self, viewpoint_position, viewpoint_orientation): + agent_state = habitat_sim.AgentState() + agent_state.position = viewpoint_position + agent_state.rotation = viewpoint_orientation + self.agent.set_state(agent_state) + viewpoint_observations = self.sim.get_sensor_observations(agent_ids=0) + _append_camera_parameters( + viewpoint_observations, self.hfov, viewpoint_position, viewpoint_orientation + ) + return viewpoint_observations + + def __getitem__(self, useless_idx): + ref_position, ref_orientation, nav_point = self.sample_random_viewpoint() + ref_observations = self.render_viewpoint(ref_position, ref_orientation) + # Extract point cloud + ref_pointcloud = compute_pointcloud( + depthmap=ref_observations["depth"], + hfov=self.hfov, + camera_position=ref_position, + camera_rotation=ref_orientation, + ) + + pixels_count = self.resolution[0] * self.resolution[1] + ref_valid_fraction = len(ref_pointcloud) / pixels_count + assert ref_valid_fraction <= 1.0 and ref_valid_fraction >= 0.0 + if ref_valid_fraction < self.minimum_valid_fraction: + # This should produce a recursion error at some point when something is very wrong. + return self[0] + # Pick an reference observed point in the point cloud + observed_point = np.mean(ref_pointcloud, axis=0) + + # Add the first image as reference + viewpoints_observations = [ref_observations] + viewpoints_covisibility = [ref_valid_fraction] + viewpoints_positions = [ref_position] + viewpoints_orientations = [quaternion.as_float_array(ref_orientation)] + viewpoints_clouds = [ref_pointcloud] + viewpoints_valid_fractions = [ref_valid_fraction] + + for _ in range(self.views_count - 1): + # Generate an other viewpoint using some dummy random walk + successful_sampling = False + for sampling_attempt in range(self.max_attempts_count): + position, rotation, _ = self.sample_other_random_viewpoint( + observed_point, nav_point + ) + # Observation + other_viewpoint_observations = self.render_viewpoint(position, rotation) + other_pointcloud = compute_pointcloud( + other_viewpoint_observations["depth"], self.hfov, position, rotation + ) + + is_valid, valid_fraction, covisibility = ( + self.is_other_pointcloud_overlapping( + ref_pointcloud, other_pointcloud + ) + ) + if is_valid: + successful_sampling = True + break + if not successful_sampling: + print("WARNING: Maximum number of attempts reached.") + # Dirty hack, try using a novel original viewpoint + return self[0] + viewpoints_observations.append(other_viewpoint_observations) + viewpoints_covisibility.append(covisibility) + viewpoints_positions.append(position) + viewpoints_orientations.append( + quaternion.as_float_array(rotation) + ) # WXYZ convention for the quaternion encoding. + viewpoints_clouds.append(other_pointcloud) + viewpoints_valid_fractions.append(valid_fraction) + + # Estimate relations between all pairs of images + pairwise_visibility_ratios = np.ones( + (len(viewpoints_observations), len(viewpoints_observations)) + ) + for i in range(len(viewpoints_observations)): + pairwise_visibility_ratios[i, i] = viewpoints_valid_fractions[i] + for j in range(i + 1, len(viewpoints_observations)): + overlap = compute_pointcloud_overlaps_scikit( + viewpoints_clouds[i], + viewpoints_clouds[j], + self.distance_threshold, + compute_symmetric=True, + ) + pairwise_visibility_ratios[i, j] = ( + overlap["intersection1"] / pixels_count + ) + pairwise_visibility_ratios[j, i] = ( + overlap["intersection2"] / pixels_count + ) + + # IoU is relative to the image 0 + data = { + "observations": viewpoints_observations, + "positions": np.asarray(viewpoints_positions), + "orientations": np.asarray(viewpoints_orientations), + "covisibility_ratios": np.asarray(viewpoints_covisibility), + "valid_fractions": np.asarray(viewpoints_valid_fractions, dtype=float), + "pairwise_visibility_ratios": np.asarray( + pairwise_visibility_ratios, dtype=float + ), + } + + if self.transform is not None: + data = self.transform(data) + return data + + def generate_random_spiral_trajectory( + self, + images_count=100, + max_radius=0.5, + half_turns=5, + use_constant_orientation=False, + ): + """ + Return a list of images corresponding to a spiral trajectory from a random starting point. + Useful to generate nice visualisations. + Use an even number of half turns to get a nice "C1-continuous" loop effect + """ + ref_position, ref_orientation, navpoint = self.sample_random_viewpoint() + ref_observations = self.render_viewpoint(ref_position, ref_orientation) + ref_pointcloud = compute_pointcloud( + depthmap=ref_observations["depth"], + hfov=self.hfov, + camera_position=ref_position, + camera_rotation=ref_orientation, + ) + pixels_count = self.resolution[0] * self.resolution[1] + if len(ref_pointcloud) / pixels_count < self.minimum_valid_fraction: + # Dirty hack: ensure that the valid part of the image is significant + return self.generate_random_spiral_trajectory( + images_count, max_radius, half_turns, use_constant_orientation + ) + + # Pick an observed point in the point cloud + observed_point = np.mean(ref_pointcloud, axis=0) + ref_R, ref_t = compute_camera_pose_opencv_convention( + ref_position, ref_orientation + ) + + images = [] + is_valid = [] + # Spiral trajectory, use_constant orientation + for i, alpha in enumerate(np.linspace(0, 1, images_count)): + r = max_radius * np.abs( + np.sin(alpha * np.pi) + ) # Increase then decrease the radius + theta = alpha * half_turns * np.pi + x = r * np.cos(theta) + y = r * np.sin(theta) + z = 0.0 + position = ( + ref_position + (ref_R @ np.asarray([x, y, z]).reshape(3, 1)).flatten() + ) + if use_constant_orientation: + orientation = ref_orientation + else: + # trajectory looking at a mean point in front of the ref observation + orientation, position = look_at_for_habitat( + eye=position, center=observed_point, up=habitat_sim.geo.UP + ) + observations = self.render_viewpoint(position, orientation) + images.append(observations["color"][..., :3]) + _is_valid, valid_fraction, iou = self.is_other_viewpoint_overlapping( + ref_pointcloud, observations, position, orientation + ) + is_valid.append(_is_valid) + return images, np.all(is_valid) diff --git a/croco/datasets/habitat_sim/pack_metadata_files.py b/croco/datasets/habitat_sim/pack_metadata_files.py new file mode 100644 index 0000000000000000000000000000000000000000..9bd8234dfaa491d5f25f7c778406255116a8b392 --- /dev/null +++ b/croco/datasets/habitat_sim/pack_metadata_files.py @@ -0,0 +1,80 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +""" +Utility script to pack metadata files of the dataset in order to be able to re-generate it elsewhere. +""" +import os +import glob +from tqdm import tqdm +import shutil +import json +from datasets.habitat_sim.paths import * +import argparse +import collections + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("input_dir") + parser.add_argument("output_dir") + args = parser.parse_args() + + input_dirname = args.input_dir + output_dirname = args.output_dir + + input_metadata_filenames = glob.iglob( + f"{input_dirname}/**/metadata.json", recursive=True + ) + + images_count = collections.defaultdict(lambda: 0) + + os.makedirs(output_dirname) + for input_filename in tqdm(input_metadata_filenames): + # Ignore empty files + with open(input_filename, "r") as f: + original_metadata = json.load(f) + if ( + "multiviews" not in original_metadata + or len(original_metadata["multiviews"]) == 0 + ): + print("No views in", input_filename) + continue + + relpath = os.path.relpath(input_filename, input_dirname) + print(relpath) + + # Copy metadata, while replacing scene paths by generic keys depending on the dataset, for portability. + # Data paths are sorted by decreasing length to avoid potential bugs due to paths starting by the same string pattern. + scenes_dataset_paths = dict( + sorted(SCENES_DATASET.items(), key=lambda x: len(x[1]), reverse=True) + ) + metadata = dict() + for key, value in original_metadata.items(): + if key in ("scene_dataset_config_file", "scene", "navmesh") and value != "": + known_path = False + for dataset, dataset_path in scenes_dataset_paths.items(): + if value.startswith(dataset_path): + value = os.path.join( + dataset, os.path.relpath(value, dataset_path) + ) + known_path = True + break + if not known_path: + raise KeyError("Unknown path:" + value) + metadata[key] = value + + # Compile some general statistics while packing data + scene_split = metadata["scene"].split("/") + upper_level = ( + "/".join(scene_split[:2]) if scene_split[0] == "hm3d" else scene_split[0] + ) + images_count[upper_level] += len(metadata["multiviews"]) + + output_filename = os.path.join(output_dirname, relpath) + os.makedirs(os.path.dirname(output_filename), exist_ok=True) + with open(output_filename, "w") as f: + json.dump(metadata, f) + + # Print statistics + print("Images count:") + for upper_level, count in images_count.items(): + print(f"- {upper_level}: {count}") diff --git a/croco/datasets/habitat_sim/paths.py b/croco/datasets/habitat_sim/paths.py new file mode 100644 index 0000000000000000000000000000000000000000..87389fcff93d220d6f205dc21119da3c56c3abb9 --- /dev/null +++ b/croco/datasets/habitat_sim/paths.py @@ -0,0 +1,179 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +""" +Paths to Habitat-Sim scenes +""" + +import os +import json +import collections +from tqdm import tqdm + + +# Hardcoded path to the different scene datasets +SCENES_DATASET = { + "hm3d": "./data/habitat-sim-data/scene_datasets/hm3d/", + "gibson": "./data/habitat-sim-data/scene_datasets/gibson/", + "habitat-test-scenes": "./data/habitat-sim/scene_datasets/habitat-test-scenes/", + "replica_cad_baked_lighting": "./data/habitat-sim/scene_datasets/replica_cad_baked_lighting/", + "replica_cad": "./data/habitat-sim/scene_datasets/replica_cad/", + "replica": "./data/habitat-sim/scene_datasets/ReplicaDataset/", + "scannet": "./data/habitat-sim/scene_datasets/scannet/", +} + +SceneData = collections.namedtuple( + "SceneData", ["scene_dataset_config_file", "scene", "navmesh", "output_dir"] +) + + +def list_replicacad_scenes(base_output_dir, base_path=SCENES_DATASET["replica_cad"]): + scene_dataset_config_file = os.path.join( + base_path, "replicaCAD.scene_dataset_config.json" + ) + scenes = [f"apt_{i}" for i in range(6)] + ["empty_stage"] + navmeshes = [f"navmeshes/apt_{i}_static_furniture.navmesh" for i in range(6)] + [ + "empty_stage.navmesh" + ] + scenes_data = [] + for idx in range(len(scenes)): + output_dir = os.path.join(base_output_dir, "ReplicaCAD", scenes[idx]) + # Add scene + data = SceneData( + scene_dataset_config_file=scene_dataset_config_file, + scene=scenes[idx] + ".scene_instance.json", + navmesh=os.path.join(base_path, navmeshes[idx]), + output_dir=output_dir, + ) + scenes_data.append(data) + return scenes_data + + +def list_replica_cad_baked_lighting_scenes( + base_output_dir, base_path=SCENES_DATASET["replica_cad_baked_lighting"] +): + scene_dataset_config_file = os.path.join( + base_path, "replicaCAD_baked.scene_dataset_config.json" + ) + scenes = sum( + [[f"Baked_sc{i}_staging_{j:02}" for i in range(5)] for j in range(21)], [] + ) + navmeshes = "" # [f"navmeshes/apt_{i}_static_furniture.navmesh" for i in range(6)] + ["empty_stage.navmesh"] + scenes_data = [] + for idx in range(len(scenes)): + output_dir = os.path.join( + base_output_dir, "replica_cad_baked_lighting", scenes[idx] + ) + data = SceneData( + scene_dataset_config_file=scene_dataset_config_file, + scene=scenes[idx], + navmesh="", + output_dir=output_dir, + ) + scenes_data.append(data) + return scenes_data + + +def list_replica_scenes(base_output_dir, base_path): + scenes_data = [] + for scene_id in os.listdir(base_path): + scene = os.path.join(base_path, scene_id, "mesh.ply") + navmesh = os.path.join( + base_path, scene_id, "habitat/mesh_preseg_semantic.navmesh" + ) # Not sure if I should use it + scene_dataset_config_file = "" + output_dir = os.path.join(base_output_dir, scene_id) + # Add scene only if it does not exist already, or if exist_ok + data = SceneData( + scene_dataset_config_file=scene_dataset_config_file, + scene=scene, + navmesh=navmesh, + output_dir=output_dir, + ) + scenes_data.append(data) + return scenes_data + + +def list_scenes(base_output_dir, base_path): + """ + Generic method iterating through a base_path folder to find scenes. + """ + scenes_data = [] + for root, dirs, files in os.walk(base_path, followlinks=True): + folder_scenes_data = [] + for file in files: + name, ext = os.path.splitext(file) + if ext == ".glb": + scene = os.path.join(root, name + ".glb") + navmesh = os.path.join(root, name + ".navmesh") + if not os.path.exists(navmesh): + navmesh = "" + relpath = os.path.relpath(root, base_path) + output_dir = os.path.abspath( + os.path.join(base_output_dir, relpath, name) + ) + data = SceneData( + scene_dataset_config_file="", + scene=scene, + navmesh=navmesh, + output_dir=output_dir, + ) + folder_scenes_data.append(data) + + # Specific check for HM3D: + # When two meshesxxxx.basis.glb and xxxx.glb are present, use the 'basis' version. + basis_scenes = [ + data.scene[: -len(".basis.glb")] + for data in folder_scenes_data + if data.scene.endswith(".basis.glb") + ] + if len(basis_scenes) != 0: + folder_scenes_data = [ + data + for data in folder_scenes_data + if not (data.scene[: -len(".glb")] in basis_scenes) + ] + + scenes_data.extend(folder_scenes_data) + return scenes_data + + +def list_scenes_available(base_output_dir, scenes_dataset_paths=SCENES_DATASET): + scenes_data = [] + + # HM3D + for split in ("minival", "train", "val", "examples"): + scenes_data += list_scenes( + base_output_dir=os.path.join(base_output_dir, f"hm3d/{split}/"), + base_path=f"{scenes_dataset_paths['hm3d']}/{split}", + ) + + # Gibson + scenes_data += list_scenes( + base_output_dir=os.path.join(base_output_dir, "gibson"), + base_path=scenes_dataset_paths["gibson"], + ) + + # Habitat test scenes (just a few) + scenes_data += list_scenes( + base_output_dir=os.path.join(base_output_dir, "habitat-test-scenes"), + base_path=scenes_dataset_paths["habitat-test-scenes"], + ) + + # ReplicaCAD (baked lightning) + scenes_data += list_replica_cad_baked_lighting_scenes( + base_output_dir=base_output_dir + ) + + # ScanNet + scenes_data += list_scenes( + base_output_dir=os.path.join(base_output_dir, "scannet"), + base_path=scenes_dataset_paths["scannet"], + ) + + # Replica + list_replica_scenes( + base_output_dir=os.path.join(base_output_dir, "replica"), + base_path=scenes_dataset_paths["replica"], + ) + return scenes_data diff --git a/croco/datasets/pairs_dataset.py b/croco/datasets/pairs_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..066bb9510332255edd211f98f2beb6670abff4f9 --- /dev/null +++ b/croco/datasets/pairs_dataset.py @@ -0,0 +1,162 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +import os +from torch.utils.data import Dataset +from PIL import Image + +from datasets.transforms import get_pair_transforms + + +def load_image(impath): + return Image.open(impath) + + +def load_pairs_from_cache_file(fname, root=""): + assert os.path.isfile( + fname + ), "cannot parse pairs from {:s}, file does not exist".format(fname) + with open(fname, "r") as fid: + lines = fid.read().strip().splitlines() + pairs = [ + (os.path.join(root, l.split()[0]), os.path.join(root, l.split()[1])) + for l in lines + ] + return pairs + + +def load_pairs_from_list_file(fname, root=""): + assert os.path.isfile( + fname + ), "cannot parse pairs from {:s}, file does not exist".format(fname) + with open(fname, "r") as fid: + lines = fid.read().strip().splitlines() + pairs = [ + (os.path.join(root, l + "_1.jpg"), os.path.join(root, l + "_2.jpg")) + for l in lines + if not l.startswith("#") + ] + return pairs + + +def write_cache_file(fname, pairs, root=""): + if len(root) > 0: + if not root.endswith("/"): + root += "/" + assert os.path.isdir(root) + s = "" + for im1, im2 in pairs: + if len(root) > 0: + assert im1.startswith(root), im1 + assert im2.startswith(root), im2 + s += "{:s} {:s}\n".format(im1[len(root) :], im2[len(root) :]) + with open(fname, "w") as fid: + fid.write(s[:-1]) + + +def parse_and_cache_all_pairs(dname, data_dir="./data/"): + if dname == "habitat_release": + dirname = os.path.join(data_dir, "habitat_release") + assert os.path.isdir(dirname), ( + "cannot find folder for habitat_release pairs: " + dirname + ) + cache_file = os.path.join(dirname, "pairs.txt") + assert not os.path.isfile(cache_file), ( + "cache file already exists: " + cache_file + ) + + print("Parsing pairs for dataset: " + dname) + pairs = [] + for root, dirs, files in os.walk(dirname): + if "val" in root: + continue + dirs.sort() + pairs += [ + ( + os.path.join(root, f), + os.path.join(root, f[: -len("_1.jpeg")] + "_2.jpeg"), + ) + for f in sorted(files) + if f.endswith("_1.jpeg") + ] + print("Found {:,} pairs".format(len(pairs))) + print("Writing cache to: " + cache_file) + write_cache_file(cache_file, pairs, root=dirname) + + else: + raise NotImplementedError("Unknown dataset: " + dname) + + +def dnames_to_image_pairs(dnames, data_dir="./data/"): + """ + dnames: list of datasets with image pairs, separated by + + """ + all_pairs = [] + for dname in dnames.split("+"): + if dname == "habitat_release": + dirname = os.path.join(data_dir, "habitat_release") + assert os.path.isdir(dirname), ( + "cannot find folder for habitat_release pairs: " + dirname + ) + cache_file = os.path.join(dirname, "pairs.txt") + assert os.path.isfile(cache_file), ( + "cannot find cache file for habitat_release pairs, please first create the cache file, see instructions. " + + cache_file + ) + pairs = load_pairs_from_cache_file(cache_file, root=dirname) + elif dname in ["ARKitScenes", "MegaDepth", "3DStreetView", "IndoorVL"]: + dirname = os.path.join(data_dir, dname + "_crops") + assert os.path.isdir( + dirname + ), "cannot find folder for {:s} pairs: {:s}".format(dname, dirname) + list_file = os.path.join(dirname, "listing.txt") + assert os.path.isfile( + list_file + ), "cannot find list file for {:s} pairs, see instructions. {:s}".format( + dname, list_file + ) + pairs = load_pairs_from_list_file(list_file, root=dirname) + print(" {:s}: {:,} pairs".format(dname, len(pairs))) + all_pairs += pairs + if "+" in dnames: + print(" Total: {:,} pairs".format(len(all_pairs))) + return all_pairs + + +class PairsDataset(Dataset): + + def __init__( + self, dnames, trfs="", totensor=True, normalize=True, data_dir="./data/" + ): + super().__init__() + self.image_pairs = dnames_to_image_pairs(dnames, data_dir=data_dir) + self.transforms = get_pair_transforms( + transform_str=trfs, totensor=totensor, normalize=normalize + ) + + def __len__(self): + return len(self.image_pairs) + + def __getitem__(self, index): + im1path, im2path = self.image_pairs[index] + im1 = load_image(im1path) + im2 = load_image(im2path) + if self.transforms is not None: + im1, im2 = self.transforms(im1, im2) + return im1, im2 + + +if __name__ == "__main__": + import argparse + + parser = argparse.ArgumentParser( + prog="Computing and caching list of pairs for a given dataset" + ) + parser.add_argument( + "--data_dir", default="./data/", type=str, help="path where data are stored" + ) + parser.add_argument( + "--dataset", default="habitat_release", type=str, help="name of the dataset" + ) + args = parser.parse_args() + parse_and_cache_all_pairs(dname=args.dataset, data_dir=args.data_dir) diff --git a/croco/datasets/transforms.py b/croco/datasets/transforms.py new file mode 100644 index 0000000000000000000000000000000000000000..5dc89dd1092293f63035afd70e9ef9f907696f44 --- /dev/null +++ b/croco/datasets/transforms.py @@ -0,0 +1,135 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +import torch +import torchvision.transforms +import torchvision.transforms.functional as F + +# "Pair": apply a transform on a pair +# "Both": apply the exact same transform to both images + + +class ComposePair(torchvision.transforms.Compose): + def __call__(self, img1, img2): + for t in self.transforms: + img1, img2 = t(img1, img2) + return img1, img2 + + +class NormalizeBoth(torchvision.transforms.Normalize): + def forward(self, img1, img2): + img1 = super().forward(img1) + img2 = super().forward(img2) + return img1, img2 + + +class ToTensorBoth(torchvision.transforms.ToTensor): + def __call__(self, img1, img2): + img1 = super().__call__(img1) + img2 = super().__call__(img2) + return img1, img2 + + +class RandomCropPair(torchvision.transforms.RandomCrop): + # the crop will be intentionally different for the two images with this class + def forward(self, img1, img2): + img1 = super().forward(img1) + img2 = super().forward(img2) + return img1, img2 + + +class ColorJitterPair(torchvision.transforms.ColorJitter): + # can be symmetric (same for both images) or assymetric (different jitter params for each image) depending on assymetric_prob + def __init__(self, assymetric_prob, **kwargs): + super().__init__(**kwargs) + self.assymetric_prob = assymetric_prob + + def jitter_one( + self, + img, + fn_idx, + brightness_factor, + contrast_factor, + saturation_factor, + hue_factor, + ): + for fn_id in fn_idx: + if fn_id == 0 and brightness_factor is not None: + img = F.adjust_brightness(img, brightness_factor) + elif fn_id == 1 and contrast_factor is not None: + img = F.adjust_contrast(img, contrast_factor) + elif fn_id == 2 and saturation_factor is not None: + img = F.adjust_saturation(img, saturation_factor) + elif fn_id == 3 and hue_factor is not None: + img = F.adjust_hue(img, hue_factor) + return img + + def forward(self, img1, img2): + + fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = ( + self.get_params(self.brightness, self.contrast, self.saturation, self.hue) + ) + img1 = self.jitter_one( + img1, + fn_idx, + brightness_factor, + contrast_factor, + saturation_factor, + hue_factor, + ) + if torch.rand(1) < self.assymetric_prob: # assymetric: + ( + fn_idx, + brightness_factor, + contrast_factor, + saturation_factor, + hue_factor, + ) = self.get_params( + self.brightness, self.contrast, self.saturation, self.hue + ) + img2 = self.jitter_one( + img2, + fn_idx, + brightness_factor, + contrast_factor, + saturation_factor, + hue_factor, + ) + return img1, img2 + + +def get_pair_transforms(transform_str, totensor=True, normalize=True): + # transform_str is eg crop224+color + trfs = [] + for s in transform_str.split("+"): + if s.startswith("crop"): + size = int(s[len("crop") :]) + trfs.append(RandomCropPair(size)) + elif s == "acolor": + trfs.append( + ColorJitterPair( + assymetric_prob=1.0, + brightness=(0.6, 1.4), + contrast=(0.6, 1.4), + saturation=(0.6, 1.4), + hue=0.0, + ) + ) + elif s == "": # if transform_str was "" + pass + else: + raise NotImplementedError("Unknown augmentation: " + s) + + if totensor: + trfs.append(ToTensorBoth()) + if normalize: + trfs.append( + NormalizeBoth(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) + ) + + if len(trfs) == 0: + return None + elif len(trfs) == 1: + return trfs + else: + return ComposePair(trfs) diff --git a/croco/interactive_demo.ipynb b/croco/interactive_demo.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..6cfc960af5baac9a69029c29a16eea4e24123a71 --- /dev/null +++ b/croco/interactive_demo.ipynb @@ -0,0 +1,271 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "# Interactive demo of Cross-view Completion." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "# Copyright (C) 2022-present Naver Corporation. All rights reserved.\n", + "# Licensed under CC BY-NC-SA 4.0 (non-commercial use only)." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "import torch\n", + "import numpy as np\n", + "from models.croco import CroCoNet\n", + "from ipywidgets import interact, interactive, fixed, interact_manual\n", + "import ipywidgets as widgets\n", + "import matplotlib.pyplot as plt\n", + "import quaternion\n", + "import models.masking" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "### Load CroCo model" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "ckpt = torch.load('pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth', 'cpu')\n", + "model = CroCoNet( **ckpt.get('croco_kwargs',{}))\n", + "msg = model.load_state_dict(ckpt['model'], strict=True)\n", + "use_gpu = torch.cuda.is_available() and torch.cuda.device_count()>0\n", + "device = torch.device('cuda:0' if use_gpu else 'cpu')\n", + "model = model.eval()\n", + "model = model.to(device=device)\n", + "print(msg)\n", + "\n", + "def process_images(ref_image, target_image, masking_ratio, reconstruct_unmasked_patches=False):\n", + " \"\"\"\n", + " Perform Cross-View completion using two input images, specified using Numpy arrays.\n", + " \"\"\"\n", + " # Replace the mask generator\n", + " model.mask_generator = models.masking.RandomMask(model.patch_embed.num_patches, masking_ratio)\n", + "\n", + " # ImageNet-1k color normalization\n", + " imagenet_mean = torch.as_tensor([0.485, 0.456, 0.406]).reshape(1,3,1,1).to(device)\n", + " imagenet_std = torch.as_tensor([0.229, 0.224, 0.225]).reshape(1,3,1,1).to(device)\n", + "\n", + " normalize_input_colors = True\n", + " is_output_normalized = True\n", + " with torch.no_grad():\n", + " # Cast data to torch\n", + " target_image = (torch.as_tensor(target_image, dtype=torch.float, device=device).permute(2,0,1) / 255)[None]\n", + " ref_image = (torch.as_tensor(ref_image, dtype=torch.float, device=device).permute(2,0,1) / 255)[None]\n", + "\n", + " if normalize_input_colors:\n", + " ref_image = (ref_image - imagenet_mean) / imagenet_std\n", + " target_image = (target_image - imagenet_mean) / imagenet_std\n", + "\n", + " out, mask, _ = model(target_image, ref_image)\n", + " # # get target\n", + " if not is_output_normalized:\n", + " predicted_image = model.unpatchify(out)\n", + " else:\n", + " # The output only contains higher order information,\n", + " # we retrieve mean and standard deviation from the actual target image\n", + " patchified = model.patchify(target_image)\n", + " mean = patchified.mean(dim=-1, keepdim=True)\n", + " var = patchified.var(dim=-1, keepdim=True)\n", + " pred_renorm = out * (var + 1.e-6)**.5 + mean\n", + " predicted_image = model.unpatchify(pred_renorm)\n", + "\n", + " image_masks = model.unpatchify(model.patchify(torch.ones_like(ref_image)) * mask[:,:,None])\n", + " masked_target_image = (1 - image_masks) * target_image\n", + " \n", + " if not reconstruct_unmasked_patches:\n", + " # Replace unmasked patches by their actual values\n", + " predicted_image = predicted_image * image_masks + masked_target_image\n", + "\n", + " # Unapply color normalization\n", + " if normalize_input_colors:\n", + " predicted_image = predicted_image * imagenet_std + imagenet_mean\n", + " masked_target_image = masked_target_image * imagenet_std + imagenet_mean\n", + " \n", + " # Cast to Numpy\n", + " masked_target_image = np.asarray(torch.clamp(masked_target_image.squeeze(0).permute(1,2,0) * 255, 0, 255).cpu().numpy(), dtype=np.uint8)\n", + " predicted_image = np.asarray(torch.clamp(predicted_image.squeeze(0).permute(1,2,0) * 255, 0, 255).cpu().numpy(), dtype=np.uint8)\n", + " return masked_target_image, predicted_image" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "### Use the Habitat simulator to render images from arbitrary viewpoints (requires habitat_sim to be installed)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "import os\n", + "os.environ[\"MAGNUM_LOG\"]=\"quiet\"\n", + "os.environ[\"HABITAT_SIM_LOG\"]=\"quiet\"\n", + "import habitat_sim\n", + "\n", + "scene = \"habitat-sim-data/scene_datasets/habitat-test-scenes/skokloster-castle.glb\"\n", + "navmesh = \"habitat-sim-data/scene_datasets/habitat-test-scenes/skokloster-castle.navmesh\"\n", + "\n", + "sim_cfg = habitat_sim.SimulatorConfiguration()\n", + "if use_gpu: sim_cfg.gpu_device_id = 0\n", + "sim_cfg.scene_id = scene\n", + "sim_cfg.load_semantic_mesh = False\n", + "rgb_sensor_spec = habitat_sim.CameraSensorSpec()\n", + "rgb_sensor_spec.uuid = \"color\"\n", + "rgb_sensor_spec.sensor_type = habitat_sim.SensorType.COLOR\n", + "rgb_sensor_spec.resolution = (224,224)\n", + "rgb_sensor_spec.hfov = 56.56\n", + "rgb_sensor_spec.position = [0.0, 0.0, 0.0]\n", + "rgb_sensor_spec.orientation = [0, 0, 0]\n", + "agent_cfg = habitat_sim.agent.AgentConfiguration(sensor_specifications=[rgb_sensor_spec])\n", + "\n", + "\n", + "cfg = habitat_sim.Configuration(sim_cfg, [agent_cfg])\n", + "sim = habitat_sim.Simulator(cfg)\n", + "if navmesh is not None:\n", + " sim.pathfinder.load_nav_mesh(navmesh)\n", + "agent = sim.initialize_agent(agent_id=0)\n", + "\n", + "def sample_random_viewpoint():\n", + " \"\"\" Sample a random viewpoint using the navmesh \"\"\"\n", + " nav_point = sim.pathfinder.get_random_navigable_point()\n", + " # Sample a random viewpoint height\n", + " viewpoint_height = np.random.uniform(1.0, 1.6)\n", + " viewpoint_position = nav_point + viewpoint_height * habitat_sim.geo.UP\n", + " viewpoint_orientation = quaternion.from_rotation_vector(np.random.uniform(-np.pi, np.pi) * habitat_sim.geo.UP)\n", + " return viewpoint_position, viewpoint_orientation\n", + "\n", + "def render_viewpoint(position, orientation):\n", + " agent_state = habitat_sim.AgentState()\n", + " agent_state.position = position\n", + " agent_state.rotation = orientation\n", + " agent.set_state(agent_state)\n", + " viewpoint_observations = sim.get_sensor_observations(agent_ids=0)\n", + " image = viewpoint_observations['color'][:,:,:3]\n", + " image = np.asarray(np.clip(1.5 * np.asarray(image, dtype=float), 0, 255), dtype=np.uint8)\n", + " return image" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "### Sample a random reference view" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "ref_position, ref_orientation = sample_random_viewpoint()\n", + "ref_image = render_viewpoint(ref_position, ref_orientation)\n", + "plt.clf()\n", + "fig, axes = plt.subplots(1,1, squeeze=False, num=1)\n", + "axes[0,0].imshow(ref_image)\n", + "for ax in axes.flatten():\n", + " ax.set_xticks([])\n", + " ax.set_yticks([])" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "### Interactive cross-view completion using CroCo" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "reconstruct_unmasked_patches = False\n", + "\n", + "def show_demo(masking_ratio, x, y, z, panorama, elevation):\n", + " R = quaternion.as_rotation_matrix(ref_orientation)\n", + " target_position = ref_position + x * R[:,0] + y * R[:,1] + z * R[:,2]\n", + " target_orientation = (ref_orientation\n", + " * quaternion.from_rotation_vector(-elevation * np.pi/180 * habitat_sim.geo.LEFT) \n", + " * quaternion.from_rotation_vector(-panorama * np.pi/180 * habitat_sim.geo.UP))\n", + " \n", + " ref_image = render_viewpoint(ref_position, ref_orientation)\n", + " target_image = render_viewpoint(target_position, target_orientation)\n", + "\n", + " masked_target_image, predicted_image = process_images(ref_image, target_image, masking_ratio, reconstruct_unmasked_patches)\n", + "\n", + " fig, axes = plt.subplots(1,4, squeeze=True, dpi=300)\n", + " axes[0].imshow(ref_image)\n", + " axes[0].set_xlabel(\"Reference\")\n", + " axes[1].imshow(masked_target_image)\n", + " axes[1].set_xlabel(\"Masked target\")\n", + " axes[2].imshow(predicted_image)\n", + " axes[2].set_xlabel(\"Reconstruction\") \n", + " axes[3].imshow(target_image)\n", + " axes[3].set_xlabel(\"Target\")\n", + " for ax in axes.flatten():\n", + " ax.set_xticks([])\n", + " ax.set_yticks([])\n", + "\n", + "interact(show_demo,\n", + " masking_ratio=widgets.FloatSlider(description='masking', value=0.9, min=0.0, max=1.0),\n", + " x=widgets.FloatSlider(value=0.0, min=-0.5, max=0.5, step=0.05),\n", + " y=widgets.FloatSlider(value=0.0, min=-0.5, max=0.5, step=0.05),\n", + " z=widgets.FloatSlider(value=0.0, min=-0.5, max=0.5, step=0.05),\n", + " panorama=widgets.FloatSlider(value=0.0, min=-20, max=20, step=0.5),\n", + " elevation=widgets.FloatSlider(value=0.0, min=-20, max=20, step=0.5));" + ] + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 3 (ipykernel)", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.7.13" + }, + "vscode": { + "interpreter": { + "hash": "f9237820cd248d7e07cb4fb9f0e4508a85d642f19d831560c0a4b61f3e907e67" + } + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/croco/models/__pycache__/blocks.cpython-310.pyc b/croco/models/__pycache__/blocks.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2eaa51e1d73dce66aa75ab5468ad53b3703583bc Binary files /dev/null and b/croco/models/__pycache__/blocks.cpython-310.pyc differ diff --git a/croco/models/__pycache__/blocks.cpython-311.pyc b/croco/models/__pycache__/blocks.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..af7b6d413284e285dc55a9528708a77c1ba44eba Binary files /dev/null and b/croco/models/__pycache__/blocks.cpython-311.pyc differ diff --git a/croco/models/__pycache__/blocks.cpython-312.pyc b/croco/models/__pycache__/blocks.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3a4f7e03e93f0fcad6ebeb8719599342c9a2fdd2 Binary files /dev/null and b/croco/models/__pycache__/blocks.cpython-312.pyc differ diff --git a/croco/models/__pycache__/croco.cpython-310.pyc b/croco/models/__pycache__/croco.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f8a3d431cfc9fea6634d9321f6b5815ba4c72172 Binary files /dev/null and b/croco/models/__pycache__/croco.cpython-310.pyc differ diff --git a/croco/models/__pycache__/croco.cpython-311.pyc b/croco/models/__pycache__/croco.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..332737c7d97e05bec7e4d945c81c9692a97dc8a3 Binary files /dev/null and b/croco/models/__pycache__/croco.cpython-311.pyc differ diff --git a/croco/models/__pycache__/croco.cpython-312.pyc b/croco/models/__pycache__/croco.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5f9ea20998d02164ae81b72867ed462a8351dcf5 Binary files /dev/null and b/croco/models/__pycache__/croco.cpython-312.pyc differ diff --git a/croco/models/__pycache__/dpt_block.cpython-310.pyc b/croco/models/__pycache__/dpt_block.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a7d45e605ad47ba8e41838c442b584ef1fc6c4fa Binary files /dev/null and b/croco/models/__pycache__/dpt_block.cpython-310.pyc differ diff --git a/croco/models/__pycache__/dpt_block.cpython-311.pyc b/croco/models/__pycache__/dpt_block.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..db7d73cfe36a873e90f79f69d2d5f0e7fefca412 Binary files /dev/null and b/croco/models/__pycache__/dpt_block.cpython-311.pyc differ diff --git a/croco/models/__pycache__/dpt_block.cpython-312.pyc b/croco/models/__pycache__/dpt_block.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1db52714a17a41854bffc963272f5d90a689e210 Binary files /dev/null and b/croco/models/__pycache__/dpt_block.cpython-312.pyc differ diff --git a/croco/models/__pycache__/masking.cpython-310.pyc b/croco/models/__pycache__/masking.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0cb3add1716c7a658c0a0fdd3d65e23e04627fdc Binary files /dev/null and b/croco/models/__pycache__/masking.cpython-310.pyc differ diff --git a/croco/models/__pycache__/masking.cpython-311.pyc b/croco/models/__pycache__/masking.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a09f86d8b543acf67a31bcfa8b85b69d4d38d671 Binary files /dev/null and b/croco/models/__pycache__/masking.cpython-311.pyc differ diff --git a/croco/models/__pycache__/masking.cpython-312.pyc b/croco/models/__pycache__/masking.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..03827359d19d0c3cabd972137fa78eb98bd48811 Binary files /dev/null and b/croco/models/__pycache__/masking.cpython-312.pyc differ diff --git a/croco/models/__pycache__/pos_embed.cpython-310.pyc b/croco/models/__pycache__/pos_embed.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7814df62f4c29b22a8a6e16e7056149662015836 Binary files /dev/null and b/croco/models/__pycache__/pos_embed.cpython-310.pyc differ diff --git a/croco/models/__pycache__/pos_embed.cpython-311.pyc b/croco/models/__pycache__/pos_embed.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fb6e6e541b674b92b20be5d75dec8a24b5767c02 Binary files /dev/null and b/croco/models/__pycache__/pos_embed.cpython-311.pyc differ diff --git a/croco/models/__pycache__/pos_embed.cpython-312.pyc b/croco/models/__pycache__/pos_embed.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6393cf8bdfdc73b980554ebe184a2ff40f6042ba Binary files /dev/null and b/croco/models/__pycache__/pos_embed.cpython-312.pyc differ diff --git a/croco/models/blocks.py b/croco/models/blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..aa85a431b44d276e3bba9a33fdfd7097f02bc330 --- /dev/null +++ b/croco/models/blocks.py @@ -0,0 +1,385 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + + +# -------------------------------------------------------- +# Main encoder/decoder blocks +# -------------------------------------------------------- +# References: +# timm +# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py +# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/helpers.py +# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py +# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/mlp.py +# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/patch_embed.py + + +import torch +import torch.nn as nn + +from itertools import repeat +import collections.abc +from torch.nn.functional import scaled_dot_product_attention + + +def _ntuple(n): + def parse(x): + if isinstance(x, collections.abc.Iterable) and not isinstance(x, str): + return x + return tuple(repeat(x, n)) + + return parse + + +to_2tuple = _ntuple(2) + + +def drop_path( + x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True +): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).""" + if drop_prob == 0.0 or not training: + return x + keep_prob = 1 - drop_prob + shape = (x.shape[0],) + (1,) * ( + x.ndim - 1 + ) # work with diff dim tensors, not just 2D ConvNets + random_tensor = x.new_empty(shape).bernoulli_(keep_prob) + if keep_prob > 0.0 and scale_by_keep: + random_tensor.div_(keep_prob) + return x * random_tensor + + +class DropPath(nn.Module): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).""" + + def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True): + super(DropPath, self).__init__() + self.drop_prob = drop_prob + self.scale_by_keep = scale_by_keep + + def forward(self, x): + return drop_path(x, self.drop_prob, self.training, self.scale_by_keep) + + def extra_repr(self): + return f"drop_prob={round(self.drop_prob,3):0.3f}" + + +class Mlp(nn.Module): + """MLP as used in Vision Transformer, MLP-Mixer and related networks""" + + def __init__( + self, + in_features, + hidden_features=None, + out_features=None, + act_layer=nn.GELU, + bias=True, + drop=0.0, + ): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + bias = to_2tuple(bias) + drop_probs = to_2tuple(drop) + + self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0]) + self.act = act_layer() + self.drop1 = nn.Dropout(drop_probs[0]) + self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1]) + self.drop2 = nn.Dropout(drop_probs[1]) + + def forward(self, x): + return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x))))) + + +class Attention(nn.Module): + + def __init__( + self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0 + ): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = head_dim**-0.5 + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + self.rope = rope.float() if rope is not None else None + + def forward(self, x, xpos): + B, N, C = x.shape + + qkv = ( + self.qkv(x) + .reshape(B, N, 3, self.num_heads, C // self.num_heads) + .transpose(1, 3) + ) + q, k, v = [qkv[:, :, i] for i in range(3)] + # q,k,v = qkv.unbind(2) # make torchscript happy (cannot use tensor as tuple) + + q_type = q.dtype + k_type = k.dtype + if self.rope is not None: + q = q.to(torch.float16) + k = k.to(torch.float16) + with torch.autocast(device_type="cuda", enabled=False): + q = self.rope(q, xpos) + k = self.rope(k, xpos) + q = q.to(q_type) + k = k.to(k_type) + + # attn = (q @ k.transpose(-2, -1)) * self.scale + # attn = attn.softmax(dim=-1) + # attn = self.attn_drop(attn) + + # x = (attn @ v).transpose(1, 2).reshape(B, N, C) + # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, N, C) + x = ( + scaled_dot_product_attention( + query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale + ) + .transpose(1, 2) + .reshape(B, N, C) + ) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class Block(nn.Module): + + def __init__( + self, + dim, + num_heads, + mlp_ratio=4.0, + qkv_bias=False, + drop=0.0, + attn_drop=0.0, + drop_path=0.0, + act_layer=nn.GELU, + norm_layer=nn.LayerNorm, + rope=None, + ): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp( + in_features=dim, + hidden_features=mlp_hidden_dim, + act_layer=act_layer, + drop=drop, + ) + + def forward(self, x, xpos): + x = x + self.drop_path(self.attn(self.norm1(x), xpos)) + x = x + self.drop_path(self.mlp(self.norm2(x))) + return x + + +class CrossAttention(nn.Module): + + def __init__( + self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0 + ): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = head_dim**-0.5 + + self.projq = nn.Linear(dim, dim, bias=qkv_bias) + self.projk = nn.Linear(dim, dim, bias=qkv_bias) + self.projv = nn.Linear(dim, dim, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + self.rope = rope.float() if rope is not None else None + + def forward(self, query, key, value, qpos, kpos): + B, Nq, C = query.shape + Nk = key.shape[1] + Nv = value.shape[1] + + q = ( + self.projq(query) + .reshape(B, Nq, self.num_heads, C // self.num_heads) + .permute(0, 2, 1, 3) + ) + k = ( + self.projk(key) + .reshape(B, Nk, self.num_heads, C // self.num_heads) + .permute(0, 2, 1, 3) + ) + v = ( + self.projv(value) + .reshape(B, Nv, self.num_heads, C // self.num_heads) + .permute(0, 2, 1, 3) + ) + + q_type = q.dtype + k_type = k.dtype + if self.rope is not None: + if qpos is not None: + q = q.to(torch.float16) + with torch.autocast(device_type="cuda", enabled=False): + q = self.rope(q, qpos) + q = q.to(q_type) + + if kpos is not None: + k = k.to(torch.float16) + with torch.autocast(device_type="cuda", enabled=False): + k = self.rope(k, kpos) + k = k.to(k_type) + + # attn = (q @ k.transpose(-2, -1)) * self.scale + # attn = attn.softmax(dim=-1) + # attn = self.attn_drop(attn) + + # x = (attn @ v).transpose(1, 2).reshape(B, Nq, C) + + # x = memory_efficient_attention(query=q.permute(0, 2, 1, 3), key=k.permute(0, 2, 1, 3), value=v.permute(0, 2, 1, 3), p=self.attn_drop.p, scale=self.scale).reshape(B, Nq, C) + x = ( + scaled_dot_product_attention( + query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale + ) + .transpose(1, 2) + .reshape(B, Nq, C) + ) + + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class DecoderBlock(nn.Module): + + def __init__( + self, + dim, + num_heads, + mlp_ratio=4.0, + qkv_bias=False, + drop=0.0, + attn_drop=0.0, + drop_path=0.0, + act_layer=nn.GELU, + norm_layer=nn.LayerNorm, + norm_mem=True, + rope=None, + ): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + self.cross_attn = CrossAttention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.norm2 = norm_layer(dim) + self.norm3 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp( + in_features=dim, + hidden_features=mlp_hidden_dim, + act_layer=act_layer, + drop=drop, + ) + self.norm_y = norm_layer(dim) if norm_mem else nn.Identity() + + def forward(self, x, y, xpos, ypos): + x = x + self.drop_path(self.attn(self.norm1(x), xpos)) + y_ = self.norm_y(y) + x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos)) + x = x + self.drop_path(self.mlp(self.norm3(x))) + return x, y + + +# patch embedding +class PositionGetter(object): + """return positions of patches""" + + def __init__(self): + self.cache_positions = {} + + def __call__(self, b, h, w, device): + if not (h, w) in self.cache_positions: + x = torch.arange(w, device=device) + y = torch.arange(h, device=device) + self.cache_positions[h, w] = torch.cartesian_prod(y, x) # (h, w, 2) + pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone() + return pos + + +class PatchEmbed(nn.Module): + """just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed""" + + def __init__( + self, + img_size=224, + patch_size=16, + in_chans=3, + embed_dim=768, + norm_layer=None, + flatten=True, + ): + super().__init__() + img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + self.img_size = img_size + self.patch_size = patch_size + self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1]) + self.num_patches = self.grid_size[0] * self.grid_size[1] + self.flatten = flatten + + self.proj = nn.Conv2d( + in_chans, embed_dim, kernel_size=patch_size, stride=patch_size + ) + self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity() + + self.position_getter = PositionGetter() + + def forward(self, x): + B, C, H, W = x.shape + torch._assert( + H == self.img_size[0], + f"Input image height ({H}) doesn't match model ({self.img_size[0]}).", + ) + torch._assert( + W == self.img_size[1], + f"Input image width ({W}) doesn't match model ({self.img_size[1]}).", + ) + x = self.proj(x) + pos = self.position_getter(B, x.size(2), x.size(3), x.device) + if self.flatten: + x = x.flatten(2).transpose(1, 2) # BCHW -> BNC + x = self.norm(x) + return x, pos + + def _init_weights(self): + w = self.proj.weight.data + torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1])) diff --git a/croco/models/criterion.py b/croco/models/criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..af94f572499c976ad9cfd87d4728b8b517cdfd39 --- /dev/null +++ b/croco/models/criterion.py @@ -0,0 +1,38 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# Criterion to train CroCo +# -------------------------------------------------------- +# References: +# MAE: https://github.com/facebookresearch/mae +# -------------------------------------------------------- + +import torch + + +class MaskedMSE(torch.nn.Module): + + def __init__(self, norm_pix_loss=False, masked=True): + """ + norm_pix_loss: normalize each patch by their pixel mean and variance + masked: compute loss over the masked patches only + """ + super().__init__() + self.norm_pix_loss = norm_pix_loss + self.masked = masked + + def forward(self, pred, mask, target): + + if self.norm_pix_loss: + mean = target.mean(dim=-1, keepdim=True) + var = target.var(dim=-1, keepdim=True) + target = (target - mean) / (var + 1.0e-6) ** 0.5 + + loss = (pred - target) ** 2 + loss = loss.mean(dim=-1) # [N, L], mean loss per patch + if self.masked: + loss = (loss * mask).sum() / mask.sum() # mean loss on masked patches + else: + loss = loss.mean() # mean loss + return loss diff --git a/croco/models/croco.py b/croco/models/croco.py new file mode 100644 index 0000000000000000000000000000000000000000..64b2410e9b52ab34bc66f1d7d768d0e91c8cf30b --- /dev/null +++ b/croco/models/croco.py @@ -0,0 +1,330 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + + +# -------------------------------------------------------- +# CroCo model during pretraining +# -------------------------------------------------------- + + +import torch +import torch.nn as nn + +torch.backends.cuda.matmul.allow_tf32 = True # for gpu >= Ampere and pytorch >= 1.12 +from functools import partial + +from models.blocks import Block, DecoderBlock, PatchEmbed +from models.pos_embed import get_2d_sincos_pos_embed, RoPE2D +from models.masking import RandomMask + +from transformers import PretrainedConfig +from transformers import PreTrainedModel + + +class CrocoConfig(PretrainedConfig): + model_type = "croco" + + def __init__( + self, + img_size=224, # input image size + patch_size=16, # patch_size + mask_ratio=0.9, # ratios of masked tokens + enc_embed_dim=768, # encoder feature dimension + enc_depth=12, # encoder depth + enc_num_heads=12, # encoder number of heads in the transformer block + dec_embed_dim=512, # decoder feature dimension + dec_depth=8, # decoder depth + dec_num_heads=16, # decoder number of heads in the transformer block + mlp_ratio=4, + norm_layer=partial(nn.LayerNorm, eps=1e-6), + norm_im2_in_dec=True, # whether to apply normalization of the 'memory' = (second image) in the decoder + pos_embed="cosine", # positional embedding (either cosine or RoPE100) + ): + super().__init__() + self.img_size = img_size + self.patch_size = patch_size + self.mask_ratio = mask_ratio + self.enc_embed_dim = enc_embed_dim + self.enc_depth = enc_depth + self.enc_num_heads = enc_num_heads + self.dec_embed_dim = dec_embed_dim + self.dec_depth = dec_depth + self.dec_num_heads = dec_num_heads + self.mlp_ratio = mlp_ratio + self.norm_layer = norm_layer + self.norm_im2_in_dec = norm_im2_in_dec + self.pos_embed = pos_embed + + +class CroCoNet(PreTrainedModel): + + config_class = CrocoConfig + base_model_prefix = "croco" + + def __init__(self, config: CrocoConfig): + + super().__init__(config) + + # patch embeddings (with initialization done as in MAE) + self._set_patch_embed(config.img_size, config.patch_size, config.enc_embed_dim) + + # mask generations + self._set_mask_generator(self.patch_embed.num_patches, config.mask_ratio) + + self.pos_embed = config.pos_embed + if config.pos_embed == "cosine": + # positional embedding of the encoder + enc_pos_embed = get_2d_sincos_pos_embed( + config.enc_embed_dim, + int(self.patch_embed.num_patches**0.5), + n_cls_token=0, + ) + self.register_buffer( + "enc_pos_embed", torch.from_numpy(enc_pos_embed).float() + ) + # positional embedding of the decoder + dec_pos_embed = get_2d_sincos_pos_embed( + config.dec_embed_dim, + int(self.patch_embed.num_patches**0.5), + n_cls_token=0, + ) + self.register_buffer( + "dec_pos_embed", torch.from_numpy(dec_pos_embed).float() + ) + # pos embedding in each block + self.rope = None # nothing for cosine + elif config.pos_embed.startswith("RoPE"): # eg RoPE100 + self.enc_pos_embed = None # nothing to add in the encoder with RoPE + self.dec_pos_embed = None # nothing to add in the decoder with RoPE + if RoPE2D is None: + raise ImportError( + "Cannot find cuRoPE2D, please install it following the README instructions" + ) + freq = float(config.pos_embed[len("RoPE") :]) + self.rope = RoPE2D(freq=freq) + else: + raise NotImplementedError("Unknown pos_embed " + config.pos_embed) + + # transformer for the encoder + self.enc_depth = config.enc_depth + self.enc_embed_dim = config.enc_embed_dim + self.enc_blocks = nn.ModuleList( + [ + Block( + config.enc_embed_dim, + config.enc_num_heads, + config.mlp_ratio, + qkv_bias=True, + norm_layer=config.norm_layer, + rope=self.rope, + ) + for i in range(config.enc_depth) + ] + ) + self.enc_norm = config.norm_layer(config.enc_embed_dim) + + # masked tokens + # self._set_mask_token(config.dec_embed_dim) + self.mask_token = None + + # decoder + self._set_decoder( + config.enc_embed_dim, + config.dec_embed_dim, + config.dec_num_heads, + config.dec_depth, + config.mlp_ratio, + config.norm_layer, + config.norm_im2_in_dec, + ) + + # prediction head + self._set_prediction_head(config.dec_embed_dim, config.patch_size) + + # initializer weights + self.initialize_weights() + + def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768): + self.patch_embed = PatchEmbed(img_size, patch_size, 3, enc_embed_dim) + + def _set_mask_generator(self, num_patches, mask_ratio): + self.mask_generator = RandomMask(num_patches, mask_ratio) + + def _set_mask_token(self, dec_embed_dim): + self.mask_token = nn.Parameter(torch.zeros(1, 1, dec_embed_dim)) + + def _set_decoder( + self, + enc_embed_dim, + dec_embed_dim, + dec_num_heads, + dec_depth, + mlp_ratio, + norm_layer, + norm_im2_in_dec, + ): + self.dec_depth = dec_depth + self.dec_embed_dim = dec_embed_dim + # transfer from encoder to decoder + self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True) + # transformer for the decoder + self.dec_blocks = nn.ModuleList( + [ + DecoderBlock( + dec_embed_dim, + dec_num_heads, + mlp_ratio=mlp_ratio, + qkv_bias=True, + norm_layer=norm_layer, + norm_mem=norm_im2_in_dec, + rope=self.rope, + ) + for i in range(dec_depth) + ] + ) + # final norm layer + self.dec_norm = norm_layer(dec_embed_dim) + + def _set_prediction_head(self, dec_embed_dim, patch_size): + self.prediction_head = nn.Linear(dec_embed_dim, patch_size**2 * 3, bias=True) + + def initialize_weights(self): + # patch embed + self.patch_embed._init_weights() + # mask tokens + if self.mask_token is not None: + torch.nn.init.normal_(self.mask_token, std=0.02) + # linears and layer norms + self.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + # we use xavier_uniform following official JAX ViT: + torch.nn.init.xavier_uniform_(m.weight) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + def _encode_image(self, image, do_mask=False, return_all_blocks=False): + """ + image has B x 3 x img_size x img_size + do_mask: whether to perform masking or not + return_all_blocks: if True, return the features at the end of every block + instead of just the features from the last block (eg for some prediction heads) + """ + # embed the image into patches (x has size B x Npatches x C) + # and get position if each return patch (pos has size B x Npatches x 2) + x, pos = self.patch_embed(image) + # add positional embedding without cls token + if self.enc_pos_embed is not None: + x = x + self.enc_pos_embed[None, ...] + # apply masking + B, N, C = x.size() + if do_mask: + masks = self.mask_generator(x) + x = x[~masks].view(B, -1, C) + posvis = pos[~masks].view(B, -1, 2) + else: + B, N, C = x.size() + masks = torch.zeros((B, N), dtype=bool) + posvis = pos + # now apply the transformer encoder and normalization + if return_all_blocks: + out = [] + for blk in self.enc_blocks: + x = blk(x, posvis) + out.append(x) + out[-1] = self.enc_norm(out[-1]) + return out, pos, masks + else: + for blk in self.enc_blocks: + x = blk(x, posvis) + x = self.enc_norm(x) + return x, pos, masks + + def _decoder(self, feat1, pos1, masks1, feat2, pos2, return_all_blocks=False): + """ + return_all_blocks: if True, return the features at the end of every block + instead of just the features from the last block (eg for some prediction heads) + + masks1 can be None => assume image1 fully visible + """ + # encoder to decoder layer + visf1 = self.decoder_embed(feat1) + f2 = self.decoder_embed(feat2) + # append masked tokens to the sequence + B, Nenc, C = visf1.size() + if masks1 is None: # downstreams + f1_ = visf1 + else: # pretraining + Ntotal = masks1.size(1) + f1_ = self.mask_token.repeat(B, Ntotal, 1).to(dtype=visf1.dtype) + f1_[~masks1] = visf1.view(B * Nenc, C) + # add positional embedding + if self.dec_pos_embed is not None: + f1_ = f1_ + self.dec_pos_embed + f2 = f2 + self.dec_pos_embed + # apply Transformer blocks + out = f1_ + out2 = f2 + if return_all_blocks: + _out, out = out, [] + for blk in self.dec_blocks: + _out, out2 = blk(_out, out2, pos1, pos2) + out.append(_out) + out[-1] = self.dec_norm(out[-1]) + else: + for blk in self.dec_blocks: + out, out2 = blk(out, out2, pos1, pos2) + out = self.dec_norm(out) + return out + + def patchify(self, imgs): + """ + imgs: (B, 3, H, W) + x: (B, L, patch_size**2 *3) + """ + p = self.patch_embed.patch_size[0] + assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0 + + h = w = imgs.shape[2] // p + x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p)) + x = torch.einsum("nchpwq->nhwpqc", x) + x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3)) + + return x + + def unpatchify(self, x, channels=3): + """ + x: (N, L, patch_size**2 *channels) + imgs: (N, 3, H, W) + """ + patch_size = self.patch_embed.patch_size[0] + h = w = int(x.shape[1] ** 0.5) + assert h * w == x.shape[1] + x = x.reshape(shape=(x.shape[0], h, w, patch_size, patch_size, channels)) + x = torch.einsum("nhwpqc->nchpwq", x) + imgs = x.reshape(shape=(x.shape[0], channels, h * patch_size, h * patch_size)) + return imgs + + # def forward(self, img1, img2): + # """ + # img1: tensor of size B x 3 x img_size x img_size + # img2: tensor of size B x 3 x img_size x img_size + + # out will be B x N x (3*patch_size*patch_size) + # masks are also returned as B x N just in case + # """ + # # encoder of the masked first image + # feat1, pos1, mask1 = self._encode_image(img1, do_mask=True) + # # encoder of the second image + # feat2, pos2, _ = self._encode_image(img2, do_mask=False) + # # decoder + # decfeat = self._decoder(feat1, pos1, mask1, feat2, pos2) + # # prediction head + # out = self.prediction_head(decfeat) + # # get target + # target = self.patchify(img1) + # return out, mask1, target diff --git a/croco/models/croco_downstream.py b/croco/models/croco_downstream.py new file mode 100644 index 0000000000000000000000000000000000000000..cd59dca45d403c16d60610640b4156b151f46c9b --- /dev/null +++ b/croco/models/croco_downstream.py @@ -0,0 +1,141 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# CroCo model for downstream tasks +# -------------------------------------------------------- + +import torch + +from .croco import CroCoNet + + +def croco_args_from_ckpt(ckpt): + if "croco_kwargs" in ckpt: # CroCo v2 released models + return ckpt["croco_kwargs"] + elif "args" in ckpt and hasattr( + ckpt["args"], "model" + ): # pretrained using the official code release + s = ckpt[ + "args" + ].model # eg "CroCoNet(enc_embed_dim=1024, enc_num_heads=16, enc_depth=24)" + assert s.startswith("CroCoNet(") + return eval( + "dict" + s[len("CroCoNet") :] + ) # transform it into the string of a dictionary and evaluate it + else: # CroCo v1 released models + return dict() + + +class CroCoDownstreamMonocularEncoder(CroCoNet): + + def __init__(self, head, **kwargs): + """Build network for monocular downstream task, only using the encoder. + It takes an extra argument head, that is called with the features + and a dictionary img_info containing 'width' and 'height' keys + The head is setup with the croconet arguments in this init function + NOTE: It works by *calling super().__init__() but with redefined setters + + """ + super(CroCoDownstreamMonocularEncoder, self).__init__(**kwargs) + head.setup(self) + self.head = head + + def _set_mask_generator(self, *args, **kwargs): + """No mask generator""" + return + + def _set_mask_token(self, *args, **kwargs): + """No mask token""" + self.mask_token = None + return + + def _set_decoder(self, *args, **kwargs): + """No decoder""" + return + + def _set_prediction_head(self, *args, **kwargs): + """No 'prediction head' for downstream tasks.""" + return + + def forward(self, img): + """ + img if of size batch_size x 3 x h x w + """ + B, C, H, W = img.size() + img_info = {"height": H, "width": W} + need_all_layers = ( + hasattr(self.head, "return_all_blocks") and self.head.return_all_blocks + ) + out, _, _ = self._encode_image( + img, do_mask=False, return_all_blocks=need_all_layers + ) + return self.head(out, img_info) + + +class CroCoDownstreamBinocular(CroCoNet): + + def __init__(self, head, **kwargs): + """Build network for binocular downstream task + It takes an extra argument head, that is called with the features + and a dictionary img_info containing 'width' and 'height' keys + The head is setup with the croconet arguments in this init function + """ + super(CroCoDownstreamBinocular, self).__init__(**kwargs) + head.setup(self) + self.head = head + + def _set_mask_generator(self, *args, **kwargs): + """No mask generator""" + return + + def _set_mask_token(self, *args, **kwargs): + """No mask token""" + self.mask_token = None + return + + def _set_prediction_head(self, *args, **kwargs): + """No prediction head for downstream tasks, define your own head""" + return + + def encode_image_pairs(self, img1, img2, return_all_blocks=False): + """run encoder for a pair of images + it is actually ~5% faster to concatenate the images along the batch dimension + than to encode them separately + """ + ## the two commented lines below is the naive version with separate encoding + # out, pos, _ = self._encode_image(img1, do_mask=False, return_all_blocks=return_all_blocks) + # out2, pos2, _ = self._encode_image(img2, do_mask=False, return_all_blocks=False) + ## and now the faster version + out, pos, _ = self._encode_image( + torch.cat((img1, img2), dim=0), + do_mask=False, + return_all_blocks=return_all_blocks, + ) + if return_all_blocks: + out, out2 = list(map(list, zip(*[o.chunk(2, dim=0) for o in out]))) + out2 = out2[-1] + else: + out, out2 = out.chunk(2, dim=0) + pos, pos2 = pos.chunk(2, dim=0) + return out, out2, pos, pos2 + + def forward(self, img1, img2): + B, C, H, W = img1.size() + img_info = {"height": H, "width": W} + return_all_blocks = ( + hasattr(self.head, "return_all_blocks") and self.head.return_all_blocks + ) + out, out2, pos, pos2 = self.encode_image_pairs( + img1, img2, return_all_blocks=return_all_blocks + ) + if return_all_blocks: + decout = self._decoder( + out[-1], pos, None, out2, pos2, return_all_blocks=return_all_blocks + ) + decout = out + decout + else: + decout = self._decoder( + out, pos, None, out2, pos2, return_all_blocks=return_all_blocks + ) + return self.head(decout, img_info) diff --git a/croco/models/curope/__init__.py b/croco/models/curope/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..25e3d48a162760260826080f6366838e83e26878 --- /dev/null +++ b/croco/models/curope/__init__.py @@ -0,0 +1,4 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +from .curope2d import cuRoPE2D diff --git a/croco/models/curope/__pycache__/__init__.cpython-310.pyc b/croco/models/curope/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9c86d1e5c48709032291b3e99e3d7d34081597dd Binary files /dev/null and b/croco/models/curope/__pycache__/__init__.cpython-310.pyc differ diff --git a/croco/models/curope/__pycache__/__init__.cpython-311.pyc b/croco/models/curope/__pycache__/__init__.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ab1863b30f94225f248efcc9f469fa7549601bf6 Binary files /dev/null and b/croco/models/curope/__pycache__/__init__.cpython-311.pyc differ diff --git a/croco/models/curope/__pycache__/__init__.cpython-312.pyc b/croco/models/curope/__pycache__/__init__.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4993f6a00dd3c9ccf81a827d9b804a6003daa09b Binary files /dev/null and b/croco/models/curope/__pycache__/__init__.cpython-312.pyc differ diff --git a/croco/models/curope/__pycache__/curope2d.cpython-310.pyc b/croco/models/curope/__pycache__/curope2d.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bc761cc88644bd180e9949c53d07faffb8ad2905 Binary files /dev/null and b/croco/models/curope/__pycache__/curope2d.cpython-310.pyc differ diff --git a/croco/models/curope/__pycache__/curope2d.cpython-311.pyc b/croco/models/curope/__pycache__/curope2d.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6071022c45facd7b093e712dd5cb9973603b6efe Binary files /dev/null and b/croco/models/curope/__pycache__/curope2d.cpython-311.pyc differ diff --git a/croco/models/curope/__pycache__/curope2d.cpython-312.pyc b/croco/models/curope/__pycache__/curope2d.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..75b76cdfa78f1f6b90cb41485e1b06aff5aab707 Binary files /dev/null and b/croco/models/curope/__pycache__/curope2d.cpython-312.pyc differ diff --git a/croco/models/curope/curope.cpp b/croco/models/curope/curope.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8fe9058e05aa1bf3f37b0d970edc7312bc68455b --- /dev/null +++ b/croco/models/curope/curope.cpp @@ -0,0 +1,69 @@ +/* + Copyright (C) 2022-present Naver Corporation. All rights reserved. + Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +*/ + +#include + +// forward declaration +void rope_2d_cuda( torch::Tensor tokens, const torch::Tensor pos, const float base, const float fwd ); + +void rope_2d_cpu( torch::Tensor tokens, const torch::Tensor positions, const float base, const float fwd ) +{ + const int B = tokens.size(0); + const int N = tokens.size(1); + const int H = tokens.size(2); + const int D = tokens.size(3) / 4; + + auto tok = tokens.accessor(); + auto pos = positions.accessor(); + + for (int b = 0; b < B; b++) { + for (int x = 0; x < 2; x++) { // y and then x (2d) + for (int n = 0; n < N; n++) { + + // grab the token position + const int p = pos[b][n][x]; + + for (int h = 0; h < H; h++) { + for (int d = 0; d < D; d++) { + // grab the two values + float u = tok[b][n][h][d+0+x*2*D]; + float v = tok[b][n][h][d+D+x*2*D]; + + // grab the cos,sin + const float inv_freq = fwd * p / powf(base, d/float(D)); + float c = cosf(inv_freq); + float s = sinf(inv_freq); + + // write the result + tok[b][n][h][d+0+x*2*D] = u*c - v*s; + tok[b][n][h][d+D+x*2*D] = v*c + u*s; + } + } + } + } + } +} + +void rope_2d( torch::Tensor tokens, // B,N,H,D + const torch::Tensor positions, // B,N,2 + const float base, + const float fwd ) +{ + TORCH_CHECK(tokens.dim() == 4, "tokens must have 4 dimensions"); + TORCH_CHECK(positions.dim() == 3, "positions must have 3 dimensions"); + TORCH_CHECK(tokens.size(0) == positions.size(0), "batch size differs between tokens & positions"); + TORCH_CHECK(tokens.size(1) == positions.size(1), "seq_length differs between tokens & positions"); + TORCH_CHECK(positions.size(2) == 2, "positions.shape[2] must be equal to 2"); + TORCH_CHECK(tokens.is_cuda() == positions.is_cuda(), "tokens and positions are not on the same device" ); + + if (tokens.is_cuda()) + rope_2d_cuda( tokens, positions, base, fwd ); + else + rope_2d_cpu( tokens, positions, base, fwd ); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("rope_2d", &rope_2d, "RoPE 2d forward/backward"); +} diff --git a/croco/models/curope/curope2d.py b/croco/models/curope/curope2d.py new file mode 100644 index 0000000000000000000000000000000000000000..7e0345c31bd3925be91dde5b9cfc64432f7bf516 --- /dev/null +++ b/croco/models/curope/curope2d.py @@ -0,0 +1,40 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +import torch + +try: + import curope as _kernels # run `python setup.py install` +except ModuleNotFoundError: + from . import curope as _kernels # run `python setup.py build_ext --inplace` + + +class cuRoPE2D_func(torch.autograd.Function): + + @staticmethod + def forward(ctx, tokens, positions, base, F0=1): + ctx.save_for_backward(positions) + ctx.saved_base = base + ctx.saved_F0 = F0 + # tokens = tokens.clone() # uncomment this if inplace doesn't work + _kernels.rope_2d(tokens, positions, base, F0) + ctx.mark_dirty(tokens) + return tokens + + @staticmethod + def backward(ctx, grad_res): + positions, base, F0 = ctx.saved_tensors[0], ctx.saved_base, ctx.saved_F0 + _kernels.rope_2d(grad_res, positions, base, -F0) + ctx.mark_dirty(grad_res) + return grad_res, None, None, None + + +class cuRoPE2D(torch.nn.Module): + def __init__(self, freq=100.0, F0=1.0): + super().__init__() + self.base = freq + self.F0 = F0 + + def forward(self, tokens, positions): + cuRoPE2D_func.apply(tokens.transpose(1, 2), positions, self.base, self.F0) + return tokens diff --git a/croco/models/curope/kernels.cu b/croco/models/curope/kernels.cu new file mode 100644 index 0000000000000000000000000000000000000000..7156cd1bb935cb1f0be45e58add53f9c21505c20 --- /dev/null +++ b/croco/models/curope/kernels.cu @@ -0,0 +1,108 @@ +/* + Copyright (C) 2022-present Naver Corporation. All rights reserved. + Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +*/ + +#include +#include +#include +#include + +#define CHECK_CUDA(tensor) {\ + TORCH_CHECK((tensor).is_cuda(), #tensor " is not in cuda memory"); \ + TORCH_CHECK((tensor).is_contiguous(), #tensor " is not contiguous"); } +void CHECK_KERNEL() {auto error = cudaGetLastError(); TORCH_CHECK( error == cudaSuccess, cudaGetErrorString(error));} + + +template < typename scalar_t > +__global__ void rope_2d_cuda_kernel( + //scalar_t* __restrict__ tokens, + torch::PackedTensorAccessor32 tokens, + const int64_t* __restrict__ pos, + const float base, + const float fwd ) + // const int N, const int H, const int D ) +{ + // tokens shape = (B, N, H, D) + const int N = tokens.size(1); + const int H = tokens.size(2); + const int D = tokens.size(3); + + // each block update a single token, for all heads + // each thread takes care of a single output + extern __shared__ float shared[]; + float* shared_inv_freq = shared + D; + + const int b = blockIdx.x / N; + const int n = blockIdx.x % N; + + const int Q = D / 4; + // one token = [0..Q : Q..2Q : 2Q..3Q : 3Q..D] + // u_Y v_Y u_X v_X + + // shared memory: first, compute inv_freq + if (threadIdx.x < Q) + shared_inv_freq[threadIdx.x] = fwd / powf(base, threadIdx.x/float(Q)); + __syncthreads(); + + // start of X or Y part + const int X = threadIdx.x < D/2 ? 0 : 1; + const int m = (X*D/2) + (threadIdx.x % Q); // index of u_Y or u_X + + // grab the cos,sin appropriate for me + const float freq = pos[blockIdx.x*2+X] * shared_inv_freq[threadIdx.x % Q]; + const float cos = cosf(freq); + const float sin = sinf(freq); + /* + float* shared_cos_sin = shared + D + D/4; + if ((threadIdx.x % (D/2)) < Q) + shared_cos_sin[m+0] = cosf(freq); + else + shared_cos_sin[m+Q] = sinf(freq); + __syncthreads(); + const float cos = shared_cos_sin[m+0]; + const float sin = shared_cos_sin[m+Q]; + */ + + for (int h = 0; h < H; h++) + { + // then, load all the token for this head in shared memory + shared[threadIdx.x] = tokens[b][n][h][threadIdx.x]; + __syncthreads(); + + const float u = shared[m]; + const float v = shared[m+Q]; + + // write output + if ((threadIdx.x % (D/2)) < Q) + tokens[b][n][h][threadIdx.x] = u*cos - v*sin; + else + tokens[b][n][h][threadIdx.x] = v*cos + u*sin; + } +} + +void rope_2d_cuda( torch::Tensor tokens, const torch::Tensor pos, const float base, const float fwd ) +{ + const int B = tokens.size(0); // batch size + const int N = tokens.size(1); // sequence length + const int H = tokens.size(2); // number of heads + const int D = tokens.size(3); // dimension per head + + TORCH_CHECK(tokens.stride(3) == 1 && tokens.stride(2) == D, "tokens are not contiguous"); + TORCH_CHECK(pos.is_contiguous(), "positions are not contiguous"); + TORCH_CHECK(pos.size(0) == B && pos.size(1) == N && pos.size(2) == 2, "bad pos.shape"); + TORCH_CHECK(D % 4 == 0, "token dim must be multiple of 4"); + + // one block for each layer, one thread per local-max + const int THREADS_PER_BLOCK = D; + const int N_BLOCKS = B * N; // each block takes care of H*D values + const int SHARED_MEM = sizeof(float) * (D + D/4); + + AT_DISPATCH_FLOATING_TYPES_AND_HALF(tokens.type(), "rope_2d_cuda", ([&] { + rope_2d_cuda_kernel <<>> ( + //tokens.data_ptr(), + tokens.packed_accessor32(), + pos.data_ptr(), + base, fwd); //, N, H, D ); + })); +} diff --git a/croco/models/curope/setup.py b/croco/models/curope/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..02ddb0912370a67a49fd2bb91164cf2f1da8648e --- /dev/null +++ b/croco/models/curope/setup.py @@ -0,0 +1,34 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +from setuptools import setup +from torch import cuda +from torch.utils.cpp_extension import BuildExtension, CUDAExtension + +# compile for all possible CUDA architectures +all_cuda_archs = cuda.get_gencode_flags().replace("compute=", "arch=").split() +# alternatively, you can list cuda archs that you want, eg: +# all_cuda_archs = [ +# '-gencode', 'arch=compute_70,code=sm_70', +# '-gencode', 'arch=compute_75,code=sm_75', +# '-gencode', 'arch=compute_80,code=sm_80', +# '-gencode', 'arch=compute_86,code=sm_86' +# ] + +setup( + name="curope", + ext_modules=[ + CUDAExtension( + name="curope", + sources=[ + "curope.cpp", + "kernels.cu", + ], + extra_compile_args=dict( + nvcc=["-O3", "--ptxas-options=-v", "--use_fast_math"] + all_cuda_archs, + cxx=["-O3"], + ), + ) + ], + cmdclass={"build_ext": BuildExtension}, +) diff --git a/croco/models/dpt_block.py b/croco/models/dpt_block.py new file mode 100644 index 0000000000000000000000000000000000000000..b470d91c9c86af8f3b3947e3abcf96d49ab3e06d --- /dev/null +++ b/croco/models/dpt_block.py @@ -0,0 +1,513 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# DPT head for ViTs +# -------------------------------------------------------- +# References: +# https://github.com/isl-org/DPT +# https://github.com/EPFL-VILAB/MultiMAE/blob/main/multimae/output_adapters.py + +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange, repeat +from typing import Union, Tuple, Iterable, List, Optional, Dict + + +def pair(t): + return t if isinstance(t, tuple) else (t, t) + + +def make_scratch(in_shape, out_shape, groups=1, expand=False): + scratch = nn.Module() + + out_shape1 = out_shape + out_shape2 = out_shape + out_shape3 = out_shape + out_shape4 = out_shape + if expand == True: + out_shape1 = out_shape + out_shape2 = out_shape * 2 + out_shape3 = out_shape * 4 + out_shape4 = out_shape * 8 + + scratch.layer1_rn = nn.Conv2d( + in_shape[0], + out_shape1, + kernel_size=3, + stride=1, + padding=1, + bias=False, + groups=groups, + ) + scratch.layer2_rn = nn.Conv2d( + in_shape[1], + out_shape2, + kernel_size=3, + stride=1, + padding=1, + bias=False, + groups=groups, + ) + scratch.layer3_rn = nn.Conv2d( + in_shape[2], + out_shape3, + kernel_size=3, + stride=1, + padding=1, + bias=False, + groups=groups, + ) + scratch.layer4_rn = nn.Conv2d( + in_shape[3], + out_shape4, + kernel_size=3, + stride=1, + padding=1, + bias=False, + groups=groups, + ) + + scratch.layer_rn = nn.ModuleList( + [ + scratch.layer1_rn, + scratch.layer2_rn, + scratch.layer3_rn, + scratch.layer4_rn, + ] + ) + + return scratch + + +class ResidualConvUnit_custom(nn.Module): + """Residual convolution module.""" + + def __init__(self, features, activation, bn): + """Init. + Args: + features (int): number of features + """ + super().__init__() + + self.bn = bn + + self.groups = 1 + + self.conv1 = nn.Conv2d( + features, + features, + kernel_size=3, + stride=1, + padding=1, + bias=not self.bn, + groups=self.groups, + ) + + self.conv2 = nn.Conv2d( + features, + features, + kernel_size=3, + stride=1, + padding=1, + bias=not self.bn, + groups=self.groups, + ) + + if self.bn == True: + self.bn1 = nn.BatchNorm2d(features) + self.bn2 = nn.BatchNorm2d(features) + + self.activation = activation + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, x): + """Forward pass. + Args: + x (tensor): input + Returns: + tensor: output + """ + + out = self.activation(x) + out = self.conv1(out) + if self.bn == True: + out = self.bn1(out) + + out = self.activation(out) + out = self.conv2(out) + if self.bn == True: + out = self.bn2(out) + + if self.groups > 1: + out = self.conv_merge(out) + + return self.skip_add.add(out, x) + + +class FeatureFusionBlock_custom(nn.Module): + """Feature fusion block.""" + + def __init__( + self, + features, + activation, + deconv=False, + bn=False, + expand=False, + align_corners=True, + width_ratio=1, + ): + """Init. + Args: + features (int): number of features + """ + super(FeatureFusionBlock_custom, self).__init__() + self.width_ratio = width_ratio + + self.deconv = deconv + self.align_corners = align_corners + + self.groups = 1 + + self.expand = expand + out_features = features + if self.expand == True: + out_features = features // 2 + + self.out_conv = nn.Conv2d( + features, + out_features, + kernel_size=1, + stride=1, + padding=0, + bias=True, + groups=1, + ) + + self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn) + self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn) + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, *xs): + """Forward pass. + Returns: + tensor: output + """ + output = xs[0] + + if len(xs) == 2: + res = self.resConfUnit1(xs[1]) + if self.width_ratio != 1: + res = F.interpolate( + res, size=(output.shape[2], output.shape[3]), mode="bilinear" + ) + + output = self.skip_add.add(output, res) + # output += res + + output = self.resConfUnit2(output) + + if self.width_ratio != 1: + # and output.shape[3] < self.width_ratio * output.shape[2] + # size=(image.shape[]) + if (output.shape[3] / output.shape[2]) < (2 / 3) * self.width_ratio: + shape = 3 * output.shape[3] + else: + shape = int(self.width_ratio * 2 * output.shape[2]) + output = F.interpolate( + output, size=(2 * output.shape[2], shape), mode="bilinear" + ) + else: + output = nn.functional.interpolate( + output, + scale_factor=2, + mode="bilinear", + align_corners=self.align_corners, + ) + output = self.out_conv(output) + return output + + +def make_fusion_block(features, use_bn, width_ratio=1): + return FeatureFusionBlock_custom( + features, + nn.ReLU(False), + deconv=False, + bn=use_bn, + expand=False, + align_corners=True, + width_ratio=width_ratio, + ) + + +class Interpolate(nn.Module): + """Interpolation module.""" + + def __init__(self, scale_factor, mode, align_corners=False): + """Init. + Args: + scale_factor (float): scaling + mode (str): interpolation mode + """ + super(Interpolate, self).__init__() + + self.interp = nn.functional.interpolate + self.scale_factor = scale_factor + self.mode = mode + self.align_corners = align_corners + + def forward(self, x): + """Forward pass. + Args: + x (tensor): input + Returns: + tensor: interpolated data + """ + + x = self.interp( + x, + scale_factor=self.scale_factor, + mode=self.mode, + align_corners=self.align_corners, + ) + + return x + + +class DPTOutputAdapter(nn.Module): + """DPT output adapter. + + :param num_cahnnels: Number of output channels + :param stride_level: tride level compared to the full-sized image. + E.g. 4 for 1/4th the size of the image. + :param patch_size_full: Int or tuple of the patch size over the full image size. + Patch size for smaller inputs will be computed accordingly. + :param hooks: Index of intermediate layers + :param layer_dims: Dimension of intermediate layers + :param feature_dim: Feature dimension + :param last_dim: out_channels/in_channels for the last two Conv2d when head_type == regression + :param use_bn: If set to True, activates batch norm + :param dim_tokens_enc: Dimension of tokens coming from encoder + """ + + def __init__( + self, + num_channels: int = 1, + stride_level: int = 1, + patch_size: Union[int, Tuple[int, int]] = 16, + main_tasks: Iterable[str] = ("rgb",), + hooks: List[int] = [2, 5, 8, 11], + layer_dims: List[int] = [96, 192, 384, 768], + feature_dim: int = 256, + last_dim: int = 32, + use_bn: bool = False, + dim_tokens_enc: Optional[int] = None, + head_type: str = "regression", + output_width_ratio=1, + **kwargs + ): + super().__init__() + self.num_channels = num_channels + self.stride_level = stride_level + self.patch_size = pair(patch_size) + self.main_tasks = main_tasks + self.hooks = hooks + self.layer_dims = layer_dims + self.feature_dim = feature_dim + self.dim_tokens_enc = ( + dim_tokens_enc * len(self.main_tasks) + if dim_tokens_enc is not None + else None + ) + self.head_type = head_type + + # Actual patch height and width, taking into account stride of input + self.P_H = max(1, self.patch_size[0] // stride_level) + self.P_W = max(1, self.patch_size[1] // stride_level) + + self.scratch = make_scratch(layer_dims, feature_dim, groups=1, expand=False) + + self.scratch.refinenet1 = make_fusion_block( + feature_dim, use_bn, output_width_ratio + ) + self.scratch.refinenet2 = make_fusion_block( + feature_dim, use_bn, output_width_ratio + ) + self.scratch.refinenet3 = make_fusion_block( + feature_dim, use_bn, output_width_ratio + ) + self.scratch.refinenet4 = make_fusion_block( + feature_dim, use_bn, output_width_ratio + ) + + if self.head_type == "regression": + # The "DPTDepthModel" head + self.head = nn.Sequential( + nn.Conv2d( + feature_dim, feature_dim // 2, kernel_size=3, stride=1, padding=1 + ), + Interpolate(scale_factor=2, mode="bilinear", align_corners=True), + nn.Conv2d( + feature_dim // 2, last_dim, kernel_size=3, stride=1, padding=1 + ), + nn.ReLU(True), + nn.Conv2d( + last_dim, self.num_channels, kernel_size=1, stride=1, padding=0 + ), + ) + elif self.head_type == "semseg": + # The "DPTSegmentationModel" head + self.head = nn.Sequential( + nn.Conv2d( + feature_dim, feature_dim, kernel_size=3, padding=1, bias=False + ), + nn.BatchNorm2d(feature_dim) if use_bn else nn.Identity(), + nn.ReLU(True), + nn.Dropout(0.1, False), + nn.Conv2d(feature_dim, self.num_channels, kernel_size=1), + Interpolate(scale_factor=2, mode="bilinear", align_corners=True), + ) + else: + raise ValueError('DPT head_type must be "regression" or "semseg".') + + if self.dim_tokens_enc is not None: + self.init(dim_tokens_enc=dim_tokens_enc) + + def init(self, dim_tokens_enc=768): + """ + Initialize parts of decoder that are dependent on dimension of encoder tokens. + Should be called when setting up MultiMAE. + + :param dim_tokens_enc: Dimension of tokens coming from encoder + """ + # print(dim_tokens_enc) + + # Set up activation postprocessing layers + if isinstance(dim_tokens_enc, int): + dim_tokens_enc = 4 * [dim_tokens_enc] + + self.dim_tokens_enc = [dt * len(self.main_tasks) for dt in dim_tokens_enc] + + self.act_1_postprocess = nn.Sequential( + nn.Conv2d( + in_channels=self.dim_tokens_enc[0], + out_channels=self.layer_dims[0], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=self.layer_dims[0], + out_channels=self.layer_dims[0], + kernel_size=4, + stride=4, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + self.act_2_postprocess = nn.Sequential( + nn.Conv2d( + in_channels=self.dim_tokens_enc[1], + out_channels=self.layer_dims[1], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=self.layer_dims[1], + out_channels=self.layer_dims[1], + kernel_size=2, + stride=2, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + self.act_3_postprocess = nn.Sequential( + nn.Conv2d( + in_channels=self.dim_tokens_enc[2], + out_channels=self.layer_dims[2], + kernel_size=1, + stride=1, + padding=0, + ) + ) + + self.act_4_postprocess = nn.Sequential( + nn.Conv2d( + in_channels=self.dim_tokens_enc[3], + out_channels=self.layer_dims[3], + kernel_size=1, + stride=1, + padding=0, + ), + nn.Conv2d( + in_channels=self.layer_dims[3], + out_channels=self.layer_dims[3], + kernel_size=3, + stride=2, + padding=1, + ), + ) + + self.act_postprocess = nn.ModuleList( + [ + self.act_1_postprocess, + self.act_2_postprocess, + self.act_3_postprocess, + self.act_4_postprocess, + ] + ) + + def adapt_tokens(self, encoder_tokens): + # Adapt tokens + x = [] + x.append(encoder_tokens[:, :]) + x = torch.cat(x, dim=-1) + return x + + def forward(self, encoder_tokens: List[torch.Tensor], image_size): + # input_info: Dict): + assert ( + self.dim_tokens_enc is not None + ), "Need to call init(dim_tokens_enc) function first" + H, W = image_size + + # Number of patches in height and width + N_H = H // (self.stride_level * self.P_H) + N_W = W // (self.stride_level * self.P_W) + + # Hook decoder onto 4 layers from specified ViT layers + layers = [encoder_tokens[hook] for hook in self.hooks] + + # Extract only task-relevant tokens and ignore global tokens. + layers = [self.adapt_tokens(l) for l in layers] + + # Reshape tokens to spatial representation + layers = [ + rearrange(l, "b (nh nw) c -> b c nh nw", nh=N_H, nw=N_W) for l in layers + ] + + layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)] + # Project layers to chosen feature dim + layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)] + + # Fuse layers using refinement stages + path_4 = self.scratch.refinenet4(layers[3]) + path_3 = self.scratch.refinenet3(path_4, layers[2]) + path_2 = self.scratch.refinenet2(path_3, layers[1]) + path_1 = self.scratch.refinenet1(path_2, layers[0]) + + # Output head + out = self.head(path_1) + + return out diff --git a/croco/models/head_downstream.py b/croco/models/head_downstream.py new file mode 100644 index 0000000000000000000000000000000000000000..384afcbd6ac9d4b5729c0219dd8534b5123d2b17 --- /dev/null +++ b/croco/models/head_downstream.py @@ -0,0 +1,83 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Heads for downstream tasks +# -------------------------------------------------------- + +""" +A head is a module where the __init__ defines only the head hyperparameters. +A method setup(croconet) takes a CroCoNet and set all layers according to the head and croconet attributes. +The forward takes the features as well as a dictionary img_info containing the keys 'width' and 'height' +""" + +import torch +import torch.nn as nn +from .dpt_block import DPTOutputAdapter + + +class PixelwiseTaskWithDPT(nn.Module): + """DPT module for CroCo. + by default, hooks_idx will be equal to: + * for encoder-only: 4 equally spread layers + * for encoder+decoder: last encoder + 3 equally spread layers of the decoder + """ + + def __init__( + self, + *, + hooks_idx=None, + layer_dims=[96, 192, 384, 768], + output_width_ratio=1, + num_channels=1, + postprocess=None, + **kwargs, + ): + super(PixelwiseTaskWithDPT, self).__init__() + self.return_all_blocks = True # backbone needs to return all layers + self.postprocess = postprocess + self.output_width_ratio = output_width_ratio + self.num_channels = num_channels + self.hooks_idx = hooks_idx + self.layer_dims = layer_dims + + def setup(self, croconet): + dpt_args = { + "output_width_ratio": self.output_width_ratio, + "num_channels": self.num_channels, + } + if self.hooks_idx is None: + if hasattr(croconet, "dec_blocks"): # encoder + decoder + step = {8: 3, 12: 4, 24: 8}[croconet.dec_depth] + hooks_idx = [ + croconet.dec_depth + croconet.enc_depth - 1 - i * step + for i in range(3, -1, -1) + ] + else: # encoder only + step = croconet.enc_depth // 4 + hooks_idx = [ + croconet.enc_depth - 1 - i * step for i in range(3, -1, -1) + ] + self.hooks_idx = hooks_idx + print( + f" PixelwiseTaskWithDPT: automatically setting hook_idxs={self.hooks_idx}" + ) + dpt_args["hooks"] = self.hooks_idx + dpt_args["layer_dims"] = self.layer_dims + self.dpt = DPTOutputAdapter(**dpt_args) + dim_tokens = [ + ( + croconet.enc_embed_dim + if hook < croconet.enc_depth + else croconet.dec_embed_dim + ) + for hook in self.hooks_idx + ] + dpt_init_args = {"dim_tokens_enc": dim_tokens} + self.dpt.init(**dpt_init_args) + + def forward(self, x, img_info): + out = self.dpt(x, image_size=(img_info["height"], img_info["width"])) + if self.postprocess: + out = self.postprocess(out) + return out diff --git a/croco/models/masking.py b/croco/models/masking.py new file mode 100644 index 0000000000000000000000000000000000000000..ae18f927ae82e4075c2246ce722007c69a4da344 --- /dev/null +++ b/croco/models/masking.py @@ -0,0 +1,26 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + + +# -------------------------------------------------------- +# Masking utils +# -------------------------------------------------------- + +import torch +import torch.nn as nn + + +class RandomMask(nn.Module): + """ + random masking + """ + + def __init__(self, num_patches, mask_ratio): + super().__init__() + self.num_patches = num_patches + self.num_mask = int(mask_ratio * self.num_patches) + + def __call__(self, x): + noise = torch.rand(x.size(0), self.num_patches, device=x.device) + argsort = torch.argsort(noise, dim=1) + return argsort < self.num_mask diff --git a/croco/models/pos_embed.py b/croco/models/pos_embed.py new file mode 100644 index 0000000000000000000000000000000000000000..0f76e4d5be2222d446f14d7fb24a047b686cb328 --- /dev/null +++ b/croco/models/pos_embed.py @@ -0,0 +1,179 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + + +# -------------------------------------------------------- +# Position embedding utils +# -------------------------------------------------------- + + +import numpy as np + +import torch + + +# -------------------------------------------------------- +# 2D sine-cosine position embedding +# References: +# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py +# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py +# MoCo v3: https://github.com/facebookresearch/moco-v3 +# -------------------------------------------------------- +def get_2d_sincos_pos_embed(embed_dim, grid_size, n_cls_token=0): + """ + grid_size: int of the grid height and width + return: + pos_embed: [grid_size*grid_size, embed_dim] or [n_cls_token+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) + """ + grid_h = np.arange(grid_size, dtype=np.float32) + grid_w = np.arange(grid_size, dtype=np.float32) + grid = np.meshgrid(grid_w, grid_h) # here w goes first + grid = np.stack(grid, axis=0) + + grid = grid.reshape([2, 1, grid_size, grid_size]) + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + if n_cls_token > 0: + pos_embed = np.concatenate( + [np.zeros([n_cls_token, embed_dim]), pos_embed], axis=0 + ) + return pos_embed + + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + assert embed_dim % 2 == 0 + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) + return emb + + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + embed_dim: output dimension for each position + pos: a list of positions to be encoded: size (M,) + out: (M, D) + """ + assert embed_dim % 2 == 0 + omega = np.arange(embed_dim // 2, dtype=float) + omega /= embed_dim / 2.0 + omega = 1.0 / 10000**omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product + + emb_sin = np.sin(out) # (M, D/2) + emb_cos = np.cos(out) # (M, D/2) + + emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) + return emb + + +# -------------------------------------------------------- +# Interpolate position embeddings for high-resolution +# References: +# MAE: https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py +# DeiT: https://github.com/facebookresearch/deit +# -------------------------------------------------------- +def interpolate_pos_embed(model, checkpoint_model): + if "pos_embed" in checkpoint_model: + pos_embed_checkpoint = checkpoint_model["pos_embed"] + embedding_size = pos_embed_checkpoint.shape[-1] + num_patches = model.patch_embed.num_patches + num_extra_tokens = model.pos_embed.shape[-2] - num_patches + # height (== width) for the checkpoint position embedding + orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5) + # height (== width) for the new position embedding + new_size = int(num_patches**0.5) + # class_token and dist_token are kept unchanged + if orig_size != new_size: + print( + "Position interpolate from %dx%d to %dx%d" + % (orig_size, orig_size, new_size, new_size) + ) + extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] + # only the position tokens are interpolated + pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] + pos_tokens = pos_tokens.reshape( + -1, orig_size, orig_size, embedding_size + ).permute(0, 3, 1, 2) + pos_tokens = torch.nn.functional.interpolate( + pos_tokens, + size=(new_size, new_size), + mode="bicubic", + align_corners=False, + ) + pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2) + new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) + checkpoint_model["pos_embed"] = new_pos_embed + + +# ---------------------------------------------------------- +# RoPE2D: RoPE implementation in 2D +# ---------------------------------------------------------- + +try: + from models.curope import cuRoPE2D + + RoPE2D = cuRoPE2D +except ImportError: + print( + "Warning, cannot find cuda-compiled version of RoPE2D, using a slow pytorch version instead" + ) + + class RoPE2D(torch.nn.Module): + + def __init__(self, freq=100.0, F0=1.0): + super().__init__() + self.base = freq + self.F0 = F0 + self.cache = {} + + def get_cos_sin(self, D, seq_len, device, dtype): + if (D, seq_len, device, dtype) not in self.cache: + inv_freq = 1.0 / ( + self.base ** (torch.arange(0, D, 2).float().to(device) / D) + ) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.einsum("i,j->ij", t, inv_freq).to(dtype) + freqs = torch.cat((freqs, freqs), dim=-1) + cos = freqs.cos() # (Seq, Dim) + sin = freqs.sin() + self.cache[D, seq_len, device, dtype] = (cos, sin) + return self.cache[D, seq_len, device, dtype] + + @staticmethod + def rotate_half(x): + x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + def apply_rope1d(self, tokens, pos1d, cos, sin): + assert pos1d.ndim == 2 + cos = torch.nn.functional.embedding(pos1d, cos)[:, None, :, :] + sin = torch.nn.functional.embedding(pos1d, sin)[:, None, :, :] + return (tokens * cos) + (self.rotate_half(tokens) * sin) + + def forward(self, tokens, positions): + """ + input: + * tokens: batch_size x nheads x ntokens x dim + * positions: batch_size x ntokens x 2 (y and x position of each token) + output: + * tokens after appplying RoPE2D (batch_size x nheads x ntokens x dim) + """ + assert ( + tokens.size(3) % 2 == 0 + ), "number of dimensions should be a multiple of two" + D = tokens.size(3) // 2 + assert positions.ndim == 3 and positions.shape[-1] == 2 # Batch, Seq, 2 + cos, sin = self.get_cos_sin( + D, int(positions.max()) + 1, tokens.device, tokens.dtype + ) + # split features into two along the feature dimension, and apply rope1d on each half + y, x = tokens.chunk(2, dim=-1) + y = self.apply_rope1d(y, positions[:, :, 0], cos, sin) + x = self.apply_rope1d(x, positions[:, :, 1], cos, sin) + tokens = torch.cat((y, x), dim=-1) + return tokens diff --git a/croco/pretrain.py b/croco/pretrain.py new file mode 100644 index 0000000000000000000000000000000000000000..fef4ff2a0b7cb865a68741ac0e76d43d50ee4659 --- /dev/null +++ b/croco/pretrain.py @@ -0,0 +1,391 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# Pre-training CroCo +# -------------------------------------------------------- +# References: +# MAE: https://github.com/facebookresearch/mae +# DeiT: https://github.com/facebookresearch/deit +# BEiT: https://github.com/microsoft/unilm/tree/master/beit +# -------------------------------------------------------- +import argparse +import datetime +import json +import numpy as np +import os +import sys +import time +import math +from pathlib import Path +from typing import Iterable + +import torch +import torch.distributed as dist +import torch.backends.cudnn as cudnn +from torch.utils.tensorboard import SummaryWriter +import torchvision.transforms as transforms +import torchvision.datasets as datasets + +import utils.misc as misc +from utils.misc import NativeScalerWithGradNormCount as NativeScaler +from models.croco import CroCoNet +from models.criterion import MaskedMSE +from datasets.pairs_dataset import PairsDataset + + +def get_args_parser(): + parser = argparse.ArgumentParser("CroCo pre-training", add_help=False) + # model and criterion + parser.add_argument( + "--model", + default="CroCoNet()", + type=str, + help="string containing the model to build", + ) + parser.add_argument( + "--norm_pix_loss", + default=1, + choices=[0, 1], + help="apply per-patch mean/std normalization before applying the loss", + ) + # dataset + parser.add_argument( + "--dataset", default="habitat_release", type=str, help="training set" + ) + parser.add_argument( + "--transforms", default="crop224+acolor", type=str, help="transforms to apply" + ) # in the paper, we also use some homography and rotation, but find later that they were not useful or even harmful + # training + parser.add_argument("--seed", default=0, type=int, help="Random seed") + parser.add_argument( + "--batch_size", + default=64, + type=int, + help="Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus", + ) + parser.add_argument( + "--epochs", + default=800, + type=int, + help="Maximum number of epochs for the scheduler", + ) + parser.add_argument( + "--max_epoch", default=400, type=int, help="Stop training at this epoch" + ) + parser.add_argument( + "--accum_iter", + default=1, + type=int, + help="Accumulate gradient iterations (for increasing the effective batch size under memory constraints)", + ) + parser.add_argument( + "--weight_decay", type=float, default=0.05, help="weight decay (default: 0.05)" + ) + parser.add_argument( + "--lr", + type=float, + default=None, + metavar="LR", + help="learning rate (absolute lr)", + ) + parser.add_argument( + "--blr", + type=float, + default=1.5e-4, + metavar="LR", + help="base learning rate: absolute_lr = base_lr * total_batch_size / 256", + ) + parser.add_argument( + "--min_lr", + type=float, + default=0.0, + metavar="LR", + help="lower lr bound for cyclic schedulers that hit 0", + ) + parser.add_argument( + "--warmup_epochs", type=int, default=40, metavar="N", help="epochs to warmup LR" + ) + parser.add_argument( + "--amp", + type=int, + default=1, + choices=[0, 1], + help="Use Automatic Mixed Precision for pretraining", + ) + # others + parser.add_argument("--num_workers", default=8, type=int) + parser.add_argument( + "--world_size", default=1, type=int, help="number of distributed processes" + ) + parser.add_argument("--local_rank", default=-1, type=int) + parser.add_argument( + "--dist_url", default="env://", help="url used to set up distributed training" + ) + parser.add_argument( + "--save_freq", + default=1, + type=int, + help="frequence (number of epochs) to save checkpoint in checkpoint-last.pth", + ) + parser.add_argument( + "--keep_freq", + default=20, + type=int, + help="frequence (number of epochs) to save checkpoint in checkpoint-%d.pth", + ) + parser.add_argument( + "--print_freq", + default=20, + type=int, + help="frequence (number of iterations) to print infos while training", + ) + # paths + parser.add_argument( + "--output_dir", + default="./output/", + type=str, + help="path where to save the output", + ) + parser.add_argument( + "--data_dir", default="./data/", type=str, help="path where data are stored" + ) + return parser + + +def main(args): + misc.init_distributed_mode(args) + global_rank = misc.get_rank() + world_size = misc.get_world_size() + + print("output_dir: " + args.output_dir) + if args.output_dir: + Path(args.output_dir).mkdir(parents=True, exist_ok=True) + + # auto resume + last_ckpt_fname = os.path.join(args.output_dir, f"checkpoint-last.pth") + args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None + + print("job dir: {}".format(os.path.dirname(os.path.realpath(__file__)))) + print("{}".format(args).replace(", ", ",\n")) + + device = "cuda" if torch.cuda.is_available() else "cpu" + device = torch.device(device) + + # fix the seed + seed = args.seed + misc.get_rank() + torch.manual_seed(seed) + np.random.seed(seed) + + cudnn.benchmark = True + + ## training dataset and loader + print( + "Building dataset for {:s} with transforms {:s}".format( + args.dataset, args.transforms + ) + ) + dataset = PairsDataset(args.dataset, trfs=args.transforms, data_dir=args.data_dir) + if world_size > 1: + sampler_train = torch.utils.data.DistributedSampler( + dataset, num_replicas=world_size, rank=global_rank, shuffle=True + ) + print("Sampler_train = %s" % str(sampler_train)) + else: + sampler_train = torch.utils.data.RandomSampler(dataset) + data_loader_train = torch.utils.data.DataLoader( + dataset, + sampler=sampler_train, + batch_size=args.batch_size, + num_workers=args.num_workers, + pin_memory=True, + drop_last=True, + ) + + ## model + print("Loading model: {:s}".format(args.model)) + model = eval(args.model) + print( + "Loading criterion: MaskedMSE(norm_pix_loss={:s})".format( + str(bool(args.norm_pix_loss)) + ) + ) + criterion = MaskedMSE(norm_pix_loss=bool(args.norm_pix_loss)) + + model.to(device) + model_without_ddp = model + print("Model = %s" % str(model_without_ddp)) + + eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size() + if args.lr is None: # only base_lr is specified + args.lr = args.blr * eff_batch_size / 256 + print("base lr: %.2e" % (args.lr * 256 / eff_batch_size)) + print("actual lr: %.2e" % args.lr) + print("accumulate grad iterations: %d" % args.accum_iter) + print("effective batch size: %d" % eff_batch_size) + + if args.distributed: + model = torch.nn.parallel.DistributedDataParallel( + model, device_ids=[args.gpu], find_unused_parameters=True, static_graph=True + ) + model_without_ddp = model.module + + param_groups = misc.get_parameter_groups( + model_without_ddp, args.weight_decay + ) # following timm: set wd as 0 for bias and norm layers + optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95)) + print(optimizer) + loss_scaler = NativeScaler() + + misc.load_model( + args=args, + model_without_ddp=model_without_ddp, + optimizer=optimizer, + loss_scaler=loss_scaler, + ) + + if global_rank == 0 and args.output_dir is not None: + log_writer = SummaryWriter(log_dir=args.output_dir) + else: + log_writer = None + + print(f"Start training until {args.max_epoch} epochs") + start_time = time.time() + for epoch in range(args.start_epoch, args.max_epoch): + if world_size > 1: + data_loader_train.sampler.set_epoch(epoch) + + train_stats = train_one_epoch( + model, + criterion, + data_loader_train, + optimizer, + device, + epoch, + loss_scaler, + log_writer=log_writer, + args=args, + ) + + if args.output_dir and epoch % args.save_freq == 0: + misc.save_model( + args=args, + model_without_ddp=model_without_ddp, + optimizer=optimizer, + loss_scaler=loss_scaler, + epoch=epoch, + fname="last", + ) + + if ( + args.output_dir + and (epoch % args.keep_freq == 0 or epoch + 1 == args.max_epoch) + and (epoch > 0 or args.max_epoch == 1) + ): + misc.save_model( + args=args, + model_without_ddp=model_without_ddp, + optimizer=optimizer, + loss_scaler=loss_scaler, + epoch=epoch, + ) + + log_stats = { + **{f"train_{k}": v for k, v in train_stats.items()}, + "epoch": epoch, + } + + if args.output_dir and misc.is_main_process(): + if log_writer is not None: + log_writer.flush() + with open( + os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8" + ) as f: + f.write(json.dumps(log_stats) + "\n") + + total_time = time.time() - start_time + total_time_str = str(datetime.timedelta(seconds=int(total_time))) + print("Training time {}".format(total_time_str)) + + +def train_one_epoch( + model: torch.nn.Module, + criterion: torch.nn.Module, + data_loader: Iterable, + optimizer: torch.optim.Optimizer, + device: torch.device, + epoch: int, + loss_scaler, + log_writer=None, + args=None, +): + model.train(True) + metric_logger = misc.MetricLogger(delimiter=" ") + metric_logger.add_meter("lr", misc.SmoothedValue(window_size=1, fmt="{value:.6f}")) + header = "Epoch: [{}]".format(epoch) + accum_iter = args.accum_iter + + optimizer.zero_grad() + + if log_writer is not None: + print("log_dir: {}".format(log_writer.log_dir)) + + for data_iter_step, (image1, image2) in enumerate( + metric_logger.log_every(data_loader, args.print_freq, header) + ): + + # we use a per iteration lr scheduler + if data_iter_step % accum_iter == 0: + misc.adjust_learning_rate( + optimizer, data_iter_step / len(data_loader) + epoch, args + ) + + image1 = image1.to(device, non_blocking=True) + image2 = image2.to(device, non_blocking=True) + with torch.cuda.amp.autocast(enabled=bool(args.amp)): + out, mask, target = model(image1, image2) + loss = criterion(out, mask, target) + + loss_value = loss.item() + + if not math.isfinite(loss_value): + print("Loss is {}, stopping training".format(loss_value)) + sys.exit(1) + + loss /= accum_iter + loss_scaler( + loss, + optimizer, + parameters=model.parameters(), + update_grad=(data_iter_step + 1) % accum_iter == 0, + ) + if (data_iter_step + 1) % accum_iter == 0: + optimizer.zero_grad() + + torch.cuda.synchronize() + + metric_logger.update(loss=loss_value) + + lr = optimizer.param_groups[0]["lr"] + metric_logger.update(lr=lr) + + loss_value_reduce = misc.all_reduce_mean(loss_value) + if ( + log_writer is not None + and ((data_iter_step + 1) % (accum_iter * args.print_freq)) == 0 + ): + # x-axis is based on epoch_1000x in the tensorboard, calibrating differences curves when batch size changes + epoch_1000x = int((data_iter_step / len(data_loader) + epoch) * 1000) + log_writer.add_scalar("train_loss", loss_value_reduce, epoch_1000x) + log_writer.add_scalar("lr", lr, epoch_1000x) + + # gather the stats from all processes + metric_logger.synchronize_between_processes() + print("Averaged stats:", metric_logger) + return {k: meter.global_avg for k, meter in metric_logger.meters.items()} + + +if __name__ == "__main__": + args = get_args_parser() + args = args.parse_args() + main(args) diff --git a/croco/stereoflow/README.MD b/croco/stereoflow/README.MD new file mode 100644 index 0000000000000000000000000000000000000000..81595380fadd274b523e0cf77921b1b65cbedb34 --- /dev/null +++ b/croco/stereoflow/README.MD @@ -0,0 +1,318 @@ +## CroCo-Stereo and CroCo-Flow + +This README explains how to use CroCo-Stereo and CroCo-Flow as well as how they were trained. +All commands should be launched from the root directory. + +### Simple inference example + +We provide a simple inference exemple for CroCo-Stereo and CroCo-Flow in the Totebook `croco-stereo-flow-demo.ipynb`. +Before running it, please download the trained models with: +``` +bash stereoflow/download_model.sh crocostereo.pth +bash stereoflow/download_model.sh crocoflow.pth +``` + +### Prepare data for training or evaluation + +Put the datasets used for training/evaluation in `./data/stereoflow` (or update the paths at the top of `stereoflow/datasets_stereo.py` and `stereoflow/datasets_flow.py`). +Please find below on the file structure should look for each dataset: +
+FlyingChairs + +``` +./data/stereoflow/FlyingChairs/ +└───chairs_split.txt +└───data/ + └─── ... +``` +
+ +
+MPI-Sintel + +``` +./data/stereoflow/MPI-Sintel/ +└───training/ +│ └───clean/ +│ └───final/ +│ └───flow/ +└───test/ + └───clean/ + └───final/ +``` +
+ +
+SceneFlow (including FlyingThings) + +``` +./data/stereoflow/SceneFlow/ +└───Driving/ +│ └───disparity/ +│ └───frames_cleanpass/ +│ └───frames_finalpass/ +└───FlyingThings/ +│ └───disparity/ +│ └───frames_cleanpass/ +│ └───frames_finalpass/ +│ └───optical_flow/ +└───Monkaa/ + └───disparity/ + └───frames_cleanpass/ + └───frames_finalpass/ +``` +
+ +
+TartanAir + +``` +./data/stereoflow/TartanAir/ +└───abandonedfactory/ +│ └───.../ +└───abandonedfactory_night/ +│ └───.../ +└───.../ +``` +
+ +
+Booster + +``` +./data/stereoflow/booster_gt/ +└───train/ + └───balanced/ + └───Bathroom/ + └───Bedroom/ + └───... +``` +
+ +
+CREStereo + +``` +./data/stereoflow/crenet_stereo_trainset/ +└───stereo_trainset/ + └───crestereo/ + └───hole/ + └───reflective/ + └───shapenet/ + └───tree/ +``` +
+ +
+ETH3D Two-view Low-res + +``` +./data/stereoflow/eth3d_lowres/ +└───test/ +│ └───lakeside_1l/ +│ └───... +└───train/ +│ └───delivery_area_1l/ +│ └───... +└───train_gt/ + └───delivery_area_1l/ + └───... +``` +
+ +
+KITTI 2012 + +``` +./data/stereoflow/kitti-stereo-2012/ +└───testing/ +│ └───colored_0/ +│ └───colored_1/ +└───training/ + └───colored_0/ + └───colored_1/ + └───disp_occ/ + └───flow_occ/ +``` +
+ +
+KITTI 2015 + +``` +./data/stereoflow/kitti-stereo-2015/ +└───testing/ +│ └───image_2/ +│ └───image_3/ +└───training/ + └───image_2/ + └───image_3/ + └───disp_occ_0/ + └───flow_occ/ +``` +
+ +
+Middlebury + +``` +./data/stereoflow/middlebury +└───2005/ +│ └───train/ +│ └───Art/ +│ └───... +└───2006/ +│ └───Aloe/ +│ └───Baby1/ +│ └───... +└───2014/ +│ └───Adirondack-imperfect/ +│ └───Adirondack-perfect/ +│ └───... +└───2021/ +│ └───data/ +│ └───artroom1/ +│ └───artroom2/ +│ └───... +└───MiddEval3_F/ + └───test/ + │ └───Australia/ + │ └───... + └───train/ + └───Adirondack/ + └───... +``` +
+ +
+Spring + +``` +./data/stereoflow/spring/ +└───test/ +│ └───0003/ +│ └───... +└───train/ + └───0001/ + └───... +``` +
+ + +### CroCo-Stereo + +##### Main model + +The main training of CroCo-Stereo was performed on a series of datasets, and it was used as it for Middlebury v3 benchmark. + +``` +# Download the model +bash stereoflow/download_model.sh crocostereo.pth +# Middlebury v3 submission +python stereoflow/test.py --model stereoflow_models/crocostereo.pth --dataset "MdEval3('all_full')" --save submission --tile_overlap 0.9 +# Training command that was used, using checkpoint-last.pth +python -u stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('train')+50*Md05('train')+50*Md06('train')+50*Md14('train')+50*Md21('train')+50*MdEval3('train_full')+Booster('train_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 6 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main/ +# or it can be launched on multiple gpus (while maintaining the effective batch size), e.g. on 3 gpus: +torchrun --nproc_per_node 3 stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('train')+50*Md05('train')+50*Md06('train')+50*Md14('train')+50*Md21('train')+50*MdEval3('train_full')+Booster('train_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 2 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main/ +``` + +For evaluation of validation set, we also provide the model trained on the `subtrain` subset of the training sets. + +``` +# Download the model +bash stereoflow/download_model.sh crocostereo_subtrain.pth +# Evaluation on validation sets +python stereoflow/test.py --model stereoflow_models/crocostereo_subtrain.pth --dataset "MdEval3('subval_full')+ETH3DLowRes('subval')+SceneFlow('test_finalpass')+SceneFlow('test_cleanpass')" --save metrics --tile_overlap 0.9 +# Training command that was used (same as above but on subtrain, using checkpoint-best.pth), can also be launched on multiple gpus +python -u stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('subtrain')+50*Md05('subtrain')+50*Md06('subtrain')+50*Md14('subtrain')+50*Md21('subtrain')+50*MdEval3('subtrain_full')+Booster('subtrain_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 6 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main_subtrain/ +``` + +##### Other models + +
+ Model for ETH3D + The model used for the submission on ETH3D is trained with the same command but using an unbounded Laplacian loss. + + # Download the model + bash stereoflow/download_model.sh crocostereo_eth3d.pth + # ETH3D submission + python stereoflow/test.py --model stereoflow_models/crocostereo_eth3d.pth --dataset "ETH3DLowRes('all')" --save submission --tile_overlap 0.9 + # Training command that was used + python -u stereoflow/train.py stereo --criterion "LaplacianLoss()" --tile_conf_mode conf_expbeta3 --dataset "CREStereo('train')+SceneFlow('train_allpass')+30*ETH3DLowRes('train')+50*Md05('train')+50*Md06('train')+50*Md14('train')+50*Md21('train')+50*MdEval3('train_full')+Booster('train_balanced')" --val_dataset "SceneFlow('test1of100_finalpass')+SceneFlow('test1of100_cleanpass')+ETH3DLowRes('subval')+Md05('subval')+Md06('subval')+Md14('subval')+Md21('subval')+MdEval3('subval_full')+Booster('subval_balanced')" --lr 3e-5 --batch_size 6 --epochs 32 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocostereo/main_eth3d/ + +
+ +
+ Main model finetuned on Kitti + + # Download the model + bash stereoflow/download_model.sh crocostereo_finetune_kitti.pth + # Kitti submission + python stereoflow/test.py --model stereoflow_models/crocostereo_finetune_kitti.pth --dataset "Kitti15('test')" --save submission --tile_overlap 0.9 + # Training that was used + python -u stereoflow/train.py stereo --crop 352 1216 --criterion "LaplacianLossBounded2()" --dataset "Kitti12('train')+Kitti15('train')" --lr 3e-5 --batch_size 1 --accum_iter 6 --epochs 20 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocostereo.pth --output_dir xps/crocostereo/finetune_kitti/ --save_every 5 +
+ +
+ Main model finetuned on Spring + + # Download the model + bash stereoflow/download_model.sh crocostereo_finetune_spring.pth + # Spring submission + python stereoflow/test.py --model stereoflow_models/crocostereo_finetune_spring.pth --dataset "Spring('test')" --save submission --tile_overlap 0.9 + # Training command that was used + python -u stereoflow/train.py stereo --criterion "LaplacianLossBounded2()" --dataset "Spring('train')" --lr 3e-5 --batch_size 6 --epochs 8 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocostereo.pth --output_dir xps/crocostereo/finetune_spring/ +
+ +
+ Smaller models + To train CroCo-Stereo with smaller CroCo pretrained models, simply replace the --pretrained argument. To download the smaller CroCo-Stereo models based on CroCo v2 pretraining with ViT-Base encoder and Small encoder, use bash stereoflow/download_model.sh crocostereo_subtrain_vitb_smalldecoder.pth, and for the model with a ViT-Base encoder and a Base decoder, use bash stereoflow/download_model.sh crocostereo_subtrain_vitb_basedecoder.pth. +
+ + +### CroCo-Flow + +##### Main model + +The main training of CroCo-Flow was performed on the FlyingThings, FlyingChairs, MPI-Sintel and TartanAir datasets. +It was used for our submission to the MPI-Sintel benchmark. + +``` +# Download the model +bash stereoflow/download_model.sh crocoflow.pth +# Evaluation +python stereoflow/test.py --model stereoflow_models/crocoflow.pth --dataset "MPISintel('subval_cleanpass')+MPISintel('subval_finalpass')" --save metrics --tile_overlap 0.9 +# Sintel submission +python stereoflow/test.py --model stereoflow_models/crocoflow.pth --dataset "MPISintel('test_allpass')" --save submission --tile_overlap 0.9 +# Training command that was used, with checkpoint-best.pth +python -u stereoflow/train.py flow --criterion "LaplacianLossBounded()" --dataset "40*MPISintel('subtrain_cleanpass')+40*MPISintel('subtrain_finalpass')+4*FlyingThings('train_allpass')+4*FlyingChairs('train')+TartanAir('train')" --val_dataset "MPISintel('subval_cleanpass')+MPISintel('subval_finalpass')" --lr 2e-5 --batch_size 8 --epochs 240 --img_per_epoch 30000 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --output_dir xps/crocoflow/main/ +``` + +##### Other models + +
+ Main model finetuned on Kitti + + # Download the model + bash stereoflow/download_model.sh crocoflow_finetune_kitti.pth + # Kitti submission + python stereoflow/test.py --model stereoflow_models/crocoflow_finetune_kitti.pth --dataset "Kitti15('test')" --save submission --tile_overlap 0.99 + # Training that was used, with checkpoint-last.pth + python -u stereoflow/train.py flow --crop 352 1216 --criterion "LaplacianLossBounded()" --dataset "Kitti15('train')+Kitti12('train')" --lr 2e-5 --batch_size 1 --accum_iter 8 --epochs 150 --save_every 5 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocoflow.pth --output_dir xps/crocoflow/finetune_kitti/ +
+ +
+ Main model finetuned on Spring + + # Download the model + bash stereoflow/download_model.sh crocoflow_finetune_spring.pth + # Spring submission + python stereoflow/test.py --model stereoflow_models/crocoflow_finetune_spring.pth --dataset "Spring('test')" --save submission --tile_overlap 0.9 + # Training command that was used, with checkpoint-last.pth + python -u stereoflow/train.py flow --criterion "LaplacianLossBounded()" --dataset "Spring('train')" --lr 2e-5 --batch_size 8 --epochs 12 --pretrained pretrained_models/CroCo_V2_ViTLarge_BaseDecoder.pth --start_from stereoflow_models/crocoflow.pth --output_dir xps/crocoflow/finetune_spring/ +
+ +
+ Smaller models + To train CroCo-Flow with smaller CroCo pretrained models, simply replace the --pretrained argument. To download the smaller CroCo-Flow models based on CroCo v2 pretraining with ViT-Base encoder and Small encoder, use bash stereoflow/download_model.sh crocoflow_vitb_smalldecoder.pth, and for the model with a ViT-Base encoder and a Base decoder, use bash stereoflow/download_model.sh crocoflow_vitb_basedecoder.pth. +
diff --git a/croco/stereoflow/augmentor.py b/croco/stereoflow/augmentor.py new file mode 100644 index 0000000000000000000000000000000000000000..aac818df45d927ac383a41978ff92dc5f2899890 --- /dev/null +++ b/croco/stereoflow/augmentor.py @@ -0,0 +1,396 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Data augmentation for training stereo and flow +# -------------------------------------------------------- + +# References +# https://github.com/autonomousvision/unimatch/blob/master/dataloader/stereo/transforms.py +# https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/transforms.py + + +import numpy as np +import random +from PIL import Image + +import cv2 + +cv2.setNumThreads(0) +cv2.ocl.setUseOpenCL(False) + +import torch +from torchvision.transforms import ColorJitter +import torchvision.transforms.functional as FF + + +class StereoAugmentor(object): + + def __init__( + self, + crop_size, + scale_prob=0.5, + scale_xonly=True, + lhth=800.0, + lminscale=0.0, + lmaxscale=1.0, + hminscale=-0.2, + hmaxscale=0.4, + scale_interp_nearest=True, + rightjitterprob=0.5, + v_flip_prob=0.5, + color_aug_asym=True, + color_choice_prob=0.5, + ): + self.crop_size = crop_size + self.scale_prob = scale_prob + self.scale_xonly = scale_xonly + self.lhth = lhth + self.lminscale = lminscale + self.lmaxscale = lmaxscale + self.hminscale = hminscale + self.hmaxscale = hmaxscale + self.scale_interp_nearest = scale_interp_nearest + self.rightjitterprob = rightjitterprob + self.v_flip_prob = v_flip_prob + self.color_aug_asym = color_aug_asym + self.color_choice_prob = color_choice_prob + + def _random_scale(self, img1, img2, disp): + ch, cw = self.crop_size + h, w = img1.shape[:2] + if self.scale_prob > 0.0 and np.random.rand() < self.scale_prob: + min_scale, max_scale = ( + (self.lminscale, self.lmaxscale) + if min(h, w) < self.lhth + else (self.hminscale, self.hmaxscale) + ) + scale_x = 2.0 ** np.random.uniform(min_scale, max_scale) + scale_x = np.clip(scale_x, (cw + 8) / float(w), None) + scale_y = 1.0 + if not self.scale_xonly: + scale_y = scale_x + scale_y = np.clip(scale_y, (ch + 8) / float(h), None) + img1 = cv2.resize( + img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR + ) + img2 = cv2.resize( + img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR + ) + disp = ( + cv2.resize( + disp, + None, + fx=scale_x, + fy=scale_y, + interpolation=( + cv2.INTER_LINEAR + if not self.scale_interp_nearest + else cv2.INTER_NEAREST + ), + ) + * scale_x + ) + else: # check if we need to resize to be able to crop + h, w = img1.shape[:2] + clip_scale = (cw + 8) / float(w) + if clip_scale > 1.0: + scale_x = clip_scale + scale_y = scale_x if not self.scale_xonly else 1.0 + img1 = cv2.resize( + img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR + ) + img2 = cv2.resize( + img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR + ) + disp = ( + cv2.resize( + disp, + None, + fx=scale_x, + fy=scale_y, + interpolation=( + cv2.INTER_LINEAR + if not self.scale_interp_nearest + else cv2.INTER_NEAREST + ), + ) + * scale_x + ) + return img1, img2, disp + + def _random_crop(self, img1, img2, disp): + h, w = img1.shape[:2] + ch, cw = self.crop_size + assert ch <= h and cw <= w, (img1.shape, h, w, ch, cw) + offset_x = np.random.randint(w - cw + 1) + offset_y = np.random.randint(h - ch + 1) + img1 = img1[offset_y : offset_y + ch, offset_x : offset_x + cw] + img2 = img2[offset_y : offset_y + ch, offset_x : offset_x + cw] + disp = disp[offset_y : offset_y + ch, offset_x : offset_x + cw] + return img1, img2, disp + + def _random_vflip(self, img1, img2, disp): + # vertical flip + if self.v_flip_prob > 0 and np.random.rand() < self.v_flip_prob: + img1 = np.copy(np.flipud(img1)) + img2 = np.copy(np.flipud(img2)) + disp = np.copy(np.flipud(disp)) + return img1, img2, disp + + def _random_rotate_shift_right(self, img2): + if self.rightjitterprob > 0.0 and np.random.rand() < self.rightjitterprob: + angle, pixel = 0.1, 2 + px = np.random.uniform(-pixel, pixel) + ag = np.random.uniform(-angle, angle) + image_center = ( + np.random.uniform(0, img2.shape[0]), + np.random.uniform(0, img2.shape[1]), + ) + rot_mat = cv2.getRotationMatrix2D(image_center, ag, 1.0) + img2 = cv2.warpAffine( + img2, rot_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR + ) + trans_mat = np.float32([[1, 0, 0], [0, 1, px]]) + img2 = cv2.warpAffine( + img2, trans_mat, img2.shape[1::-1], flags=cv2.INTER_LINEAR + ) + return img2 + + def _random_color_contrast(self, img1, img2): + if np.random.random() < 0.5: + contrast_factor = np.random.uniform(0.8, 1.2) + img1 = FF.adjust_contrast(img1, contrast_factor) + if self.color_aug_asym and np.random.random() < 0.5: + contrast_factor = np.random.uniform(0.8, 1.2) + img2 = FF.adjust_contrast(img2, contrast_factor) + return img1, img2 + + def _random_color_gamma(self, img1, img2): + if np.random.random() < 0.5: + gamma = np.random.uniform(0.7, 1.5) + img1 = FF.adjust_gamma(img1, gamma) + if self.color_aug_asym and np.random.random() < 0.5: + gamma = np.random.uniform(0.7, 1.5) + img2 = FF.adjust_gamma(img2, gamma) + return img1, img2 + + def _random_color_brightness(self, img1, img2): + if np.random.random() < 0.5: + brightness = np.random.uniform(0.5, 2.0) + img1 = FF.adjust_brightness(img1, brightness) + if self.color_aug_asym and np.random.random() < 0.5: + brightness = np.random.uniform(0.5, 2.0) + img2 = FF.adjust_brightness(img2, brightness) + return img1, img2 + + def _random_color_hue(self, img1, img2): + if np.random.random() < 0.5: + hue = np.random.uniform(-0.1, 0.1) + img1 = FF.adjust_hue(img1, hue) + if self.color_aug_asym and np.random.random() < 0.5: + hue = np.random.uniform(-0.1, 0.1) + img2 = FF.adjust_hue(img2, hue) + return img1, img2 + + def _random_color_saturation(self, img1, img2): + if np.random.random() < 0.5: + saturation = np.random.uniform(0.8, 1.2) + img1 = FF.adjust_saturation(img1, saturation) + if self.color_aug_asym and np.random.random() < 0.5: + saturation = np.random.uniform(-0.8, 1.2) + img2 = FF.adjust_saturation(img2, saturation) + return img1, img2 + + def _random_color(self, img1, img2): + trfs = [ + self._random_color_contrast, + self._random_color_gamma, + self._random_color_brightness, + self._random_color_hue, + self._random_color_saturation, + ] + img1 = Image.fromarray(img1.astype("uint8")) + img2 = Image.fromarray(img2.astype("uint8")) + if np.random.random() < self.color_choice_prob: + # A single transform + t = random.choice(trfs) + img1, img2 = t(img1, img2) + else: + # Combination of trfs + # Random order + random.shuffle(trfs) + for t in trfs: + img1, img2 = t(img1, img2) + img1 = np.array(img1).astype(np.float32) + img2 = np.array(img2).astype(np.float32) + return img1, img2 + + def __call__(self, img1, img2, disp, dataset_name): + img1, img2, disp = self._random_scale(img1, img2, disp) + img1, img2, disp = self._random_crop(img1, img2, disp) + img1, img2, disp = self._random_vflip(img1, img2, disp) + img2 = self._random_rotate_shift_right(img2) + img1, img2 = self._random_color(img1, img2) + return img1, img2, disp + + +class FlowAugmentor: + + def __init__( + self, + crop_size, + min_scale=-0.2, + max_scale=0.5, + spatial_aug_prob=0.8, + stretch_prob=0.8, + max_stretch=0.2, + h_flip_prob=0.5, + v_flip_prob=0.1, + asymmetric_color_aug_prob=0.2, + ): + + # spatial augmentation params + self.crop_size = crop_size + self.min_scale = min_scale + self.max_scale = max_scale + self.spatial_aug_prob = spatial_aug_prob + self.stretch_prob = stretch_prob + self.max_stretch = max_stretch + + # flip augmentation params + self.h_flip_prob = h_flip_prob + self.v_flip_prob = v_flip_prob + + # photometric augmentation params + self.photo_aug = ColorJitter( + brightness=0.4, contrast=0.4, saturation=0.4, hue=0.5 / 3.14 + ) + + self.asymmetric_color_aug_prob = asymmetric_color_aug_prob + + def color_transform(self, img1, img2): + """Photometric augmentation""" + + # asymmetric + if np.random.rand() < self.asymmetric_color_aug_prob: + img1 = np.array(self.photo_aug(Image.fromarray(img1)), dtype=np.uint8) + img2 = np.array(self.photo_aug(Image.fromarray(img2)), dtype=np.uint8) + + # symmetric + else: + image_stack = np.concatenate([img1, img2], axis=0) + image_stack = np.array( + self.photo_aug(Image.fromarray(image_stack)), dtype=np.uint8 + ) + img1, img2 = np.split(image_stack, 2, axis=0) + + return img1, img2 + + def _resize_flow(self, flow, scale_x, scale_y, factor=1.0): + if np.all(np.isfinite(flow)): + flow = cv2.resize( + flow, + None, + fx=scale_x / factor, + fy=scale_y / factor, + interpolation=cv2.INTER_LINEAR, + ) + flow = flow * [scale_x, scale_y] + else: # sparse version + fx, fy = scale_x, scale_y + ht, wd = flow.shape[:2] + coords = np.meshgrid(np.arange(wd), np.arange(ht)) + coords = np.stack(coords, axis=-1) + + coords = coords.reshape(-1, 2).astype(np.float32) + flow = flow.reshape(-1, 2).astype(np.float32) + valid = np.isfinite(flow[:, 0]) + + coords0 = coords[valid] + flow0 = flow[valid] + + ht1 = int(round(ht * fy / factor)) + wd1 = int(round(wd * fx / factor)) + + rescale = np.expand_dims(np.array([fx, fy]), axis=0) + coords1 = coords0 * rescale / factor + flow1 = flow0 * rescale + + xx = np.round(coords1[:, 0]).astype(np.int32) + yy = np.round(coords1[:, 1]).astype(np.int32) + + v = (xx > 0) & (xx < wd1) & (yy > 0) & (yy < ht1) + xx = xx[v] + yy = yy[v] + flow1 = flow1[v] + + flow = np.inf * np.ones( + [ht1, wd1, 2], dtype=np.float32 + ) # invalid value every where, before we fill it with the correct ones + flow[yy, xx] = flow1 + return flow + + def spatial_transform(self, img1, img2, flow, dname): + + if np.random.rand() < self.spatial_aug_prob: + # randomly sample scale + ht, wd = img1.shape[:2] + clip_min_scale = np.maximum( + (self.crop_size[0] + 8) / float(ht), (self.crop_size[1] + 8) / float(wd) + ) + min_scale, max_scale = self.min_scale, self.max_scale + scale = 2 ** np.random.uniform(self.min_scale, self.max_scale) + scale_x = scale + scale_y = scale + if np.random.rand() < self.stretch_prob: + scale_x *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch) + scale_y *= 2 ** np.random.uniform(-self.max_stretch, self.max_stretch) + scale_x = np.clip(scale_x, clip_min_scale, None) + scale_y = np.clip(scale_y, clip_min_scale, None) + # rescale the images + img1 = cv2.resize( + img1, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR + ) + img2 = cv2.resize( + img2, None, fx=scale_x, fy=scale_y, interpolation=cv2.INTER_LINEAR + ) + flow = self._resize_flow( + flow, scale_x, scale_y, factor=2.0 if dname == "Spring" else 1.0 + ) + elif dname == "Spring": + flow = self._resize_flow(flow, 1.0, 1.0, factor=2.0) + + if self.h_flip_prob > 0.0 and np.random.rand() < self.h_flip_prob: # h-flip + img1 = img1[:, ::-1] + img2 = img2[:, ::-1] + flow = flow[:, ::-1] * [-1.0, 1.0] + + if self.v_flip_prob > 0.0 and np.random.rand() < self.v_flip_prob: # v-flip + img1 = img1[::-1, :] + img2 = img2[::-1, :] + flow = flow[::-1, :] * [1.0, -1.0] + + # In case no cropping + if img1.shape[0] - self.crop_size[0] > 0: + y0 = np.random.randint(0, img1.shape[0] - self.crop_size[0]) + else: + y0 = 0 + if img1.shape[1] - self.crop_size[1] > 0: + x0 = np.random.randint(0, img1.shape[1] - self.crop_size[1]) + else: + x0 = 0 + + img1 = img1[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]] + img2 = img2[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]] + flow = flow[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]] + + return img1, img2, flow + + def __call__(self, img1, img2, flow, dname): + img1, img2, flow = self.spatial_transform(img1, img2, flow, dname) + img1, img2 = self.color_transform(img1, img2) + img1 = np.ascontiguousarray(img1) + img2 = np.ascontiguousarray(img2) + flow = np.ascontiguousarray(flow) + return img1, img2, flow diff --git a/croco/stereoflow/criterion.py b/croco/stereoflow/criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..f041240edb549e32f2eaa1123b07871deb322fd5 --- /dev/null +++ b/croco/stereoflow/criterion.py @@ -0,0 +1,351 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Losses, metrics per batch, metrics per dataset +# -------------------------------------------------------- + +import torch +from torch import nn +import torch.nn.functional as F + + +def _get_gtnorm(gt): + if gt.size(1) == 1: # stereo + return gt + # flow + return torch.sqrt(torch.sum(gt**2, dim=1, keepdims=True)) # Bx1xHxW + + +############ losses without confidence + + +class L1Loss(nn.Module): + + def __init__(self, max_gtnorm=None): + super().__init__() + self.max_gtnorm = max_gtnorm + self.with_conf = False + + def _error(self, gt, predictions): + return torch.abs(gt - predictions) + + def forward(self, predictions, gt, inspect=False): + mask = torch.isfinite(gt) + if self.max_gtnorm is not None: + mask *= _get_gtnorm(gt).expand(-1, gt.size(1), -1, -1) < self.max_gtnorm + if inspect: + return self._error(gt, predictions) + return self._error(gt[mask], predictions[mask]).mean() + + +############## losses with confience +## there are several parametrizations + + +class LaplacianLoss(nn.Module): # used for CroCo-Stereo on ETH3D, d'=exp(d) + + def __init__(self, max_gtnorm=None): + super().__init__() + self.max_gtnorm = max_gtnorm + self.with_conf = True + + def forward(self, predictions, gt, conf): + mask = torch.isfinite(gt) + mask = mask[:, 0, :, :] + if self.max_gtnorm is not None: + mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm + conf = conf.squeeze(1) + return ( + torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask]) + + conf[mask] + ).mean() # + torch.log(2) => which is a constant + + +class LaplacianLossBounded( + nn.Module +): # used for CroCo-Flow ; in the equation of the paper, we have a=1/b + def __init__(self, max_gtnorm=10000.0, a=0.25, b=4.0): + super().__init__() + self.max_gtnorm = max_gtnorm + self.with_conf = True + self.a, self.b = a, b + + def forward(self, predictions, gt, conf): + mask = torch.isfinite(gt) + mask = mask[:, 0, :, :] + if self.max_gtnorm is not None: + mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm + conf = conf.squeeze(1) + conf = (self.b - self.a) * torch.sigmoid(conf) + self.a + return ( + torch.abs(gt - predictions).sum(dim=1)[mask] / conf[mask] + + torch.log(conf)[mask] + ).mean() # + torch.log(2) => which is a constant + + +class LaplacianLossBounded2( + nn.Module +): # used for CroCo-Stereo (except for ETH3D) ; in the equation of the paper, we have a=b + def __init__(self, max_gtnorm=None, a=3.0, b=3.0): + super().__init__() + self.max_gtnorm = max_gtnorm + self.with_conf = True + self.a, self.b = a, b + + def forward(self, predictions, gt, conf): + mask = torch.isfinite(gt) + mask = mask[:, 0, :, :] + if self.max_gtnorm is not None: + mask *= _get_gtnorm(gt)[:, 0, :, :] < self.max_gtnorm + conf = conf.squeeze(1) + conf = 2 * self.a * (torch.sigmoid(conf / self.b) - 0.5) + return ( + torch.abs(gt - predictions).sum(dim=1)[mask] / torch.exp(conf[mask]) + + conf[mask] + ).mean() # + torch.log(2) => which is a constant + + +############## metrics per batch + + +class StereoMetrics(nn.Module): + + def __init__(self, do_quantile=False): + super().__init__() + self.bad_ths = [0.5, 1, 2, 3] + self.do_quantile = do_quantile + + def forward(self, predictions, gt): + B = predictions.size(0) + metrics = {} + gtcopy = gt.clone() + mask = torch.isfinite(gtcopy) + gtcopy[~mask] = ( + 999999.0 # we make a copy and put a non-infinite value, such that it does not become nan once multiplied by the mask value 0 + ) + Npx = mask.view(B, -1).sum(dim=1) + L1error = (torch.abs(gtcopy - predictions) * mask).view(B, -1) + L2error = (torch.square(gtcopy - predictions) * mask).view(B, -1) + # avgerr + metrics["avgerr"] = torch.mean(L1error.sum(dim=1) / Npx) + # rmse + metrics["rmse"] = torch.sqrt(L2error.sum(dim=1) / Npx).mean(dim=0) + # err > t for t in [0.5,1,2,3] + for ths in self.bad_ths: + metrics["bad@{:.1f}".format(ths)] = ( + ((L1error > ths) * mask.view(B, -1)).sum(dim=1) / Npx + ).mean(dim=0) * 100 + return metrics + + +class FlowMetrics(nn.Module): + def __init__(self): + super().__init__() + self.bad_ths = [1, 3, 5] + + def forward(self, predictions, gt): + B = predictions.size(0) + metrics = {} + mask = torch.isfinite(gt[:, 0, :, :]) # both x and y would be infinite + Npx = mask.view(B, -1).sum(dim=1) + gtcopy = ( + gt.clone() + ) # to compute L1/L2 error, we need to have non-infinite value, the error computed at this locations will be ignored + gtcopy[:, 0, :, :][~mask] = 999999.0 + gtcopy[:, 1, :, :][~mask] = 999999.0 + L1error = (torch.abs(gtcopy - predictions).sum(dim=1) * mask).view(B, -1) + L2error = ( + torch.sqrt(torch.sum(torch.square(gtcopy - predictions), dim=1)) * mask + ).view(B, -1) + metrics["L1err"] = torch.mean(L1error.sum(dim=1) / Npx) + metrics["EPE"] = torch.mean(L2error.sum(dim=1) / Npx) + for ths in self.bad_ths: + metrics["bad@{:.1f}".format(ths)] = ( + ((L2error > ths) * mask.view(B, -1)).sum(dim=1) / Npx + ).mean(dim=0) * 100 + return metrics + + +############## metrics per dataset +## we update the average and maintain the number of pixels while adding data batch per batch +## at the beggining, call reset() +## after each batch, call add_batch(...) +## at the end: call get_results() + + +class StereoDatasetMetrics(nn.Module): + + def __init__(self): + super().__init__() + self.bad_ths = [0.5, 1, 2, 3] + + def reset(self): + self.agg_N = 0 # number of pixels so far + self.agg_L1err = torch.tensor(0.0) # L1 error so far + self.agg_Nbad = [0 for _ in self.bad_ths] # counter of bad pixels + self._metrics = None + + def add_batch(self, predictions, gt): + assert predictions.size(1) == 1, predictions.size() + assert gt.size(1) == 1, gt.size() + if ( + gt.size(2) == predictions.size(2) * 2 + and gt.size(3) == predictions.size(3) * 2 + ): # special case for Spring ... + L1err = torch.minimum( + torch.minimum( + torch.minimum( + torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1), + torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1), + ), + torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1), + ), + torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1), + ) + valid = torch.isfinite(L1err) + else: + valid = torch.isfinite(gt[:, 0, :, :]) # both x and y would be infinite + L1err = torch.sum(torch.abs(gt - predictions), dim=1) + N = valid.sum() + Nnew = self.agg_N + N + self.agg_L1err = ( + float(self.agg_N) / Nnew * self.agg_L1err + + L1err[valid].mean().cpu() * float(N) / Nnew + ) + self.agg_N = Nnew + for i, th in enumerate(self.bad_ths): + self.agg_Nbad[i] += (L1err[valid] > th).sum().cpu() + + def _compute_metrics(self): + if self._metrics is not None: + return + out = {} + out["L1err"] = self.agg_L1err.item() + for i, th in enumerate(self.bad_ths): + out["bad@{:.1f}".format(th)] = ( + float(self.agg_Nbad[i]) / self.agg_N + ).item() * 100.0 + self._metrics = out + + def get_results(self): + self._compute_metrics() # to avoid recompute them multiple times + return self._metrics + + +class FlowDatasetMetrics(nn.Module): + + def __init__(self): + super().__init__() + self.bad_ths = [0.5, 1, 3, 5] + self.speed_ths = [(0, 10), (10, 40), (40, torch.inf)] + + def reset(self): + self.agg_N = 0 # number of pixels so far + self.agg_L1err = torch.tensor(0.0) # L1 error so far + self.agg_L2err = torch.tensor(0.0) # L2 (=EPE) error so far + self.agg_Nbad = [0 for _ in self.bad_ths] # counter of bad pixels + self.agg_EPEspeed = [ + torch.tensor(0.0) for _ in self.speed_ths + ] # EPE per speed bin so far + self.agg_Nspeed = [0 for _ in self.speed_ths] # N pixels per speed bin so far + self._metrics = None + self.pairname_results = {} + + def add_batch(self, predictions, gt): + assert predictions.size(1) == 2, predictions.size() + assert gt.size(1) == 2, gt.size() + if ( + gt.size(2) == predictions.size(2) * 2 + and gt.size(3) == predictions.size(3) * 2 + ): # special case for Spring ... + L1err = torch.minimum( + torch.minimum( + torch.minimum( + torch.sum(torch.abs(gt[:, :, 0::2, 0::2] - predictions), dim=1), + torch.sum(torch.abs(gt[:, :, 1::2, 0::2] - predictions), dim=1), + ), + torch.sum(torch.abs(gt[:, :, 0::2, 1::2] - predictions), dim=1), + ), + torch.sum(torch.abs(gt[:, :, 1::2, 1::2] - predictions), dim=1), + ) + L2err = torch.minimum( + torch.minimum( + torch.minimum( + torch.sqrt( + torch.sum( + torch.square(gt[:, :, 0::2, 0::2] - predictions), dim=1 + ) + ), + torch.sqrt( + torch.sum( + torch.square(gt[:, :, 1::2, 0::2] - predictions), dim=1 + ) + ), + ), + torch.sqrt( + torch.sum( + torch.square(gt[:, :, 0::2, 1::2] - predictions), dim=1 + ) + ), + ), + torch.sqrt( + torch.sum(torch.square(gt[:, :, 1::2, 1::2] - predictions), dim=1) + ), + ) + valid = torch.isfinite(L1err) + gtspeed = ( + torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 0::2]), dim=1)) + + torch.sqrt(torch.sum(torch.square(gt[:, :, 0::2, 1::2]), dim=1)) + + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 0::2]), dim=1)) + + torch.sqrt(torch.sum(torch.square(gt[:, :, 1::2, 1::2]), dim=1)) + ) / 4.0 # let's just average them + else: + valid = torch.isfinite(gt[:, 0, :, :]) # both x and y would be infinite + L1err = torch.sum(torch.abs(gt - predictions), dim=1) + L2err = torch.sqrt(torch.sum(torch.square(gt - predictions), dim=1)) + gtspeed = torch.sqrt(torch.sum(torch.square(gt), dim=1)) + N = valid.sum() + Nnew = self.agg_N + N + self.agg_L1err = ( + float(self.agg_N) / Nnew * self.agg_L1err + + L1err[valid].mean().cpu() * float(N) / Nnew + ) + self.agg_L2err = ( + float(self.agg_N) / Nnew * self.agg_L2err + + L2err[valid].mean().cpu() * float(N) / Nnew + ) + self.agg_N = Nnew + for i, th in enumerate(self.bad_ths): + self.agg_Nbad[i] += (L2err[valid] > th).sum().cpu() + for i, (th1, th2) in enumerate(self.speed_ths): + vv = (gtspeed[valid] >= th1) * (gtspeed[valid] < th2) + iNspeed = vv.sum() + if iNspeed == 0: + continue + iNnew = self.agg_Nspeed[i] + iNspeed + self.agg_EPEspeed[i] = ( + float(self.agg_Nspeed[i]) / iNnew * self.agg_EPEspeed[i] + + float(iNspeed) / iNnew * L2err[valid][vv].mean().cpu() + ) + self.agg_Nspeed[i] = iNnew + + def _compute_metrics(self): + if self._metrics is not None: + return + out = {} + out["L1err"] = self.agg_L1err.item() + out["EPE"] = self.agg_L2err.item() + for i, th in enumerate(self.bad_ths): + out["bad@{:.1f}".format(th)] = ( + float(self.agg_Nbad[i]) / self.agg_N + ).item() * 100.0 + for i, (th1, th2) in enumerate(self.speed_ths): + out["s{:d}{:s}".format(th1, "-" + str(th2) if th2 < torch.inf else "+")] = ( + self.agg_EPEspeed[i].item() + ) + self._metrics = out + + def get_results(self): + self._compute_metrics() # to avoid recompute them multiple times + return self._metrics diff --git a/croco/stereoflow/datasets_flow.py b/croco/stereoflow/datasets_flow.py new file mode 100644 index 0000000000000000000000000000000000000000..d5b1bc603b97a18e1245ec1756b74a9424d53ead --- /dev/null +++ b/croco/stereoflow/datasets_flow.py @@ -0,0 +1,936 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Dataset structure for flow +# -------------------------------------------------------- + +import os +import os.path as osp +import pickle +import numpy as np +import struct +from PIL import Image +import json +import h5py +import torch +from torch.utils import data + +from .augmentor import FlowAugmentor +from .datasets_stereo import _read_img, img_to_tensor, dataset_to_root, _read_pfm +from copy import deepcopy + +dataset_to_root = deepcopy(dataset_to_root) + +dataset_to_root.update( + **{ + "TartanAir": "./data/stereoflow/TartanAir", + "FlyingChairs": "./data/stereoflow/FlyingChairs/", + "FlyingThings": osp.join(dataset_to_root["SceneFlow"], "FlyingThings") + "/", + "MPISintel": "./data/stereoflow//MPI-Sintel/" + "/", + } +) +cache_dir = "./data/stereoflow/datasets_flow_cache/" + + +def flow_to_tensor(disp): + return torch.from_numpy(disp).float().permute(2, 0, 1) + + +class FlowDataset(data.Dataset): + + def __init__(self, split, augmentor=False, crop_size=None, totensor=True): + self.split = split + if not augmentor: + assert crop_size is None + if crop_size is not None: + assert augmentor + self.crop_size = crop_size + self.augmentor_str = augmentor + self.augmentor = FlowAugmentor(crop_size) if augmentor else None + self.totensor = totensor + self.rmul = 1 # keep track of rmul + self.has_constant_resolution = True # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time) + self._prepare_data() + self._load_or_build_cache() + + def prepare_data(self): + """ + to be defined for each dataset + """ + raise NotImplementedError + + def __len__(self): + return len( + self.pairnames + ) # each pairname is typically of the form (str, int1, int2) + + def __getitem__(self, index): + pairname = self.pairnames[index] + + # get filenames + img1name = self.pairname_to_img1name(pairname) + img2name = self.pairname_to_img2name(pairname) + flowname = ( + self.pairname_to_flowname(pairname) + if self.pairname_to_flowname is not None + else None + ) + + # load images and disparities + img1 = _read_img(img1name) + img2 = _read_img(img2name) + flow = self.load_flow(flowname) if flowname is not None else None + + # apply augmentations + if self.augmentor is not None: + img1, img2, flow = self.augmentor(img1, img2, flow, self.name) + + if self.totensor: + img1 = img_to_tensor(img1) + img2 = img_to_tensor(img2) + if flow is not None: + flow = flow_to_tensor(flow) + else: + flow = torch.tensor( + [] + ) # to allow dataloader batching with default collate_gn + pairname = str( + pairname + ) # transform potential tuple to str to be able to batch it + + return img1, img2, flow, pairname + + def __rmul__(self, v): + self.rmul *= v + self.pairnames = v * self.pairnames + return self + + def __str__(self): + return f"{self.__class__.__name__}_{self.split}" + + def __repr__(self): + s = f"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})" + if self.rmul == 1: + s += f"\n\tnum pairs: {len(self.pairnames)}" + else: + s += f"\n\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})" + return s + + def _set_root(self): + self.root = dataset_to_root[self.name] + assert os.path.isdir( + self.root + ), f"could not find root directory for dataset {self.name}: {self.root}" + + def _load_or_build_cache(self): + cache_file = osp.join(cache_dir, self.name + ".pkl") + if osp.isfile(cache_file): + with open(cache_file, "rb") as fid: + self.pairnames = pickle.load(fid)[self.split] + else: + tosave = self._build_cache() + os.makedirs(cache_dir, exist_ok=True) + with open(cache_file, "wb") as fid: + pickle.dump(tosave, fid) + self.pairnames = tosave[self.split] + + +class TartanAirDataset(FlowDataset): + + def _prepare_data(self): + self.name = "TartanAir" + self._set_root() + assert self.split in ["train"] + self.pairname_to_img1name = lambda pairname: osp.join( + self.root, pairname[0], "image_left/{:06d}_left.png".format(pairname[1]) + ) + self.pairname_to_img2name = lambda pairname: osp.join( + self.root, pairname[0], "image_left/{:06d}_left.png".format(pairname[2]) + ) + self.pairname_to_flowname = lambda pairname: osp.join( + self.root, + pairname[0], + "flow/{:06d}_{:06d}_flow.npy".format(pairname[1], pairname[2]), + ) + self.pairname_to_str = lambda pairname: os.path.join( + pairname[0][pairname[0].find("/") + 1 :], + "{:06d}_{:06d}".format(pairname[1], pairname[2]), + ) + self.load_flow = _read_numpy_flow + + def _build_cache(self): + seqs = sorted(os.listdir(self.root)) + pairs = [ + (osp.join(s, s, difficulty, Pxxx), int(a[:6]), int(a[:6]) + 1) + for s in seqs + for difficulty in ["Easy", "Hard"] + for Pxxx in sorted(os.listdir(osp.join(self.root, s, s, difficulty))) + for a in sorted( + os.listdir(osp.join(self.root, s, s, difficulty, Pxxx, "image_left/")) + )[:-1] + ] + assert len(pairs) == 306268, "incorrect parsing of pairs in TartanAir" + tosave = {"train": pairs} + return tosave + + +class FlyingChairsDataset(FlowDataset): + + def _prepare_data(self): + self.name = "FlyingChairs" + self._set_root() + assert self.split in ["train", "val"] + self.pairname_to_img1name = lambda pairname: osp.join( + self.root, "data", pairname + "_img1.ppm" + ) + self.pairname_to_img2name = lambda pairname: osp.join( + self.root, "data", pairname + "_img2.ppm" + ) + self.pairname_to_flowname = lambda pairname: osp.join( + self.root, "data", pairname + "_flow.flo" + ) + self.pairname_to_str = lambda pairname: pairname + self.load_flow = _read_flo_file + + def _build_cache(self): + split_file = osp.join(self.root, "chairs_split.txt") + split_list = np.loadtxt(split_file, dtype=np.int32) + trainpairs = ["{:05d}".format(i) for i in np.where(split_list == 1)[0] + 1] + valpairs = ["{:05d}".format(i) for i in np.where(split_list == 2)[0] + 1] + assert ( + len(trainpairs) == 22232 and len(valpairs) == 640 + ), "incorrect parsing of pairs in MPI-Sintel" + tosave = {"train": trainpairs, "val": valpairs} + return tosave + + +class FlyingThingsDataset(FlowDataset): + + def _prepare_data(self): + self.name = "FlyingThings" + self._set_root() + assert self.split in [ + f"{set_}_{pass_}pass{camstr}" + for set_ in ["train", "test", "test1024"] + for camstr in ["", "_rightcam"] + for pass_ in ["clean", "final", "all"] + ] + self.pairname_to_img1name = lambda pairname: osp.join( + self.root, + f"frames_{pairname[3]}pass", + pairname[0].replace("into_future", "").replace("into_past", ""), + "{:04d}.png".format(pairname[1]), + ) + self.pairname_to_img2name = lambda pairname: osp.join( + self.root, + f"frames_{pairname[3]}pass", + pairname[0].replace("into_future", "").replace("into_past", ""), + "{:04d}.png".format(pairname[2]), + ) + self.pairname_to_flowname = lambda pairname: osp.join( + self.root, + "optical_flow", + pairname[0], + "OpticalFlowInto{f:s}_{i:04d}_{c:s}.pfm".format( + f="Future" if "future" in pairname[0] else "Past", + i=pairname[1], + c="L" if "left" in pairname[0] else "R", + ), + ) + self.pairname_to_str = lambda pairname: os.path.join( + pairname[3] + "pass", + pairname[0], + "Into{f:s}_{i:04d}_{c:s}".format( + f="Future" if "future" in pairname[0] else "Past", + i=pairname[1], + c="L" if "left" in pairname[0] else "R", + ), + ) + self.load_flow = _read_pfm_flow + + def _build_cache(self): + tosave = {} + # train and test splits for the different passes + for set_ in ["train", "test"]: + sroot = osp.join(self.root, "optical_flow", set_.upper()) + fname_to_i = lambda f: int( + f[len("OpticalFlowIntoFuture_") : -len("_L.pfm")] + ) + pp = [ + (osp.join(set_.upper(), d, s, "into_future/left"), fname_to_i(fname)) + for d in sorted(os.listdir(sroot)) + for s in sorted(os.listdir(osp.join(sroot, d))) + for fname in sorted( + os.listdir(osp.join(sroot, d, s, "into_future/left")) + )[:-1] + ] + pairs = [(a, i, i + 1) for a, i in pp] + pairs += [(a.replace("into_future", "into_past"), i + 1, i) for a, i in pp] + assert ( + len(pairs) == {"train": 40302, "test": 7866}[set_] + ), "incorrect parsing of pairs Flying Things" + for cam in ["left", "right"]: + camstr = "" if cam == "left" else f"_{cam}cam" + for pass_ in ["final", "clean"]: + tosave[f"{set_}_{pass_}pass{camstr}"] = [ + (a.replace("left", cam), i, j, pass_) for a, i, j in pairs + ] + tosave[f"{set_}_allpass{camstr}"] = ( + tosave[f"{set_}_cleanpass{camstr}"] + + tosave[f"{set_}_finalpass{camstr}"] + ) + # test1024: this is the same split as unimatch 'validation' split + # see https://github.com/autonomousvision/unimatch/blob/master/dataloader/flow/datasets.py#L229 + test1024_nsamples = 1024 + alltest_nsamples = len(tosave["test_cleanpass"]) # 7866 + stride = alltest_nsamples // test1024_nsamples + remove = alltest_nsamples % test1024_nsamples + for cam in ["left", "right"]: + camstr = "" if cam == "left" else f"_{cam}cam" + for pass_ in ["final", "clean"]: + tosave[f"test1024_{pass_}pass{camstr}"] = sorted( + tosave[f"test_{pass_}pass{camstr}"] + )[:-remove][ + ::stride + ] # warning, it was not sorted before + assert ( + len(tosave["test1024_cleanpass"]) == 1024 + ), "incorrect parsing of pairs in Flying Things" + tosave[f"test1024_allpass{camstr}"] = ( + tosave[f"test1024_cleanpass{camstr}"] + + tosave[f"test1024_finalpass{camstr}"] + ) + return tosave + + +class MPISintelDataset(FlowDataset): + + def _prepare_data(self): + self.name = "MPISintel" + self._set_root() + assert self.split in [ + s + "_" + p + for s in ["train", "test", "subval", "subtrain"] + for p in ["cleanpass", "finalpass", "allpass"] + ] + self.pairname_to_img1name = lambda pairname: osp.join( + self.root, pairname[0], "frame_{:04d}.png".format(pairname[1]) + ) + self.pairname_to_img2name = lambda pairname: osp.join( + self.root, pairname[0], "frame_{:04d}.png".format(pairname[1] + 1) + ) + self.pairname_to_flowname = lambda pairname: ( + None + if pairname[0].startswith("test/") + else osp.join( + self.root, + pairname[0].replace("/clean/", "/flow/").replace("/final/", "/flow/"), + "frame_{:04d}.flo".format(pairname[1]), + ) + ) + self.pairname_to_str = lambda pairname: osp.join( + pairname[0], "frame_{:04d}".format(pairname[1]) + ) + self.load_flow = _read_flo_file + + def _build_cache(self): + trainseqs = sorted(os.listdir(self.root + "training/clean")) + trainpairs = [ + (osp.join("training/clean", s), i) + for s in trainseqs + for i in range(1, len(os.listdir(self.root + "training/clean/" + s))) + ] + subvalseqs = ["temple_2", "temple_3"] + subtrainseqs = [s for s in trainseqs if s not in subvalseqs] + subvalpairs = [(p, i) for p, i in trainpairs if any(s in p for s in subvalseqs)] + subtrainpairs = [ + (p, i) for p, i in trainpairs if any(s in p for s in subtrainseqs) + ] + testseqs = sorted(os.listdir(self.root + "test/clean")) + testpairs = [ + (osp.join("test/clean", s), i) + for s in testseqs + for i in range(1, len(os.listdir(self.root + "test/clean/" + s))) + ] + assert ( + len(trainpairs) == 1041 + and len(testpairs) == 552 + and len(subvalpairs) == 98 + and len(subtrainpairs) == 943 + ), "incorrect parsing of pairs in MPI-Sintel" + tosave = {} + tosave["train_cleanpass"] = trainpairs + tosave["test_cleanpass"] = testpairs + tosave["subval_cleanpass"] = subvalpairs + tosave["subtrain_cleanpass"] = subtrainpairs + for t in ["train", "test", "subval", "subtrain"]: + tosave[t + "_finalpass"] = [ + (p.replace("/clean/", "/final/"), i) + for p, i in tosave[t + "_cleanpass"] + ] + tosave[t + "_allpass"] = tosave[t + "_cleanpass"] + tosave[t + "_finalpass"] + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, _time): + assert prediction.shape[2] == 2 + outfile = os.path.join( + outdir, "submission", self.pairname_to_str(pairname) + ".flo" + ) + os.makedirs(os.path.dirname(outfile), exist_ok=True) + writeFlowFile(prediction, outfile) + + def finalize_submission(self, outdir): + assert self.split == "test_allpass" + bundle_exe = "/nfs/data/ffs-3d/datasets/StereoFlow/MPI-Sintel/bundler/linux-x64/bundler" # eg + if os.path.isfile(bundle_exe): + cmd = f'{bundle_exe} "{outdir}/submission/test/clean/" "{outdir}/submission/test/final" "{outdir}/submission/bundled.lzma"' + print(cmd) + os.system(cmd) + print(f'Done. Submission file at: "{outdir}/submission/bundled.lzma"') + else: + print("Could not find bundler executable for submission.") + print("Please download it and run:") + print( + f' "{outdir}/submission/test/clean/" "{outdir}/submission/test/final" "{outdir}/submission/bundled.lzma"' + ) + + +class SpringDataset(FlowDataset): + + def _prepare_data(self): + self.name = "Spring" + self._set_root() + assert self.split in ["train", "test", "subtrain", "subval"] + self.pairname_to_img1name = lambda pairname: osp.join( + self.root, + pairname[0], + pairname[1], + "frame_" + pairname[3], + "frame_{:s}_{:04d}.png".format(pairname[3], pairname[4]), + ) + self.pairname_to_img2name = lambda pairname: osp.join( + self.root, + pairname[0], + pairname[1], + "frame_" + pairname[3], + "frame_{:s}_{:04d}.png".format( + pairname[3], pairname[4] + (1 if pairname[2] == "FW" else -1) + ), + ) + self.pairname_to_flowname = lambda pairname: ( + None + if pairname[0] == "test" + else osp.join( + self.root, + pairname[0], + pairname[1], + f"flow_{pairname[2]}_{pairname[3]}", + f"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5", + ) + ) + self.pairname_to_str = lambda pairname: osp.join( + pairname[0], + pairname[1], + f"flow_{pairname[2]}_{pairname[3]}", + f"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}", + ) + self.load_flow = _read_hdf5_flow + + def _build_cache(self): + # train + trainseqs = sorted(os.listdir(osp.join(self.root, "train"))) + trainpairs = [] + for leftright in ["left", "right"]: + for fwbw in ["FW", "BW"]: + trainpairs += [ + ( + "train", + s, + fwbw, + leftright, + int(f[len(f"flow_{fwbw}_{leftright}_") : -len(".flo5")]), + ) + for s in trainseqs + for f in sorted( + os.listdir( + osp.join(self.root, "train", s, f"flow_{fwbw}_{leftright}") + ) + ) + ] + # test + testseqs = sorted(os.listdir(osp.join(self.root, "test"))) + testpairs = [] + for leftright in ["left", "right"]: + testpairs += [ + ( + "test", + s, + "FW", + leftright, + int(f[len(f"frame_{leftright}_") : -len(".png")]), + ) + for s in testseqs + for f in sorted( + os.listdir(osp.join(self.root, "test", s, f"frame_{leftright}")) + )[:-1] + ] + testpairs += [ + ( + "test", + s, + "BW", + leftright, + int(f[len(f"frame_{leftright}_") : -len(".png")]) + 1, + ) + for s in testseqs + for f in sorted( + os.listdir(osp.join(self.root, "test", s, f"frame_{leftright}")) + )[:-1] + ] + # subtrain / subval + subtrainpairs = [p for p in trainpairs if p[1] != "0041"] + subvalpairs = [p for p in trainpairs if p[1] == "0041"] + assert ( + len(trainpairs) == 19852 + and len(testpairs) == 3960 + and len(subtrainpairs) == 19472 + and len(subvalpairs) == 380 + ), "incorrect parsing of pairs in Spring" + tosave = { + "train": trainpairs, + "test": testpairs, + "subtrain": subtrainpairs, + "subval": subvalpairs, + } + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 3 + assert prediction.shape[2] == 2 + assert prediction.dtype == np.float32 + outfile = osp.join( + outdir, + pairname[0], + pairname[1], + f"flow_{pairname[2]}_{pairname[3]}", + f"flow_{pairname[2]}_{pairname[3]}_{pairname[4]:04d}.flo5", + ) + os.makedirs(os.path.dirname(outfile), exist_ok=True) + writeFlo5File(prediction, outfile) + + def finalize_submission(self, outdir): + assert self.split == "test" + exe = "{self.root}/flow_subsampling" + if os.path.isfile(exe): + cmd = f'cd "{outdir}/test"; {exe} .' + print(cmd) + os.system(cmd) + print(f"Done. Submission file at {outdir}/test/flow_submission.hdf5") + else: + print("Could not find flow_subsampling executable for submission.") + print("Please download it and run:") + print(f'cd "{outdir}/test"; .') + + +class Kitti12Dataset(FlowDataset): + + def _prepare_data(self): + self.name = "Kitti12" + self._set_root() + assert self.split in ["train", "test"] + self.pairname_to_img1name = lambda pairname: osp.join( + self.root, pairname + "_10.png" + ) + self.pairname_to_img2name = lambda pairname: osp.join( + self.root, pairname + "_11.png" + ) + self.pairname_to_flowname = ( + None + if self.split == "test" + else lambda pairname: osp.join( + self.root, pairname.replace("/colored_0/", "/flow_occ/") + "_10.png" + ) + ) + self.pairname_to_str = lambda pairname: pairname.replace("/colored_0/", "/") + self.load_flow = _read_kitti_flow + + def _build_cache(self): + trainseqs = ["training/colored_0/%06d" % (i) for i in range(194)] + testseqs = ["testing/colored_0/%06d" % (i) for i in range(195)] + assert ( + len(trainseqs) == 194 and len(testseqs) == 195 + ), "incorrect parsing of pairs in Kitti12" + tosave = {"train": trainseqs, "test": testseqs} + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 3 + assert prediction.shape[2] == 2 + outfile = os.path.join(outdir, pairname.split("/")[-1] + "_10.png") + os.makedirs(os.path.dirname(outfile), exist_ok=True) + writeFlowKitti(outfile, prediction) + + def finalize_submission(self, outdir): + assert self.split == "test" + cmd = f'cd {outdir}/; zip -r "kitti12_flow_results.zip" .' + print(cmd) + os.system(cmd) + print(f"Done. Submission file at {outdir}/kitti12_flow_results.zip") + + +class Kitti15Dataset(FlowDataset): + + def _prepare_data(self): + self.name = "Kitti15" + self._set_root() + assert self.split in ["train", "subtrain", "subval", "test"] + self.pairname_to_img1name = lambda pairname: osp.join( + self.root, pairname + "_10.png" + ) + self.pairname_to_img2name = lambda pairname: osp.join( + self.root, pairname + "_11.png" + ) + self.pairname_to_flowname = ( + None + if self.split == "test" + else lambda pairname: osp.join( + self.root, pairname.replace("/image_2/", "/flow_occ/") + "_10.png" + ) + ) + self.pairname_to_str = lambda pairname: pairname.replace("/image_2/", "/") + self.load_flow = _read_kitti_flow + + def _build_cache(self): + trainseqs = ["training/image_2/%06d" % (i) for i in range(200)] + subtrainseqs = trainseqs[:-10] + subvalseqs = trainseqs[-10:] + testseqs = ["testing/image_2/%06d" % (i) for i in range(200)] + assert ( + len(trainseqs) == 200 + and len(subtrainseqs) == 190 + and len(subvalseqs) == 10 + and len(testseqs) == 200 + ), "incorrect parsing of pairs in Kitti15" + tosave = { + "train": trainseqs, + "subtrain": subtrainseqs, + "subval": subvalseqs, + "test": testseqs, + } + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 3 + assert prediction.shape[2] == 2 + outfile = os.path.join(outdir, "flow", pairname.split("/")[-1] + "_10.png") + os.makedirs(os.path.dirname(outfile), exist_ok=True) + writeFlowKitti(outfile, prediction) + + def finalize_submission(self, outdir): + assert self.split == "test" + cmd = f'cd {outdir}/; zip -r "kitti15_flow_results.zip" flow' + print(cmd) + os.system(cmd) + print(f"Done. Submission file at {outdir}/kitti15_flow_results.zip") + + +import cv2 + + +def _read_numpy_flow(filename): + return np.load(filename) + + +def _read_pfm_flow(filename): + f, _ = _read_pfm(filename) + assert np.all(f[:, :, 2] == 0.0) + return np.ascontiguousarray(f[:, :, :2]) + + +TAG_FLOAT = 202021.25 # tag to check the sanity of the file +TAG_STRING = "PIEH" # string containing the tag +MIN_WIDTH = 1 +MAX_WIDTH = 99999 +MIN_HEIGHT = 1 +MAX_HEIGHT = 99999 + + +def readFlowFile(filename): + """ + readFlowFile() reads a flow file into a 2-band np.array. + if does not exist, an IOError is raised. + if does not finish by '.flo' or the tag, the width, the height or the file's size is illegal, an Expcetion is raised. + ---- PARAMETERS ---- + filename: string containg the name of the file to read a flow + ---- OUTPUTS ---- + a np.array of dimension (height x width x 2) containing the flow of type 'float32' + """ + + # check filename + if not filename.endswith(".flo"): + raise Exception( + "readFlowFile({:s}): filename must finish with '.flo'".format(filename) + ) + + # open the file and read it + with open(filename, "rb") as f: + # check tag + tag = struct.unpack("f", f.read(4))[0] + if tag != TAG_FLOAT: + raise Exception("flow_utils.readFlowFile({:s}): wrong tag".format(filename)) + # read dimension + w, h = struct.unpack("ii", f.read(8)) + if w < MIN_WIDTH or w > MAX_WIDTH: + raise Exception( + "flow_utils.readFlowFile({:s}: illegal width {:d}".format(filename, w) + ) + if h < MIN_HEIGHT or h > MAX_HEIGHT: + raise Exception( + "flow_utils.readFlowFile({:s}: illegal height {:d}".format(filename, h) + ) + flow = np.fromfile(f, "float32") + if not flow.shape == (h * w * 2,): + raise Exception( + "flow_utils.readFlowFile({:s}: illegal size of the file".format( + filename + ) + ) + flow.shape = (h, w, 2) + return flow + + +def writeFlowFile(flow, filename): + """ + writeFlowFile(flow,) write flow to the file . + if does not exist, an IOError is raised. + if does not finish with '.flo' or the flow has not 2 bands, an Exception is raised. + ---- PARAMETERS ---- + flow: np.array of dimension (height x width x 2) containing the flow to write + filename: string containg the name of the file to write a flow + """ + + # check filename + if not filename.endswith(".flo"): + raise Exception( + "flow_utils.writeFlowFile(,{:s}): filename must finish with '.flo'".format( + filename + ) + ) + + if not flow.shape[2:] == (2,): + raise Exception( + "flow_utils.writeFlowFile(,{:s}): must have 2 bands".format( + filename + ) + ) + + # open the file and write it + with open(filename, "wb") as f: + # write TAG + f.write(TAG_STRING.encode("utf-8")) + # write dimension + f.write(struct.pack("ii", flow.shape[1], flow.shape[0])) + # write the flow + + flow.astype(np.float32).tofile(f) + + +_read_flo_file = readFlowFile + + +def _read_kitti_flow(filename): + flow = cv2.imread(filename, cv2.IMREAD_ANYDEPTH | cv2.IMREAD_COLOR) + flow = flow[:, :, ::-1].astype(np.float32) + valid = flow[:, :, 2] > 0 + flow = flow[:, :, :2] + flow = (flow - 2**15) / 64.0 + flow[~valid, 0] = np.inf + flow[~valid, 1] = np.inf + return flow + + +_read_hd1k_flow = _read_kitti_flow + + +def writeFlowKitti(filename, uv): + uv = 64.0 * uv + 2**15 + valid = np.ones([uv.shape[0], uv.shape[1], 1]) + uv = np.concatenate([uv, valid], axis=-1).astype(np.uint16) + cv2.imwrite(filename, uv[..., ::-1]) + + +def writeFlo5File(flow, filename): + with h5py.File(filename, "w") as f: + f.create_dataset("flow", data=flow, compression="gzip", compression_opts=5) + + +def _read_hdf5_flow(filename): + flow = np.asarray(h5py.File(filename)["flow"]) + flow[np.isnan(flow)] = np.inf # make invalid values as +inf + return flow.astype(np.float32) + + +# flow visualization +RY = 15 +YG = 6 +GC = 4 +CB = 11 +BM = 13 +MR = 6 +UNKNOWN_THRESH = 1e9 + + +def colorTest(): + """ + flow_utils.colorTest(): display an example of image showing the color encoding scheme + """ + import matplotlib.pylab as plt + + truerange = 1 + h, w = 151, 151 + trange = truerange * 1.04 + s2 = round(h / 2) + x, y = np.meshgrid(range(w), range(h)) + u = x * trange / s2 - trange + v = y * trange / s2 - trange + img = _computeColor( + np.concatenate((u[:, :, np.newaxis], v[:, :, np.newaxis]), 2) + / trange + / np.sqrt(2) + ) + plt.imshow(img) + plt.axis("off") + plt.axhline(round(h / 2), color="k") + plt.axvline(round(w / 2), color="k") + + +def flowToColor(flow, maxflow=None, maxmaxflow=None, saturate=False): + """ + flow_utils.flowToColor(flow): return a color code flow field, normalized based on the maximum l2-norm of the flow + flow_utils.flowToColor(flow,maxflow): return a color code flow field, normalized by maxflow + ---- PARAMETERS ---- + flow: flow to display of shape (height x width x 2) + maxflow (default:None): if given, normalize the flow by its value, otherwise by the flow norm + maxmaxflow (default:None): if given, normalize the flow by the max of its value and the flow norm + ---- OUTPUT ---- + an np.array of shape (height x width x 3) of type uint8 containing a color code of the flow + """ + h, w, n = flow.shape + # check size of flow + assert n == 2, "flow_utils.flowToColor(flow): flow must have 2 bands" + # fix unknown flow + unknown_idx = np.max(np.abs(flow), 2) > UNKNOWN_THRESH + flow[unknown_idx] = 0.0 + # compute max flow if needed + if maxflow is None: + maxflow = flowMaxNorm(flow) + if maxmaxflow is not None: + maxflow = min(maxmaxflow, maxflow) + # normalize flow + eps = np.spacing(1) # minimum positive float value to avoid division by 0 + # compute the flow + img = _computeColor(flow / (maxflow + eps), saturate=saturate) + # put black pixels in unknown location + img[np.tile(unknown_idx[:, :, np.newaxis], [1, 1, 3])] = 0.0 + return img + + +def flowMaxNorm(flow): + """ + flow_utils.flowMaxNorm(flow): return the maximum of the l2-norm of the given flow + ---- PARAMETERS ---- + flow: the flow + + ---- OUTPUT ---- + a float containing the maximum of the l2-norm of the flow + """ + return np.max(np.sqrt(np.sum(np.square(flow), 2))) + + +def _computeColor(flow, saturate=True): + """ + flow_utils._computeColor(flow): compute color codes for the flow field flow + + ---- PARAMETERS ---- + flow: np.array of dimension (height x width x 2) containing the flow to display + ---- OUTPUTS ---- + an np.array of dimension (height x width x 3) containing the color conversion of the flow + """ + # set nan to 0 + nanidx = np.isnan(flow[:, :, 0]) + flow[nanidx] = 0.0 + + # colorwheel + ncols = RY + YG + GC + CB + BM + MR + nchans = 3 + colorwheel = np.zeros((ncols, nchans), "uint8") + col = 0 + # RY + colorwheel[:RY, 0] = 255 + colorwheel[:RY, 1] = [(255 * i) // RY for i in range(RY)] + col += RY + # YG + colorwheel[col : col + YG, 0] = [255 - (255 * i) // YG for i in range(YG)] + colorwheel[col : col + YG, 1] = 255 + col += YG + # GC + colorwheel[col : col + GC, 1] = 255 + colorwheel[col : col + GC, 2] = [(255 * i) // GC for i in range(GC)] + col += GC + # CB + colorwheel[col : col + CB, 1] = [255 - (255 * i) // CB for i in range(CB)] + colorwheel[col : col + CB, 2] = 255 + col += CB + # BM + colorwheel[col : col + BM, 0] = [(255 * i) // BM for i in range(BM)] + colorwheel[col : col + BM, 2] = 255 + col += BM + # MR + colorwheel[col : col + MR, 0] = 255 + colorwheel[col : col + MR, 2] = [255 - (255 * i) // MR for i in range(MR)] + + # compute utility variables + rad = np.sqrt(np.sum(np.square(flow), 2)) # magnitude + a = np.arctan2(-flow[:, :, 1], -flow[:, :, 0]) / np.pi # angle + fk = (a + 1) / 2 * (ncols - 1) # map [-1,1] to [0,ncols-1] + k0 = np.floor(fk).astype("int") + k1 = k0 + 1 + k1[k1 == ncols] = 0 + f = fk - k0 + + if not saturate: + rad = np.minimum(rad, 1) + + # compute the image + img = np.zeros((flow.shape[0], flow.shape[1], nchans), "uint8") + for i in range(nchans): + tmp = colorwheel[:, i].astype("float") + col0 = tmp[k0] / 255 + col1 = tmp[k1] / 255 + col = (1 - f) * col0 + f * col1 + idx = rad <= 1 + col[idx] = 1 - rad[idx] * (1 - col[idx]) # increase saturation with radius + col[~idx] *= 0.75 # out of range + img[:, :, i] = (255 * col * (1 - nanidx.astype("float"))).astype("uint8") + + return img + + +# flow dataset getter + + +def get_train_dataset_flow(dataset_str, augmentor=True, crop_size=None): + dataset_str = dataset_str.replace("(", "Dataset(") + if augmentor: + dataset_str = dataset_str.replace(")", ", augmentor=True)") + if crop_size is not None: + dataset_str = dataset_str.replace( + ")", ", crop_size={:s})".format(str(crop_size)) + ) + return eval(dataset_str) + + +def get_test_datasets_flow(dataset_str): + dataset_str = dataset_str.replace("(", "Dataset(") + return [eval(s) for s in dataset_str.split("+")] diff --git a/croco/stereoflow/datasets_stereo.py b/croco/stereoflow/datasets_stereo.py new file mode 100644 index 0000000000000000000000000000000000000000..60c9466ad05164fb433551dd23acb3153e6e7ea6 --- /dev/null +++ b/croco/stereoflow/datasets_stereo.py @@ -0,0 +1,991 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Dataset structure for stereo +# -------------------------------------------------------- + +import sys, os +import os.path as osp +import pickle +import numpy as np +from PIL import Image +import json +import h5py +from glob import glob +import cv2 + +import torch +from torch.utils import data + +from .augmentor import StereoAugmentor + + +dataset_to_root = { + "CREStereo": "./data/stereoflow//crenet_stereo_trainset/stereo_trainset/crestereo/", + "SceneFlow": "./data/stereoflow//SceneFlow/", + "ETH3DLowRes": "./data/stereoflow/eth3d_lowres/", + "Booster": "./data/stereoflow/booster_gt/", + "Middlebury2021": "./data/stereoflow/middlebury/2021/data/", + "Middlebury2014": "./data/stereoflow/middlebury/2014/", + "Middlebury2006": "./data/stereoflow/middlebury/2006/", + "Middlebury2005": "./data/stereoflow/middlebury/2005/train/", + "MiddleburyEval3": "./data/stereoflow/middlebury/MiddEval3/", + "Spring": "./data/stereoflow/spring/", + "Kitti15": "./data/stereoflow/kitti-stereo-2015/", + "Kitti12": "./data/stereoflow/kitti-stereo-2012/", +} +cache_dir = "./data/stereoflow/datasets_stereo_cache/" + + +in1k_mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1) +in1k_std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1) + + +def img_to_tensor(img): + img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.0 + img = (img - in1k_mean) / in1k_std + return img + + +def disp_to_tensor(disp): + return torch.from_numpy(disp)[None, :, :] + + +class StereoDataset(data.Dataset): + + def __init__(self, split, augmentor=False, crop_size=None, totensor=True): + self.split = split + if not augmentor: + assert crop_size is None + if crop_size: + assert augmentor + self.crop_size = crop_size + self.augmentor_str = augmentor + self.augmentor = StereoAugmentor(crop_size) if augmentor else None + self.totensor = totensor + self.rmul = 1 # keep track of rmul + self.has_constant_resolution = True # whether the dataset has constant resolution or not (=> don't use batch_size>1 at test time) + self._prepare_data() + self._load_or_build_cache() + + def prepare_data(self): + """ + to be defined for each dataset + """ + raise NotImplementedError + + def __len__(self): + return len(self.pairnames) + + def __getitem__(self, index): + pairname = self.pairnames[index] + + # get filenames + Limgname = self.pairname_to_Limgname(pairname) + Rimgname = self.pairname_to_Rimgname(pairname) + Ldispname = ( + self.pairname_to_Ldispname(pairname) + if self.pairname_to_Ldispname is not None + else None + ) + + # load images and disparities + Limg = _read_img(Limgname) + Rimg = _read_img(Rimgname) + disp = self.load_disparity(Ldispname) if Ldispname is not None else None + + # sanity check + if disp is not None: + assert np.all(disp > 0) or self.name == "Spring", ( + self.name, + pairname, + Ldispname, + ) + + # apply augmentations + if self.augmentor is not None: + Limg, Rimg, disp = self.augmentor(Limg, Rimg, disp, self.name) + + if self.totensor: + Limg = img_to_tensor(Limg) + Rimg = img_to_tensor(Rimg) + if disp is None: + disp = torch.tensor( + [] + ) # to allow dataloader batching with default collate_gn + else: + disp = disp_to_tensor(disp) + + return Limg, Rimg, disp, str(pairname) + + def __rmul__(self, v): + self.rmul *= v + self.pairnames = v * self.pairnames + return self + + def __str__(self): + return f"{self.__class__.__name__}_{self.split}" + + def __repr__(self): + s = f"{self.__class__.__name__}(split={self.split}, augmentor={self.augmentor_str}, crop_size={str(self.crop_size)}, totensor={self.totensor})" + if self.rmul == 1: + s += f"\n\tnum pairs: {len(self.pairnames)}" + else: + s += f"\n\tnum pairs: {len(self.pairnames)} ({len(self.pairnames)//self.rmul}x{self.rmul})" + return s + + def _set_root(self): + self.root = dataset_to_root[self.name] + assert os.path.isdir( + self.root + ), f"could not find root directory for dataset {self.name}: {self.root}" + + def _load_or_build_cache(self): + cache_file = osp.join(cache_dir, self.name + ".pkl") + if osp.isfile(cache_file): + with open(cache_file, "rb") as fid: + self.pairnames = pickle.load(fid)[self.split] + else: + tosave = self._build_cache() + os.makedirs(cache_dir, exist_ok=True) + with open(cache_file, "wb") as fid: + pickle.dump(tosave, fid) + self.pairnames = tosave[self.split] + + +class CREStereoDataset(StereoDataset): + + def _prepare_data(self): + self.name = "CREStereo" + self._set_root() + assert self.split in ["train"] + self.pairname_to_Limgname = lambda pairname: osp.join( + self.root, pairname + "_left.jpg" + ) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname + "_right.jpg" + ) + self.pairname_to_Ldispname = lambda pairname: osp.join( + self.root, pairname + "_left.disp.png" + ) + self.pairname_to_str = lambda pairname: pairname + self.load_disparity = _read_crestereo_disp + + def _build_cache(self): + allpairs = [ + s + "/" + f[: -len("_left.jpg")] + for s in sorted(os.listdir(self.root)) + for f in sorted(os.listdir(self.root + "/" + s)) + if f.endswith("_left.jpg") + ] + assert len(allpairs) == 200000, "incorrect parsing of pairs in CreStereo" + tosave = {"train": allpairs} + return tosave + + +class SceneFlowDataset(StereoDataset): + + def _prepare_data(self): + self.name = "SceneFlow" + self._set_root() + assert self.split in [ + "train_finalpass", + "train_cleanpass", + "train_allpass", + "test_finalpass", + "test_cleanpass", + "test_allpass", + "test1of100_cleanpass", + "test1of100_finalpass", + ] + self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname + ).replace("/left/", "/right/") + self.pairname_to_Ldispname = ( + lambda pairname: osp.join(self.root, pairname) + .replace("/frames_finalpass/", "/disparity/") + .replace("/frames_cleanpass/", "/disparity/")[:-4] + + ".pfm" + ) + self.pairname_to_str = lambda pairname: pairname[:-4] + self.load_disparity = _read_sceneflow_disp + + def _build_cache(self): + trainpairs = [] + # driving + pairs = sorted(glob(self.root + "Driving/frames_finalpass/*/*/*/left/*.png")) + pairs = list(map(lambda x: x[len(self.root) :], pairs)) + assert len(pairs) == 4400, "incorrect parsing of pairs in SceneFlow" + trainpairs += pairs + # monkaa + pairs = sorted(glob(self.root + "Monkaa/frames_finalpass/*/left/*.png")) + pairs = list(map(lambda x: x[len(self.root) :], pairs)) + assert len(pairs) == 8664, "incorrect parsing of pairs in SceneFlow" + trainpairs += pairs + # flyingthings + pairs = sorted( + glob(self.root + "FlyingThings/frames_finalpass/TRAIN/*/*/left/*.png") + ) + pairs = list(map(lambda x: x[len(self.root) :], pairs)) + assert len(pairs) == 22390, "incorrect parsing of pairs in SceneFlow" + trainpairs += pairs + assert len(trainpairs) == 35454, "incorrect parsing of pairs in SceneFlow" + testpairs = sorted( + glob(self.root + "FlyingThings/frames_finalpass/TEST/*/*/left/*.png") + ) + testpairs = list(map(lambda x: x[len(self.root) :], testpairs)) + assert len(testpairs) == 4370, "incorrect parsing of pairs in SceneFlow" + test1of100pairs = testpairs[::100] + assert len(test1of100pairs) == 44, "incorrect parsing of pairs in SceneFlow" + # all + tosave = { + "train_finalpass": trainpairs, + "train_cleanpass": list( + map( + lambda x: x.replace("frames_finalpass", "frames_cleanpass"), + trainpairs, + ) + ), + "test_finalpass": testpairs, + "test_cleanpass": list( + map( + lambda x: x.replace("frames_finalpass", "frames_cleanpass"), + testpairs, + ) + ), + "test1of100_finalpass": test1of100pairs, + "test1of100_cleanpass": list( + map( + lambda x: x.replace("frames_finalpass", "frames_cleanpass"), + test1of100pairs, + ) + ), + } + tosave["train_allpass"] = tosave["train_finalpass"] + tosave["train_cleanpass"] + tosave["test_allpass"] = tosave["test_finalpass"] + tosave["test_cleanpass"] + return tosave + + +class Md21Dataset(StereoDataset): + + def _prepare_data(self): + self.name = "Middlebury2021" + self._set_root() + assert self.split in ["train", "subtrain", "subval"] + self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname.replace("/im0", "/im1") + ) + self.pairname_to_Ldispname = lambda pairname: osp.join( + self.root, pairname.split("/")[0], "disp0.pfm" + ) + self.pairname_to_str = lambda pairname: pairname[:-4] + self.load_disparity = _read_middlebury_disp + + def _build_cache(self): + seqs = sorted(os.listdir(self.root)) + trainpairs = [] + for s in seqs: + # trainpairs += [s+'/im0.png'] # we should remove it, it is included as such in other lightings + trainpairs += [ + s + "/ambient/" + b + "/" + a + for b in sorted(os.listdir(osp.join(self.root, s, "ambient"))) + for a in sorted(os.listdir(osp.join(self.root, s, "ambient", b))) + if a.startswith("im0") + ] + assert len(trainpairs) == 355 + subtrainpairs = [ + p for p in trainpairs if any(p.startswith(s + "/") for s in seqs[:-2]) + ] + subvalpairs = [ + p for p in trainpairs if any(p.startswith(s + "/") for s in seqs[-2:]) + ] + assert ( + len(subtrainpairs) == 335 and len(subvalpairs) == 20 + ), "incorrect parsing of pairs in Middlebury 2021" + tosave = {"train": trainpairs, "subtrain": subtrainpairs, "subval": subvalpairs} + return tosave + + +class Md14Dataset(StereoDataset): + + def _prepare_data(self): + self.name = "Middlebury2014" + self._set_root() + assert self.split in ["train", "subtrain", "subval"] + self.pairname_to_Limgname = lambda pairname: osp.join( + self.root, osp.dirname(pairname), "im0.png" + ) + self.pairname_to_Rimgname = lambda pairname: osp.join(self.root, pairname) + self.pairname_to_Ldispname = lambda pairname: osp.join( + self.root, osp.dirname(pairname), "disp0.pfm" + ) + self.pairname_to_str = lambda pairname: pairname[:-4] + self.load_disparity = _read_middlebury_disp + self.has_constant_resolution = False + + def _build_cache(self): + seqs = sorted(os.listdir(self.root)) + trainpairs = [] + for s in seqs: + trainpairs += [s + "/im1.png", s + "/im1E.png", s + "/im1L.png"] + assert len(trainpairs) == 138 + valseqs = ["Umbrella-imperfect", "Vintage-perfect"] + assert all(s in seqs for s in valseqs) + subtrainpairs = [ + p for p in trainpairs if not any(p.startswith(s + "/") for s in valseqs) + ] + subvalpairs = [ + p for p in trainpairs if any(p.startswith(s + "/") for s in valseqs) + ] + assert ( + len(subtrainpairs) == 132 and len(subvalpairs) == 6 + ), "incorrect parsing of pairs in Middlebury 2014" + tosave = {"train": trainpairs, "subtrain": subtrainpairs, "subval": subvalpairs} + return tosave + + +class Md06Dataset(StereoDataset): + + def _prepare_data(self): + self.name = "Middlebury2006" + self._set_root() + assert self.split in ["train", "subtrain", "subval"] + self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, osp.dirname(pairname), "view5.png" + ) + self.pairname_to_Ldispname = lambda pairname: osp.join( + self.root, pairname.split("/")[0], "disp1.png" + ) + self.load_disparity = _read_middlebury20052006_disp + self.has_constant_resolution = False + + def _build_cache(self): + seqs = sorted(os.listdir(self.root)) + trainpairs = [] + for s in seqs: + for i in ["Illum1", "Illum2", "Illum3"]: + for e in ["Exp0", "Exp1", "Exp2"]: + trainpairs.append(osp.join(s, i, e, "view1.png")) + assert len(trainpairs) == 189 + valseqs = ["Rocks1", "Wood2"] + assert all(s in seqs for s in valseqs) + subtrainpairs = [ + p for p in trainpairs if not any(p.startswith(s + "/") for s in valseqs) + ] + subvalpairs = [ + p for p in trainpairs if any(p.startswith(s + "/") for s in valseqs) + ] + assert ( + len(subtrainpairs) == 171 and len(subvalpairs) == 18 + ), "incorrect parsing of pairs in Middlebury 2006" + tosave = {"train": trainpairs, "subtrain": subtrainpairs, "subval": subvalpairs} + return tosave + + +class Md05Dataset(StereoDataset): + + def _prepare_data(self): + self.name = "Middlebury2005" + self._set_root() + assert self.split in ["train", "subtrain", "subval"] + self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, osp.dirname(pairname), "view5.png" + ) + self.pairname_to_Ldispname = lambda pairname: osp.join( + self.root, pairname.split("/")[0], "disp1.png" + ) + self.pairname_to_str = lambda pairname: pairname[:-4] + self.load_disparity = _read_middlebury20052006_disp + + def _build_cache(self): + seqs = sorted(os.listdir(self.root)) + trainpairs = [] + for s in seqs: + for i in ["Illum1", "Illum2", "Illum3"]: + for e in ["Exp0", "Exp1", "Exp2"]: + trainpairs.append(osp.join(s, i, e, "view1.png")) + assert len(trainpairs) == 54, "incorrect parsing of pairs in Middlebury 2005" + valseqs = ["Reindeer"] + assert all(s in seqs for s in valseqs) + subtrainpairs = [ + p for p in trainpairs if not any(p.startswith(s + "/") for s in valseqs) + ] + subvalpairs = [ + p for p in trainpairs if any(p.startswith(s + "/") for s in valseqs) + ] + assert ( + len(subtrainpairs) == 45 and len(subvalpairs) == 9 + ), "incorrect parsing of pairs in Middlebury 2005" + tosave = {"train": trainpairs, "subtrain": subtrainpairs, "subval": subvalpairs} + return tosave + + +class MdEval3Dataset(StereoDataset): + + def _prepare_data(self): + self.name = "MiddleburyEval3" + self._set_root() + assert self.split in [ + s + "_" + r + for s in ["train", "subtrain", "subval", "test", "all"] + for r in ["full", "half", "quarter"] + ] + if self.split.endswith("_full"): + self.root = self.root.replace("/MiddEval3", "/MiddEval3_F") + elif self.split.endswith("_half"): + self.root = self.root.replace("/MiddEval3", "/MiddEval3_H") + else: + assert self.split.endswith("_quarter") + self.pairname_to_Limgname = lambda pairname: osp.join( + self.root, pairname, "im0.png" + ) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname, "im1.png" + ) + self.pairname_to_Ldispname = lambda pairname: ( + None + if pairname.startswith("test") + else osp.join(self.root, pairname, "disp0GT.pfm") + ) + self.pairname_to_str = lambda pairname: pairname + self.load_disparity = _read_middlebury_disp + # for submission only + self.submission_methodname = "CroCo-Stereo" + self.submission_sresolution = ( + "F" + if self.split.endswith("_full") + else ("H" if self.split.endswith("_half") else "Q") + ) + + def _build_cache(self): + trainpairs = ["train/" + s for s in sorted(os.listdir(self.root + "train/"))] + testpairs = ["test/" + s for s in sorted(os.listdir(self.root + "test/"))] + subvalpairs = trainpairs[-1:] + subtrainpairs = trainpairs[:-1] + allpairs = trainpairs + testpairs + assert ( + len(trainpairs) == 15 + and len(testpairs) == 15 + and len(subvalpairs) == 1 + and len(subtrainpairs) == 14 + and len(allpairs) == 30 + ), "incorrect parsing of pairs in Middlebury Eval v3" + tosave = {} + for r in ["full", "half", "quarter"]: + tosave.update( + **{ + "train_" + r: trainpairs, + "subtrain_" + r: subtrainpairs, + "subval_" + r: subvalpairs, + "test_" + r: testpairs, + "all_" + r: allpairs, + } + ) + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 2 + assert prediction.dtype == np.float32 + outfile = os.path.join( + outdir, + pairname.split("/")[0].replace("train", "training") + + self.submission_sresolution, + pairname.split("/")[1], + "disp0" + self.submission_methodname + ".pfm", + ) + os.makedirs(os.path.dirname(outfile), exist_ok=True) + writePFM(outfile, prediction) + timefile = os.path.join( + os.path.dirname(outfile), "time" + self.submission_methodname + ".txt" + ) + with open(timefile, "w") as fid: + fid.write(str(time)) + + def finalize_submission(self, outdir): + cmd = f'cd {outdir}/; zip -r "{self.submission_methodname}.zip" .' + print(cmd) + os.system(cmd) + print(f"Done. Submission file at {outdir}/{self.submission_methodname}.zip") + + +class ETH3DLowResDataset(StereoDataset): + + def _prepare_data(self): + self.name = "ETH3DLowRes" + self._set_root() + assert self.split in ["train", "test", "subtrain", "subval", "all"] + self.pairname_to_Limgname = lambda pairname: osp.join( + self.root, pairname, "im0.png" + ) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname, "im1.png" + ) + self.pairname_to_Ldispname = ( + None + if self.split == "test" + else lambda pairname: ( + None + if pairname.startswith("test/") + else osp.join( + self.root, pairname.replace("train/", "train_gt/"), "disp0GT.pfm" + ) + ) + ) + self.pairname_to_str = lambda pairname: pairname + self.load_disparity = _read_eth3d_disp + self.has_constant_resolution = False + + def _build_cache(self): + trainpairs = ["train/" + s for s in sorted(os.listdir(self.root + "train/"))] + testpairs = ["test/" + s for s in sorted(os.listdir(self.root + "test/"))] + assert ( + len(trainpairs) == 27 and len(testpairs) == 20 + ), "incorrect parsing of pairs in ETH3D Low Res" + subvalpairs = [ + "train/delivery_area_3s", + "train/electro_3l", + "train/playground_3l", + ] + assert all(p in trainpairs for p in subvalpairs) + subtrainpairs = [p for p in trainpairs if not p in subvalpairs] + assert ( + len(subvalpairs) == 3 and len(subtrainpairs) == 24 + ), "incorrect parsing of pairs in ETH3D Low Res" + tosave = { + "train": trainpairs, + "test": testpairs, + "subtrain": subtrainpairs, + "subval": subvalpairs, + "all": trainpairs + testpairs, + } + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 2 + assert prediction.dtype == np.float32 + outfile = os.path.join( + outdir, "low_res_two_view", pairname.split("/")[1] + ".pfm" + ) + os.makedirs(os.path.dirname(outfile), exist_ok=True) + writePFM(outfile, prediction) + timefile = outfile[:-4] + ".txt" + with open(timefile, "w") as fid: + fid.write("runtime " + str(time)) + + def finalize_submission(self, outdir): + cmd = f'cd {outdir}/; zip -r "eth3d_low_res_two_view_results.zip" low_res_two_view' + print(cmd) + os.system(cmd) + print(f"Done. Submission file at {outdir}/eth3d_low_res_two_view_results.zip") + + +class BoosterDataset(StereoDataset): + + def _prepare_data(self): + self.name = "Booster" + self._set_root() + assert self.split in [ + "train_balanced", + "test_balanced", + "subtrain_balanced", + "subval_balanced", + ] # we use only the balanced version + self.pairname_to_Limgname = lambda pairname: osp.join(self.root, pairname) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname + ).replace("/camera_00/", "/camera_02/") + self.pairname_to_Ldispname = lambda pairname: osp.join( + self.root, osp.dirname(pairname), "../disp_00.npy" + ) # same images with different colors, same gt per sequence + self.pairname_to_str = lambda pairname: pairname[:-4].replace( + "/camera_00/", "/" + ) + self.load_disparity = _read_booster_disp + + def _build_cache(self): + trainseqs = sorted(os.listdir(self.root + "train/balanced")) + trainpairs = [ + "train/balanced/" + s + "/camera_00/" + imname + for s in trainseqs + for imname in sorted( + os.listdir(self.root + "train/balanced/" + s + "/camera_00/") + ) + ] + testpairs = [ + "test/balanced/" + s + "/camera_00/" + imname + for s in sorted(os.listdir(self.root + "test/balanced")) + for imname in sorted( + os.listdir(self.root + "test/balanced/" + s + "/camera_00/") + ) + ] + assert len(trainpairs) == 228 and len(testpairs) == 191 + subtrainpairs = [p for p in trainpairs if any(s in p for s in trainseqs[:-2])] + subvalpairs = [p for p in trainpairs if any(s in p for s in trainseqs[-2:])] + # warning: if we do validation split, we should split scenes!!! + tosave = { + "train_balanced": trainpairs, + "test_balanced": testpairs, + "subtrain_balanced": subtrainpairs, + "subval_balanced": subvalpairs, + } + return tosave + + +class SpringDataset(StereoDataset): + + def _prepare_data(self): + self.name = "Spring" + self._set_root() + assert self.split in ["train", "test", "subtrain", "subval"] + self.pairname_to_Limgname = lambda pairname: osp.join( + self.root, pairname + ".png" + ) + self.pairname_to_Rimgname = ( + lambda pairname: osp.join(self.root, pairname + ".png") + .replace("frame_right", "") + .replace("frame_left", "frame_right") + .replace("", "frame_left") + ) + self.pairname_to_Ldispname = lambda pairname: ( + None + if pairname.startswith("test") + else osp.join(self.root, pairname + ".dsp5") + .replace("frame_left", "disp1_left") + .replace("frame_right", "disp1_right") + ) + self.pairname_to_str = lambda pairname: pairname + self.load_disparity = _read_hdf5_disp + + def _build_cache(self): + trainseqs = sorted(os.listdir(osp.join(self.root, "train"))) + trainpairs = [ + osp.join("train", s, "frame_left", f[:-4]) + for s in trainseqs + for f in sorted(os.listdir(osp.join(self.root, "train", s, "frame_left"))) + ] + testseqs = sorted(os.listdir(osp.join(self.root, "test"))) + testpairs = [ + osp.join("test", s, "frame_left", f[:-4]) + for s in testseqs + for f in sorted(os.listdir(osp.join(self.root, "test", s, "frame_left"))) + ] + testpairs += [p.replace("frame_left", "frame_right") for p in testpairs] + """maxnorm = {'0001': 32.88, '0002': 228.5, '0004': 298.2, '0005': 142.5, '0006': 113.6, '0007': 27.3, '0008': 554.5, '0009': 155.6, '0010': 126.1, '0011': 87.6, '0012': 303.2, '0013': 24.14, '0014': 82.56, '0015': 98.44, '0016': 156.9, '0017': 28.17, '0018': 21.03, '0020': 178.0, '0021': 58.06, '0022': 354.2, '0023': 8.79, '0024': 97.06, '0025': 55.16, '0026': 91.9, '0027': 156.6, '0030': 200.4, '0032': 58.66, '0033': 373.5, '0036': 149.4, '0037': 5.625, '0038': 37.0, '0039': 12.2, '0041': 453.5, '0043': 457.0, '0044': 379.5, '0045': 161.8, '0047': 105.44} # => let'use 0041""" + subtrainpairs = [p for p in trainpairs if p.split("/")[1] != "0041"] + subvalpairs = [p for p in trainpairs if p.split("/")[1] == "0041"] + assert ( + len(trainpairs) == 5000 + and len(testpairs) == 2000 + and len(subtrainpairs) == 4904 + and len(subvalpairs) == 96 + ), "incorrect parsing of pairs in Spring" + tosave = { + "train": trainpairs, + "test": testpairs, + "subtrain": subtrainpairs, + "subval": subvalpairs, + } + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 2 + assert prediction.dtype == np.float32 + outfile = ( + os.path.join(outdir, pairname + ".dsp5") + .replace("frame_left", "disp1_left") + .replace("frame_right", "disp1_right") + ) + os.makedirs(os.path.dirname(outfile), exist_ok=True) + writeDsp5File(prediction, outfile) + + def finalize_submission(self, outdir): + assert self.split == "test" + exe = "{self.root}/disp1_subsampling" + if os.path.isfile(exe): + cmd = f'cd "{outdir}/test"; {exe} .' + print(cmd) + os.system(cmd) + else: + print("Could not find disp1_subsampling executable for submission.") + print("Please download it and run:") + print(f'cd "{outdir}/test"; .') + + +class Kitti12Dataset(StereoDataset): + + def _prepare_data(self): + self.name = "Kitti12" + self._set_root() + assert self.split in ["train", "test"] + self.pairname_to_Limgname = lambda pairname: osp.join( + self.root, pairname + "_10.png" + ) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname.replace("/colored_0/", "/colored_1/") + "_10.png" + ) + self.pairname_to_Ldispname = ( + None + if self.split == "test" + else lambda pairname: osp.join( + self.root, pairname.replace("/colored_0/", "/disp_occ/") + "_10.png" + ) + ) + self.pairname_to_str = lambda pairname: pairname.replace("/colored_0/", "/") + self.load_disparity = _read_kitti_disp + + def _build_cache(self): + trainseqs = ["training/colored_0/%06d" % (i) for i in range(194)] + testseqs = ["testing/colored_0/%06d" % (i) for i in range(195)] + assert ( + len(trainseqs) == 194 and len(testseqs) == 195 + ), "incorrect parsing of pairs in Kitti12" + tosave = {"train": trainseqs, "test": testseqs} + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 2 + assert prediction.dtype == np.float32 + outfile = os.path.join(outdir, pairname.split("/")[-1] + "_10.png") + os.makedirs(os.path.dirname(outfile), exist_ok=True) + img = (prediction * 256).astype("uint16") + Image.fromarray(img).save(outfile) + + def finalize_submission(self, outdir): + assert self.split == "test" + cmd = f'cd {outdir}/; zip -r "kitti12_results.zip" .' + print(cmd) + os.system(cmd) + print(f"Done. Submission file at {outdir}/kitti12_results.zip") + + +class Kitti15Dataset(StereoDataset): + + def _prepare_data(self): + self.name = "Kitti15" + self._set_root() + assert self.split in ["train", "subtrain", "subval", "test"] + self.pairname_to_Limgname = lambda pairname: osp.join( + self.root, pairname + "_10.png" + ) + self.pairname_to_Rimgname = lambda pairname: osp.join( + self.root, pairname.replace("/image_2/", "/image_3/") + "_10.png" + ) + self.pairname_to_Ldispname = ( + None + if self.split == "test" + else lambda pairname: osp.join( + self.root, pairname.replace("/image_2/", "/disp_occ_0/") + "_10.png" + ) + ) + self.pairname_to_str = lambda pairname: pairname.replace("/image_2/", "/") + self.load_disparity = _read_kitti_disp + + def _build_cache(self): + trainseqs = ["training/image_2/%06d" % (i) for i in range(200)] + subtrainseqs = trainseqs[:-5] + subvalseqs = trainseqs[-5:] + testseqs = ["testing/image_2/%06d" % (i) for i in range(200)] + assert ( + len(trainseqs) == 200 + and len(subtrainseqs) == 195 + and len(subvalseqs) == 5 + and len(testseqs) == 200 + ), "incorrect parsing of pairs in Kitti15" + tosave = { + "train": trainseqs, + "subtrain": subtrainseqs, + "subval": subvalseqs, + "test": testseqs, + } + return tosave + + def submission_save_pairname(self, pairname, prediction, outdir, time): + assert prediction.ndim == 2 + assert prediction.dtype == np.float32 + outfile = os.path.join(outdir, "disp_0", pairname.split("/")[-1] + "_10.png") + os.makedirs(os.path.dirname(outfile), exist_ok=True) + img = (prediction * 256).astype("uint16") + Image.fromarray(img).save(outfile) + + def finalize_submission(self, outdir): + assert self.split == "test" + cmd = f'cd {outdir}/; zip -r "kitti15_results.zip" disp_0' + print(cmd) + os.system(cmd) + print(f"Done. Submission file at {outdir}/kitti15_results.zip") + + +### auxiliary functions + + +def _read_img(filename): + # convert to RGB for scene flow finalpass data + img = np.asarray(Image.open(filename).convert("RGB")) + return img + + +def _read_booster_disp(filename): + disp = np.load(filename) + disp[disp == 0.0] = np.inf + return disp + + +def _read_png_disp(filename, coef=1.0): + disp = np.asarray(Image.open(filename)) + disp = disp.astype(np.float32) / coef + disp[disp == 0.0] = np.inf + return disp + + +def _read_pfm_disp(filename): + disp = np.ascontiguousarray(_read_pfm(filename)[0]) + disp[disp <= 0] = ( + np.inf + ) # eg /nfs/data/ffs-3d/datasets/middlebury/2014/Shopvac-imperfect/disp0.pfm + return disp + + +def _read_npy_disp(filename): + return np.load(filename) + + +def _read_crestereo_disp(filename): + return _read_png_disp(filename, coef=32.0) + + +def _read_middlebury20052006_disp(filename): + return _read_png_disp(filename, coef=1.0) + + +def _read_kitti_disp(filename): + return _read_png_disp(filename, coef=256.0) + + +_read_sceneflow_disp = _read_pfm_disp +_read_eth3d_disp = _read_pfm_disp +_read_middlebury_disp = _read_pfm_disp +_read_carla_disp = _read_pfm_disp +_read_tartanair_disp = _read_npy_disp + + +def _read_hdf5_disp(filename): + disp = np.asarray(h5py.File(filename)["disparity"]) + disp[np.isnan(disp)] = np.inf # make invalid values as +inf + # disp[disp==0.0] = np.inf # make invalid values as +inf + return disp.astype(np.float32) + + +import re + + +def _read_pfm(file): + file = open(file, "rb") + + color = None + width = None + height = None + scale = None + endian = None + + header = file.readline().rstrip() + if header.decode("ascii") == "PF": + color = True + elif header.decode("ascii") == "Pf": + color = False + else: + raise Exception("Not a PFM file.") + + dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii")) + if dim_match: + width, height = list(map(int, dim_match.groups())) + else: + raise Exception("Malformed PFM header.") + + scale = float(file.readline().decode("ascii").rstrip()) + if scale < 0: # little-endian + endian = "<" + scale = -scale + else: + endian = ">" # big-endian + + data = np.fromfile(file, endian + "f") + shape = (height, width, 3) if color else (height, width) + + data = np.reshape(data, shape) + data = np.flipud(data) + return data, scale + + +def writePFM(file, image, scale=1): + file = open(file, "wb") + + color = None + + if image.dtype.name != "float32": + raise Exception("Image dtype must be float32.") + + image = np.flipud(image) + + if len(image.shape) == 3 and image.shape[2] == 3: # color image + color = True + elif ( + len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1 + ): # greyscale + color = False + else: + raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.") + + file.write("PF\n" if color else "Pf\n".encode()) + file.write("%d %d\n".encode() % (image.shape[1], image.shape[0])) + + endian = image.dtype.byteorder + + if endian == "<" or endian == "=" and sys.byteorder == "little": + scale = -scale + + file.write("%f\n".encode() % scale) + + image.tofile(file) + + +def writeDsp5File(disp, filename): + with h5py.File(filename, "w") as f: + f.create_dataset("disparity", data=disp, compression="gzip", compression_opts=5) + + +# disp visualization + + +def vis_disparity(disp, m=None, M=None): + if m is None: + m = disp.min() + if M is None: + M = disp.max() + disp_vis = (disp - m) / (M - m) * 255.0 + disp_vis = disp_vis.astype("uint8") + disp_vis = cv2.applyColorMap(disp_vis, cv2.COLORMAP_INFERNO) + return disp_vis + + +# dataset getter + + +def get_train_dataset_stereo(dataset_str, augmentor=True, crop_size=None): + dataset_str = dataset_str.replace("(", "Dataset(") + if augmentor: + dataset_str = dataset_str.replace(")", ", augmentor=True)") + if crop_size is not None: + dataset_str = dataset_str.replace( + ")", ", crop_size={:s})".format(str(crop_size)) + ) + return eval(dataset_str) + + +def get_test_datasets_stereo(dataset_str): + dataset_str = dataset_str.replace("(", "Dataset(") + return [eval(s) for s in dataset_str.split("+")] diff --git a/croco/stereoflow/download_model.sh b/croco/stereoflow/download_model.sh new file mode 100644 index 0000000000000000000000000000000000000000..533119609108c5ec3c22ff79b10e9215c1ac5098 --- /dev/null +++ b/croco/stereoflow/download_model.sh @@ -0,0 +1,12 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +model=$1 +outfile="stereoflow_models/${model}" +if [[ ! -f $outfile ]] +then + mkdir -p stereoflow_models/; + wget https://download.europe.naverlabs.com/ComputerVision/CroCo/StereoFlow_models/$1 -P stereoflow_models/; +else + echo "Model ${model} already downloaded in ${outfile}." +fi \ No newline at end of file diff --git a/croco/stereoflow/engine.py b/croco/stereoflow/engine.py new file mode 100644 index 0000000000000000000000000000000000000000..9736f2ab7c895e032893f60949baf87131a49b6e --- /dev/null +++ b/croco/stereoflow/engine.py @@ -0,0 +1,367 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Main function for training one epoch or testing +# -------------------------------------------------------- + +import math +import sys +from typing import Iterable +import numpy as np +import torch +import torchvision + +from utils import misc as misc + + +def split_prediction_conf(predictions, with_conf=False): + if not with_conf: + return predictions, None + conf = predictions[:, -1:, :, :] + predictions = predictions[:, :-1, :, :] + return predictions, conf + + +def train_one_epoch( + model: torch.nn.Module, + criterion: torch.nn.Module, + metrics: torch.nn.Module, + data_loader: Iterable, + optimizer: torch.optim.Optimizer, + device: torch.device, + epoch: int, + loss_scaler, + log_writer=None, + print_freq=20, + args=None, +): + model.train(True) + metric_logger = misc.MetricLogger(delimiter=" ") + metric_logger.add_meter("lr", misc.SmoothedValue(window_size=1, fmt="{value:.6f}")) + header = "Epoch: [{}]".format(epoch) + + accum_iter = args.accum_iter + + optimizer.zero_grad() + + details = {} + + if log_writer is not None: + print("log_dir: {}".format(log_writer.log_dir)) + + if args.img_per_epoch: + iter_per_epoch = args.img_per_epoch // args.batch_size + int( + args.img_per_epoch % args.batch_size > 0 + ) + assert ( + len(data_loader) >= iter_per_epoch + ), "Dataset is too small for so many iterations" + len_data_loader = iter_per_epoch + else: + len_data_loader, iter_per_epoch = len(data_loader), None + + for data_iter_step, (image1, image2, gt, pairname) in enumerate( + metric_logger.log_every( + data_loader, print_freq, header, max_iter=iter_per_epoch + ) + ): + + image1 = image1.to(device, non_blocking=True) + image2 = image2.to(device, non_blocking=True) + gt = gt.to(device, non_blocking=True) + + # we use a per iteration (instead of per epoch) lr scheduler + if data_iter_step % accum_iter == 0: + misc.adjust_learning_rate( + optimizer, data_iter_step / len_data_loader + epoch, args + ) + + with torch.cuda.amp.autocast(enabled=bool(args.amp)): + prediction = model(image1, image2) + prediction, conf = split_prediction_conf(prediction, criterion.with_conf) + batch_metrics = metrics(prediction.detach(), gt) + loss = ( + criterion(prediction, gt) + if conf is None + else criterion(prediction, gt, conf) + ) + + loss_value = loss.item() + if not math.isfinite(loss_value): + print("Loss is {}, stopping training".format(loss_value)) + sys.exit(1) + + loss /= accum_iter + loss_scaler( + loss, + optimizer, + parameters=model.parameters(), + update_grad=(data_iter_step + 1) % accum_iter == 0, + ) + if (data_iter_step + 1) % accum_iter == 0: + optimizer.zero_grad() + + torch.cuda.synchronize() + + metric_logger.update(loss=loss_value) + for k, v in batch_metrics.items(): + metric_logger.update(**{k: v.item()}) + lr = optimizer.param_groups[0]["lr"] + metric_logger.update(lr=lr) + + # if args.dsitributed: loss_value_reduce = misc.all_reduce_mean(loss_value) + time_to_log = (data_iter_step + 1) % ( + args.tboard_log_step * accum_iter + ) == 0 or data_iter_step == len_data_loader - 1 + loss_value_reduce = misc.all_reduce_mean(loss_value) + if log_writer is not None and time_to_log: + epoch_1000x = int((data_iter_step / len_data_loader + epoch) * 1000) + # We use epoch_1000x as the x-axis in tensorboard. This calibrates different curves when batch size changes. + log_writer.add_scalar("train/loss", loss_value_reduce, epoch_1000x) + log_writer.add_scalar("lr", lr, epoch_1000x) + for k, v in batch_metrics.items(): + log_writer.add_scalar("train/" + k, v.item(), epoch_1000x) + + # gather the stats from all processes + # if args.distributed: metric_logger.synchronize_between_processes() + print("Averaged stats:", metric_logger) + return {k: meter.global_avg for k, meter in metric_logger.meters.items()} + + +@torch.no_grad() +def validate_one_epoch( + model: torch.nn.Module, + criterion: torch.nn.Module, + metrics: torch.nn.Module, + data_loaders: list[Iterable], + device: torch.device, + epoch: int, + log_writer=None, + args=None, +): + + model.eval() + metric_loggers = [] + header = "Epoch: [{}]".format(epoch) + print_freq = 20 + + conf_mode = args.tile_conf_mode + crop = args.crop + + if log_writer is not None: + print("log_dir: {}".format(log_writer.log_dir)) + + results = {} + dnames = [] + image1, image2, gt, prediction = None, None, None, None + for didx, data_loader in enumerate(data_loaders): + dname = str(data_loader.dataset) + dnames.append(dname) + metric_loggers.append(misc.MetricLogger(delimiter=" ")) + for data_iter_step, (image1, image2, gt, pairname) in enumerate( + metric_loggers[didx].log_every(data_loader, print_freq, header) + ): + image1 = image1.to(device, non_blocking=True) + image2 = image2.to(device, non_blocking=True) + gt = gt.to(device, non_blocking=True) + if dname.startswith("Spring"): + assert ( + gt.size(2) == image1.size(2) * 2 + and gt.size(3) == image1.size(3) * 2 + ) + gt = ( + gt[:, :, 0::2, 0::2] + + gt[:, :, 0::2, 1::2] + + gt[:, :, 1::2, 0::2] + + gt[:, :, 1::2, 1::2] + ) / 4.0 # we approximate the gt based on the 2x upsampled ones + + with torch.inference_mode(): + prediction, tiled_loss, c = tiled_pred( + model, + criterion, + image1, + image2, + gt, + conf_mode=conf_mode, + overlap=args.val_overlap, + crop=crop, + with_conf=criterion.with_conf, + ) + batch_metrics = metrics(prediction.detach(), gt) + loss = ( + criterion(prediction.detach(), gt) + if not criterion.with_conf + else criterion(prediction.detach(), gt, c) + ) + loss_value = loss.item() + metric_loggers[didx].update(loss_tiled=tiled_loss.item()) + metric_loggers[didx].update(**{f"loss": loss_value}) + for k, v in batch_metrics.items(): + metric_loggers[didx].update(**{dname + "_" + k: v.item()}) + + results = { + k: meter.global_avg for ml in metric_loggers for k, meter in ml.meters.items() + } + if len(dnames) > 1: + for k in batch_metrics.keys(): + results["AVG_" + k] = sum( + results[dname + "_" + k] for dname in dnames + ) / len(dnames) + + if log_writer is not None: + epoch_1000x = int((1 + epoch) * 1000) + for k, v in results.items(): + log_writer.add_scalar("val/" + k, v, epoch_1000x) + + print("Averaged stats:", results) + return results + + +import torch.nn.functional as F + + +def _resize_img(img, new_size): + return F.interpolate(img, size=new_size, mode="bicubic", align_corners=False) + + +def _resize_stereo_or_flow(data, new_size): + assert data.ndim == 4 + assert data.size(1) in [1, 2] + scale_x = new_size[1] / float(data.size(3)) + out = F.interpolate(data, size=new_size, mode="bicubic", align_corners=False) + out[:, 0, :, :] *= scale_x + if out.size(1) == 2: + scale_y = new_size[0] / float(data.size(2)) + out[:, 1, :, :] *= scale_y + print(scale_x, new_size, data.shape) + return out + + +@torch.no_grad() +def tiled_pred( + model, + criterion, + img1, + img2, + gt, + overlap=0.5, + bad_crop_thr=0.05, + downscale=False, + crop=512, + ret="loss", + conf_mode="conf_expsigmoid_10_5", + with_conf=False, + return_time=False, +): + + # for each image, we are going to run inference on many overlapping patches + # then, all predictions will be weighted-averaged + if gt is not None: + B, C, H, W = gt.shape + else: + B, _, H, W = img1.shape + C = model.head.num_channels - int(with_conf) + win_height, win_width = crop[0], crop[1] + + # upscale to be larger than the crop + do_change_scale = H < win_height or W < win_width + if do_change_scale: + upscale_factor = max(win_width / W, win_height / W) + original_size = (H, W) + new_size = (round(H * upscale_factor), round(W * upscale_factor)) + img1 = _resize_img(img1, new_size) + img2 = _resize_img(img2, new_size) + # resize gt just for the computation of tiled losses + if gt is not None: + gt = _resize_stereo_or_flow(gt, new_size) + H, W = img1.shape[2:4] + + if conf_mode.startswith("conf_expsigmoid_"): # conf_expsigmoid_30_10 + beta, betasigmoid = map(float, conf_mode[len("conf_expsigmoid_") :].split("_")) + elif conf_mode.startswith("conf_expbeta"): # conf_expbeta3 + beta = float(conf_mode[len("conf_expbeta") :]) + else: + raise NotImplementedError(f"conf_mode {conf_mode} is not implemented") + + def crop_generator(): + for sy in _overlapping(H, win_height, overlap): + for sx in _overlapping(W, win_width, overlap): + yield sy, sx, sy, sx, True + + # keep track of weighted sum of prediction*weights and weights + accu_pred = img1.new_zeros( + (B, C, H, W) + ) # accumulate the weighted sum of predictions + accu_conf = img1.new_zeros((B, H, W)) + 1e-16 # accumulate the weights + accu_c = img1.new_zeros( + (B, H, W) + ) # accumulate the weighted sum of confidences ; not so useful except for computing some losses + + tiled_losses = [] + + if return_time: + start = torch.cuda.Event(enable_timing=True) + end = torch.cuda.Event(enable_timing=True) + start.record() + + for sy1, sx1, sy2, sx2, aligned in crop_generator(): + # compute optical flow there + pred = model(_crop(img1, sy1, sx1), _crop(img2, sy2, sx2)) + pred, predconf = split_prediction_conf(pred, with_conf=with_conf) + + if gt is not None: + gtcrop = _crop(gt, sy1, sx1) + if criterion is not None and gt is not None: + tiled_losses.append( + criterion(pred, gtcrop).item() + if predconf is None + else criterion(pred, gtcrop, predconf).item() + ) + + if conf_mode.startswith("conf_expsigmoid_"): + conf = torch.exp( + -beta * 2 * (torch.sigmoid(predconf / betasigmoid) - 0.5) + ).view(B, win_height, win_width) + elif conf_mode.startswith("conf_expbeta"): + conf = torch.exp(-beta * predconf).view(B, win_height, win_width) + else: + raise NotImplementedError + + accu_pred[..., sy1, sx1] += pred * conf[:, None, :, :] + accu_conf[..., sy1, sx1] += conf + accu_c[..., sy1, sx1] += predconf.view(B, win_height, win_width) * conf + + pred = accu_pred / accu_conf[:, None, :, :] + c = accu_c / accu_conf + assert not torch.any(torch.isnan(pred)) + + if return_time: + end.record() + torch.cuda.synchronize() + time = start.elapsed_time(end) / 1000.0 # this was in milliseconds + + if do_change_scale: + pred = _resize_stereo_or_flow(pred, original_size) + + if return_time: + return pred, torch.mean(torch.tensor(tiled_losses)), c, time + return pred, torch.mean(torch.tensor(tiled_losses)), c + + +def _overlapping(total, window, overlap=0.5): + assert total >= window and 0 <= overlap < 1, (total, window, overlap) + num_windows = 1 + int(np.ceil((total - window) / ((1 - overlap) * window))) + offsets = np.linspace(0, total - window, num_windows).round().astype(int) + yield from (slice(x, x + window) for x in offsets) + + +def _crop(img, sy, sx): + B, THREE, H, W = img.shape + if 0 <= sy.start and sy.stop <= H and 0 <= sx.start and sx.stop <= W: + return img[:, :, sy, sx] + l, r = max(0, -sx.start), max(0, sx.stop - W) + t, b = max(0, -sy.start), max(0, sy.stop - H) + img = torch.nn.functional.pad(img, (l, r, t, b), mode="constant") + return img[:, :, slice(sy.start + t, sy.stop + t), slice(sx.start + l, sx.stop + l)] diff --git a/croco/stereoflow/test.py b/croco/stereoflow/test.py new file mode 100644 index 0000000000000000000000000000000000000000..15dcf769169d460b716b05acb290340b6a197a6d --- /dev/null +++ b/croco/stereoflow/test.py @@ -0,0 +1,303 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Main test function +# -------------------------------------------------------- + +import os +import argparse +import pickle +from PIL import Image +import numpy as np +from tqdm import tqdm + +import torch +from torch.utils.data import DataLoader + +import utils.misc as misc +from models.croco_downstream import CroCoDownstreamBinocular +from models.head_downstream import PixelwiseTaskWithDPT + +from stereoflow.criterion import * +from stereoflow.datasets_stereo import get_test_datasets_stereo +from stereoflow.datasets_flow import get_test_datasets_flow +from stereoflow.engine import tiled_pred + +from stereoflow.datasets_stereo import vis_disparity +from stereoflow.datasets_flow import flowToColor + + +def get_args_parser(): + parser = argparse.ArgumentParser("Test CroCo models on stereo/flow", add_help=False) + # important argument + parser.add_argument( + "--model", required=True, type=str, help="Path to the model to evaluate" + ) + parser.add_argument( + "--dataset", + required=True, + type=str, + help="test dataset (there can be multiple dataset separated by a +)", + ) + # tiling + parser.add_argument( + "--tile_conf_mode", + type=str, + default="", + help="Weights for the tiling aggregation based on confidence (empty means use the formula from the loaded checkpoint", + ) + parser.add_argument( + "--tile_overlap", type=float, default=0.7, help="overlap between tiles" + ) + # save (it will automatically go to _/_) + parser.add_argument( + "--save", + type=str, + nargs="+", + default=[], + help="what to save: \ + metrics (pickle file), \ + pred (raw prediction save as torch tensor), \ + visu (visualization in png of each prediction), \ + err10 (visualization in png of the error clamp at 10 for each prediction), \ + submission (submission file)", + ) + # other (no impact) + parser.add_argument("--num_workers", default=4, type=int) + return parser + + +def _load_model_and_criterion(model_path, do_load_metrics, device): + print("loading model from", model_path) + assert os.path.isfile(model_path) + ckpt = torch.load(model_path, "cpu") + + ckpt_args = ckpt["args"] + task = ckpt_args.task + tile_conf_mode = ckpt_args.tile_conf_mode + num_channels = {"stereo": 1, "flow": 2}[task] + with_conf = eval(ckpt_args.criterion).with_conf + if with_conf: + num_channels += 1 + print("head: PixelwiseTaskWithDPT()") + head = PixelwiseTaskWithDPT() + head.num_channels = num_channels + print("croco_args:", ckpt_args.croco_args) + model = CroCoDownstreamBinocular(head, **ckpt_args.croco_args) + msg = model.load_state_dict(ckpt["model"], strict=True) + model.eval() + model = model.to(device) + + if do_load_metrics: + if task == "stereo": + metrics = StereoDatasetMetrics().to(device) + else: + metrics = FlowDatasetMetrics().to(device) + else: + metrics = None + + return model, metrics, ckpt_args.crop, with_conf, task, tile_conf_mode + + +def _save_batch( + pred, gt, pairnames, dataset, task, save, outdir, time, submission_dir=None +): + + for i in range(len(pairnames)): + + pairname = ( + eval(pairnames[i]) if pairnames[i].startswith("(") else pairnames[i] + ) # unbatch pairname + fname = os.path.join(outdir, dataset.pairname_to_str(pairname)) + os.makedirs(os.path.dirname(fname), exist_ok=True) + + predi = pred[i, ...] + if gt is not None: + gti = gt[i, ...] + + if "pred" in save: + torch.save(predi.squeeze(0).cpu(), fname + "_pred.pth") + + if "visu" in save: + if task == "stereo": + disparity = predi.permute((1, 2, 0)).squeeze(2).cpu().numpy() + m, M = None + if gt is not None: + mask = torch.isfinite(gti) + m = gt[mask].min() + M = gt[mask].max() + img_disparity = vis_disparity(disparity, m=m, M=M) + Image.fromarray(img_disparity).save(fname + "_pred.png") + else: + # normalize flowToColor according to the maxnorm of gt (or prediction if not available) + flowNorm = ( + torch.sqrt( + torch.sum((gti if gt is not None else predi) ** 2, dim=0) + ) + .max() + .item() + ) + imgflow = flowToColor( + predi.permute((1, 2, 0)).cpu().numpy(), maxflow=flowNorm + ) + Image.fromarray(imgflow).save(fname + "_pred.png") + + if "err10" in save: + assert gt is not None + L2err = torch.sqrt(torch.sum((gti - predi) ** 2, dim=0)) + valid = torch.isfinite(gti[0, :, :]) + L2err[~valid] = 0.0 + L2err = torch.clamp(L2err, max=10.0) + red = (L2err * 255.0 / 10.0).to(dtype=torch.uint8)[:, :, None] + zer = torch.zeros_like(red) + imgerr = torch.cat((red, zer, zer), dim=2).cpu().numpy() + Image.fromarray(imgerr).save(fname + "_err10.png") + + if "submission" in save: + assert submission_dir is not None + predi_np = ( + predi.permute(1, 2, 0).squeeze(2).cpu().numpy() + ) # transform into HxWx2 for flow or HxW for stereo + dataset.submission_save_pairname(pairname, predi_np, submission_dir, time) + + +def main(args): + + # load the pretrained model and metrics + device = ( + torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu") + ) + model, metrics, cropsize, with_conf, task, tile_conf_mode = ( + _load_model_and_criterion(args.model, "metrics" in args.save, device) + ) + if args.tile_conf_mode == "": + args.tile_conf_mode = tile_conf_mode + + # load the datasets + datasets = ( + get_test_datasets_stereo if task == "stereo" else get_test_datasets_flow + )(args.dataset) + dataloaders = [ + DataLoader( + dataset, + batch_size=1, + shuffle=False, + num_workers=args.num_workers, + pin_memory=True, + drop_last=False, + ) + for dataset in datasets + ] + + # run + for i, dataloader in enumerate(dataloaders): + dataset = datasets[i] + dstr = args.dataset.split("+")[i] + + outdir = args.model + "_" + misc.filename(dstr) + if "metrics" in args.save and len(args.save) == 1: + fname = os.path.join( + outdir, f"conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}.pkl" + ) + if os.path.isfile(fname) and len(args.save) == 1: + print(" metrics already compute in " + fname) + with open(fname, "rb") as fid: + results = pickle.load(fid) + for k, v in results.items(): + print("{:s}: {:.3f}".format(k, v)) + continue + + if "submission" in args.save: + dirname = ( + f"submission_conf_{args.tile_conf_mode}_overlap_{args.tile_overlap}" + ) + submission_dir = os.path.join(outdir, dirname) + else: + submission_dir = None + + print("") + print("saving {:s} in {:s}".format("+".join(args.save), outdir)) + print(repr(dataset)) + + if metrics is not None: + metrics.reset() + + for data_iter_step, (image1, image2, gt, pairnames) in enumerate( + tqdm(dataloader) + ): + + do_flip = ( + task == "stereo" + and dstr.startswith("Spring") + and any("right" in p for p in pairnames) + ) # we flip the images and will flip the prediction after as we assume img1 is on the left + + image1 = image1.to(device, non_blocking=True) + image2 = image2.to(device, non_blocking=True) + gt = ( + gt.to(device, non_blocking=True) if gt.numel() > 0 else None + ) # special case for test time + if do_flip: + assert all("right" in p for p in pairnames) + image1 = image1.flip( + dims=[3] + ) # this is already the right frame, let's flip it + image2 = image2.flip(dims=[3]) + gt = gt # that is ok + + with torch.inference_mode(): + pred, _, _, time = tiled_pred( + model, + None, + image1, + image2, + None if dataset.name == "Spring" else gt, + conf_mode=args.tile_conf_mode, + overlap=args.tile_overlap, + crop=cropsize, + with_conf=with_conf, + return_time=True, + ) + + if do_flip: + pred = pred.flip(dims=[3]) + + if metrics is not None: + metrics.add_batch(pred, gt) + + if any(k in args.save for k in ["pred", "visu", "err10", "submission"]): + _save_batch( + pred, + gt, + pairnames, + dataset, + task, + args.save, + outdir, + time, + submission_dir=submission_dir, + ) + + # print + if metrics is not None: + results = metrics.get_results() + for k, v in results.items(): + print("{:s}: {:.3f}".format(k, v)) + + # save if needed + if "metrics" in args.save: + os.makedirs(os.path.dirname(fname), exist_ok=True) + with open(fname, "wb") as fid: + pickle.dump(results, fid) + print("metrics saved in", fname) + + # finalize submission if needed + if "submission" in args.save: + dataset.finalize_submission(submission_dir) + + +if __name__ == "__main__": + args = get_args_parser() + args = args.parse_args() + main(args) diff --git a/croco/stereoflow/train.py b/croco/stereoflow/train.py new file mode 100644 index 0000000000000000000000000000000000000000..c349cb479267648cad4d8b4c282dafd7a8896076 --- /dev/null +++ b/croco/stereoflow/train.py @@ -0,0 +1,455 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). + +# -------------------------------------------------------- +# Main training function +# -------------------------------------------------------- + +import argparse +import datetime +import json +import numpy as np +import os +import sys +import time + +import torch +import torch.distributed as dist +import torch.backends.cudnn as cudnn +from torch.utils.tensorboard import SummaryWriter +import torchvision.transforms as transforms +import torchvision.datasets as datasets +from torch.utils.data import DataLoader + +import utils +import utils.misc as misc +from utils.misc import NativeScalerWithGradNormCount as NativeScaler +from models.croco_downstream import CroCoDownstreamBinocular, croco_args_from_ckpt +from models.pos_embed import interpolate_pos_embed +from models.head_downstream import PixelwiseTaskWithDPT + +from stereoflow.datasets_stereo import ( + get_train_dataset_stereo, + get_test_datasets_stereo, +) +from stereoflow.datasets_flow import get_train_dataset_flow, get_test_datasets_flow +from stereoflow.engine import train_one_epoch, validate_one_epoch +from stereoflow.criterion import * + + +def get_args_parser(): + # prepare subparsers + parser = argparse.ArgumentParser( + "Finetuning CroCo models on stereo or flow", add_help=False + ) + subparsers = parser.add_subparsers( + title="Task (stereo or flow)", dest="task", required=True + ) + parser_stereo = subparsers.add_parser("stereo", help="Training stereo model") + parser_flow = subparsers.add_parser("flow", help="Training flow model") + + def add_arg( + name_or_flags, default=None, default_stereo=None, default_flow=None, **kwargs + ): + if default is not None: + assert ( + default_stereo is None and default_flow is None + ), "setting default makes default_stereo and default_flow disabled" + parser_stereo.add_argument( + name_or_flags, + default=default if default is not None else default_stereo, + **kwargs, + ) + parser_flow.add_argument( + name_or_flags, + default=default if default is not None else default_flow, + **kwargs, + ) + + # output dir + add_arg( + "--output_dir", + required=True, + type=str, + help="path where to save, if empty, automatically created", + ) + # model + add_arg( + "--crop", + type=int, + nargs="+", + default_stereo=[352, 704], + default_flow=[320, 384], + help="size of the random image crops used during training.", + ) + add_arg( + "--pretrained", + required=True, + type=str, + help="Load pretrained model (required as croco arguments come from there)", + ) + # criterion + add_arg( + "--criterion", + default_stereo="LaplacianLossBounded2()", + default_flow="LaplacianLossBounded()", + type=str, + help="string to evaluate to get criterion", + ) + add_arg("--bestmetric", default_stereo="avgerr", default_flow="EPE", type=str) + # dataset + add_arg("--dataset", type=str, required=True, help="training set") + # training + add_arg("--seed", default=0, type=int, help="seed") + add_arg( + "--batch_size", + default_stereo=6, + default_flow=8, + type=int, + help="Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus", + ) + add_arg("--epochs", default=32, type=int, help="number of training epochs") + add_arg( + "--img_per_epoch", + type=int, + default=None, + help="Fix the number of images seen in an epoch (None means use all training pairs)", + ) + add_arg( + "--accum_iter", + default=1, + type=int, + help="Accumulate gradient iterations (for increasing the effective batch size under memory constraints)", + ) + add_arg( + "--weight_decay", type=float, default=0.05, help="weight decay (default: 0.05)" + ) + add_arg( + "--lr", + type=float, + default_stereo=3e-5, + default_flow=2e-5, + metavar="LR", + help="learning rate (absolute lr)", + ) + add_arg( + "--min_lr", + type=float, + default=0.0, + metavar="LR", + help="lower lr bound for cyclic schedulers that hit 0", + ) + add_arg( + "--warmup_epochs", type=int, default=1, metavar="N", help="epochs to warmup LR" + ) + add_arg( + "--optimizer", + default="AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))", + type=str, + help="Optimizer from torch.optim [ default: AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95)) ]", + ) + add_arg( + "--amp", + default=0, + type=int, + choices=[0, 1], + help="enable automatic mixed precision training", + ) + # validation + add_arg( + "--val_dataset", + type=str, + default="", + help="Validation sets, multiple separated by + (empty string means that no validation is performed)", + ) + add_arg( + "--tile_conf_mode", + type=str, + default_stereo="conf_expsigmoid_15_3", + default_flow="conf_expsigmoid_10_5", + help="Weights for tile aggregation", + ) + add_arg( + "--val_overlap", default=0.7, type=float, help="Overlap value for the tiling" + ) + # others + add_arg("--num_workers", default=8, type=int) + add_arg("--eval_every", type=int, default=1, help="Val loss evaluation frequency") + add_arg("--save_every", type=int, default=1, help="Save checkpoint frequency") + add_arg( + "--start_from", + type=str, + default=None, + help="Start training using weights from an other model (eg for finetuning)", + ) + add_arg( + "--tboard_log_step", + type=int, + default=100, + help="Log to tboard every so many steps", + ) + add_arg( + "--dist_url", default="env://", help="url used to set up distributed training" + ) + + return parser + + +def main(args): + misc.init_distributed_mode(args) + global_rank = misc.get_rank() + num_tasks = misc.get_world_size() + + assert os.path.isfile(args.pretrained) + print("output_dir: " + args.output_dir) + os.makedirs(args.output_dir, exist_ok=True) + + # fix the seed for reproducibility + seed = args.seed + misc.get_rank() + torch.manual_seed(seed) + np.random.seed(seed) + cudnn.benchmark = True + + # Metrics / criterion + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + metrics = (StereoMetrics if args.task == "stereo" else FlowMetrics)().to(device) + criterion = eval(args.criterion).to(device) + print("Criterion: ", args.criterion) + + # Prepare model + assert os.path.isfile(args.pretrained) + ckpt = torch.load(args.pretrained, "cpu") + croco_args = croco_args_from_ckpt(ckpt) + croco_args["img_size"] = (args.crop[0], args.crop[1]) + print("Croco args: " + str(croco_args)) + args.croco_args = croco_args # saved for test time + # prepare head + num_channels = {"stereo": 1, "flow": 2}[args.task] + if criterion.with_conf: + num_channels += 1 + print(f"Building head PixelwiseTaskWithDPT() with {num_channels} channel(s)") + head = PixelwiseTaskWithDPT() + head.num_channels = num_channels + # build model and load pretrained weights + model = CroCoDownstreamBinocular(head, **croco_args) + interpolate_pos_embed(model, ckpt["model"]) + msg = model.load_state_dict(ckpt["model"], strict=False) + print(msg) + + total_params = sum(p.numel() for p in model.parameters()) + total_params_trainable = sum( + p.numel() for p in model.parameters() if p.requires_grad + ) + print(f"Total params: {total_params}") + print(f"Total params trainable: {total_params_trainable}") + model_without_ddp = model.to(device) + + eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size() + print("lr: %.2e" % args.lr) + print("accumulate grad iterations: %d" % args.accum_iter) + print("effective batch size: %d" % eff_batch_size) + + if args.distributed: + model = torch.nn.parallel.DistributedDataParallel( + model, device_ids=[args.gpu], static_graph=True + ) + model_without_ddp = model.module + + # following timm: set wd as 0 for bias and norm layers + param_groups = misc.get_parameter_groups(model_without_ddp, args.weight_decay) + optimizer = eval(f"torch.optim.{args.optimizer}") + print(optimizer) + loss_scaler = NativeScaler() + + # automatic restart + last_ckpt_fname = os.path.join(args.output_dir, f"checkpoint-last.pth") + args.resume = last_ckpt_fname if os.path.isfile(last_ckpt_fname) else None + + if not args.resume and args.start_from: + print(f"Starting from an other model's weights: {args.start_from}") + best_so_far = None + args.start_epoch = 0 + ckpt = torch.load(args.start_from, "cpu") + msg = model_without_ddp.load_state_dict(ckpt["model"], strict=False) + print(msg) + else: + best_so_far = misc.load_model( + args=args, + model_without_ddp=model_without_ddp, + optimizer=optimizer, + loss_scaler=loss_scaler, + ) + + if best_so_far is None: + best_so_far = np.inf + + # tensorboard + log_writer = None + if global_rank == 0 and args.output_dir is not None: + log_writer = SummaryWriter( + log_dir=args.output_dir, purge_step=args.start_epoch * 1000 + ) + + # dataset and loader + print("Building Train Data loader for dataset: ", args.dataset) + train_dataset = ( + get_train_dataset_stereo if args.task == "stereo" else get_train_dataset_flow + )(args.dataset, crop_size=args.crop) + + def _print_repr_dataset(d): + if isinstance(d, torch.utils.data.dataset.ConcatDataset): + for dd in d.datasets: + _print_repr_dataset(dd) + else: + print(repr(d)) + + _print_repr_dataset(train_dataset) + print(" total length:", len(train_dataset)) + if args.distributed: + sampler_train = torch.utils.data.DistributedSampler( + train_dataset, num_replicas=num_tasks, rank=global_rank, shuffle=True + ) + else: + sampler_train = torch.utils.data.RandomSampler(train_dataset) + data_loader_train = torch.utils.data.DataLoader( + train_dataset, + sampler=sampler_train, + batch_size=args.batch_size, + num_workers=args.num_workers, + pin_memory=True, + drop_last=True, + ) + if args.val_dataset == "": + data_loaders_val = None + else: + print("Building Val Data loader for datasets: ", args.val_dataset) + val_datasets = ( + get_test_datasets_stereo + if args.task == "stereo" + else get_test_datasets_flow + )(args.val_dataset) + for val_dataset in val_datasets: + print(repr(val_dataset)) + data_loaders_val = [ + DataLoader( + val_dataset, + batch_size=1, + shuffle=False, + num_workers=args.num_workers, + pin_memory=True, + drop_last=False, + ) + for val_dataset in val_datasets + ] + bestmetric = ( + "AVG_" + if len(data_loaders_val) > 1 + else str(data_loaders_val[0].dataset) + "_" + ) + args.bestmetric + + print(f"Start training for {args.epochs} epochs") + start_time = time.time() + # Training Loop + for epoch in range(args.start_epoch, args.epochs): + + if args.distributed: + data_loader_train.sampler.set_epoch(epoch) + + # Train + epoch_start = time.time() + train_stats = train_one_epoch( + model, + criterion, + metrics, + data_loader_train, + optimizer, + device, + epoch, + loss_scaler, + log_writer=log_writer, + args=args, + ) + epoch_time = time.time() - epoch_start + + if args.distributed: + dist.barrier() + + # Validation (current naive implementation runs the validation on every gpu ... not smart ...) + if ( + data_loaders_val is not None + and args.eval_every > 0 + and (epoch + 1) % args.eval_every == 0 + ): + val_epoch_start = time.time() + val_stats = validate_one_epoch( + model, + criterion, + metrics, + data_loaders_val, + device, + epoch, + log_writer=log_writer, + args=args, + ) + val_epoch_time = time.time() - val_epoch_start + + val_best = val_stats[bestmetric] + + # Save best of all + if val_best <= best_so_far: + best_so_far = val_best + misc.save_model( + args=args, + model_without_ddp=model_without_ddp, + optimizer=optimizer, + loss_scaler=loss_scaler, + epoch=epoch, + best_so_far=best_so_far, + fname="best", + ) + + log_stats = { + **{f"train_{k}": v for k, v in train_stats.items()}, + "epoch": epoch, + **{f"val_{k}": v for k, v in val_stats.items()}, + } + else: + log_stats = { + **{f"train_{k}": v for k, v in train_stats.items()}, + "epoch": epoch, + } + + if args.distributed: + dist.barrier() + + # Save stuff + if args.output_dir and ( + (epoch + 1) % args.save_every == 0 or epoch + 1 == args.epochs + ): + misc.save_model( + args=args, + model_without_ddp=model_without_ddp, + optimizer=optimizer, + loss_scaler=loss_scaler, + epoch=epoch, + best_so_far=best_so_far, + fname="last", + ) + + if args.output_dir: + if log_writer is not None: + log_writer.flush() + with open( + os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8" + ) as f: + f.write(json.dumps(log_stats) + "\n") + + total_time = time.time() - start_time + total_time_str = str(datetime.timedelta(seconds=int(total_time))) + print("Training time {}".format(total_time_str)) + + +if __name__ == "__main__": + args = get_args_parser() + args = args.parse_args() + main(args) diff --git a/croco/utils/__pycache__/misc.cpython-310.pyc b/croco/utils/__pycache__/misc.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0bbec5bee96ef2fc2b2676fc9298de4abe5b3694 Binary files /dev/null and b/croco/utils/__pycache__/misc.cpython-310.pyc differ diff --git a/croco/utils/__pycache__/misc.cpython-311.pyc b/croco/utils/__pycache__/misc.cpython-311.pyc new file mode 100644 index 0000000000000000000000000000000000000000..57e5e5513a8bb1ccab72cb6d74054a45f7e982f5 Binary files /dev/null and b/croco/utils/__pycache__/misc.cpython-311.pyc differ diff --git a/croco/utils/__pycache__/misc.cpython-312.pyc b/croco/utils/__pycache__/misc.cpython-312.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a0e3dd10444e954e27c5b3a95ec0a96f5de3f7da Binary files /dev/null and b/croco/utils/__pycache__/misc.cpython-312.pyc differ diff --git a/croco/utils/misc.py b/croco/utils/misc.py new file mode 100644 index 0000000000000000000000000000000000000000..ed7674a656b7de2b431e34e835148b0e80955a7b --- /dev/null +++ b/croco/utils/misc.py @@ -0,0 +1,628 @@ +# Copyright (C) 2022-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# utilitary functions for CroCo +# -------------------------------------------------------- +# References: +# MAE: https://github.com/facebookresearch/mae +# DeiT: https://github.com/facebookresearch/deit +# BEiT: https://github.com/microsoft/unilm/tree/master/beit +# -------------------------------------------------------- + +import builtins +import datetime +import os +import time +import math +import json +from collections import defaultdict, deque +from pathlib import Path +import numpy as np + +import torch +import torch.distributed as dist +from torch import inf +from accelerate import Accelerator +from accelerate.logging import get_logger + +printer = get_logger(__name__, log_level="DEBUG") + + +class SmoothedValue(object): + """Track a series of values and provide access to smoothed values.""" + + def __init__(self, window_size=20, fmt=None): + if fmt is None: + fmt = "{median:.4f} ({global_avg:.4f})" + self.deque = deque(maxlen=window_size) + self.total = 0.0 + self.count = 0 + self.fmt = fmt + + def update(self, value, n=1): + self.deque.append(value) + self.count += n + self.total += value * n + + def synchronize_between_processes(self, accelerator: Accelerator): + """Synchronize the count and total across all processes.""" + if accelerator.num_processes == 1: + return + t = torch.tensor( + [self.count, self.total], dtype=torch.float64, device=accelerator.device + ) + accelerator.wait_for_everyone() + accelerator.reduce(t, reduction="sum") + t = t.tolist() + self.count = int(t[0]) + self.total = t[1] + + @property + def median(self): + return torch.tensor(list(self.deque)).median().item() + + @property + def avg(self): + return torch.tensor(list(self.deque), dtype=torch.float32).mean().item() + + @property + def global_avg(self): + return self.total / self.count + + @property + def max(self): + return max(self.deque) + + @property + def value(self): + return self.deque[-1] + + def __str__(self): + return self.fmt.format( + median=self.median, + avg=self.avg, + global_avg=self.global_avg, + max=self.max, + value=self.value, + ) + + +class MetricLogger(object): + def __init__(self, delimiter="\t"): + self.meters = defaultdict(SmoothedValue) + self.delimiter = delimiter + + def update(self, **kwargs): + for k, v in kwargs.items(): + if v is None: + continue + if isinstance(v, torch.Tensor): + if v.ndim > 0: + continue + v = v.item() + if isinstance(v, list): + continue + assert isinstance(v, (float, int)) + self.meters[k].update(v) + + def __getattr__(self, attr): + if attr in self.meters: + return self.meters[attr] + if attr in self.__dict__: + return self.__dict__[attr] + raise AttributeError( + "'{}' object has no attribute '{}'".format(type(self).__name__, attr) + ) + + def __str__(self): + loss_str = [] + for name, meter in self.meters.items(): + loss_str.append("{}: {}".format(name, str(meter))) + return self.delimiter.join(loss_str) + + def synchronize_between_processes(self, accelerator): + for meter in self.meters.values(): + meter.synchronize_between_processes(accelerator) + + def add_meter(self, name, meter): + self.meters[name] = meter + + def log_every( + self, iterable, print_freq, accelerator: Accelerator, header=None, max_iter=None, start_step=0, + ): + i = 0 + if not header: + header = "" + start_time = time.time() + end = time.time() + iter_time = SmoothedValue(fmt="{avg:.4f}") + data_time = SmoothedValue(fmt="{avg:.4f}") + len_iterable = min(len(iterable), max_iter) if max_iter else len(iterable) + space_fmt = ":" + str(len(str(len_iterable))) + "d" + log_msg = [ + header, + "[{0" + space_fmt + "}/{1}]", + "eta: {eta}", + "{meters}", + "time: {time}", + "data: {data}", + ] + if torch.cuda.is_available(): + log_msg.append("max mem: {memory:.0f}") + log_msg = self.delimiter.join(log_msg) + MB = 1024.0 * 1024.0 + for it, obj in enumerate(iterable): + data_time.update(time.time() - end) + yield obj + iter_time.update(time.time() - end) + if i % print_freq == 0 or i == len_iterable - 1: + eta_seconds = iter_time.global_avg * (len_iterable - i) + eta_string = str(datetime.timedelta(seconds=int(eta_seconds))) + if torch.cuda.is_available(): + if accelerator.is_main_process: + printer.info( + log_msg.format( + i, + len_iterable, + eta=eta_string, + meters=str(self), + time=str(iter_time), + data=str(data_time), + memory=torch.cuda.max_memory_allocated() / MB, + ) + ) + else: + if accelerator.is_main_process: + printer.info( + log_msg.format( + i, + len_iterable, + eta=eta_string, + meters=str(self), + time=str(iter_time), + data=str(data_time), + ) + ) + i += 1 + end = time.time() + if max_iter and it >= max_iter: + break + # if i + start_step >= len_iterable: + # break + total_time = time.time() - start_time + total_time_str = str(datetime.timedelta(seconds=int(total_time))) + if accelerator.is_main_process: + printer.info( + "{} Total time: {} ({:.4f} s / it)".format( + header, total_time_str, total_time / len_iterable + ) + ) + + +def setup_for_distributed(is_master): + """ + This function disables printing when not in master process + """ + builtin_print = builtins.print + + def print(*args, **kwargs): + force = kwargs.pop("force", False) + force = force or (get_world_size() > 8) + if is_master or force: + now = datetime.datetime.now().time() + builtin_print("[{}] ".format(now), end="") # print with time stamp + builtin_print(*args, **kwargs) + + builtins.print = print + + +def is_dist_avail_and_initialized(): + if not dist.is_available(): + return False + if not dist.is_initialized(): + return False + return True + + +def get_world_size(): + if not is_dist_avail_and_initialized(): + return 1 + return dist.get_world_size() + + +def get_rank(): + if not is_dist_avail_and_initialized(): + return 0 + return dist.get_rank() + + +def is_main_process(accelerator: Accelerator): + return accelerator.is_main_process + + +def save_on_master(accelerator: Accelerator, *args, **kwargs): + if is_main_process(accelerator): + # torch.save(*args, **kwargs) + accelerator.save(*args, **kwargs) + # unwrapped_model = accelerator.unwrap_model(model) + # accelerator.save(unwrapped_model.state_dict(), checkpoint_path) + + +def init_distributed_mode(args): + nodist = args.nodist if hasattr(args, "nodist") else False + if "RANK" in os.environ and "WORLD_SIZE" in os.environ and not nodist: + args.rank = int(os.environ["RANK"]) + args.world_size = int(os.environ["WORLD_SIZE"]) + args.gpu = int(os.environ["LOCAL_RANK"]) + else: + print("Not using distributed mode") + setup_for_distributed(is_master=True) # hack + args.distributed = False + return + + args.distributed = True + + torch.cuda.set_device(args.gpu) + args.dist_backend = "nccl" + print( + "| distributed init (rank {}): {}, gpu {}".format( + args.rank, args.dist_url, args.gpu + ), + flush=True, + ) + torch.distributed.init_process_group( + backend=args.dist_backend, + init_method=args.dist_url, + world_size=args.world_size, + rank=args.rank, + ) + torch.distributed.barrier() + setup_for_distributed(args.rank == 0) + + +class NativeScalerWithGradNormCount: + state_dict_key = "amp_scaler" + + def __init__(self, enabled=True, accelerator: Accelerator = None): + self.accelerator = accelerator + + def __call__( + self, + loss, + optimizer, + clip_grad=None, + parameters=None, + create_graph=False, + update_grad=True, + ): + self.accelerator.backward( + loss, create_graph=create_graph + ) # .backward(create_graph=create_graph) + if update_grad: + if clip_grad is not None: + assert parameters is not None + # self._scaler.unscale_(optimizer) # unscale the gradients of optimizer's assigned params in-place + norm = self.accelerator.clip_grad_norm_(parameters, clip_grad) + else: + if self.accelerator.scaler is not None: + self.accelerator.unscale_gradients() + norm = get_grad_norm_(parameters) + optimizer.step() + else: + norm = None + return norm + + def state_dict(self): + if self.accelerator.scaler is not None: + return self.accelerator.scaler.state_dict() + else: + return {} + + def load_state_dict(self, state_dict): + if self.accelerator.scaler is not None: + self.accelerator.scaler.load_state_dict(state_dict) + + +# class NativeScalerWithGradNormCount: +# state_dict_key = "amp_scaler" + +# def __init__(self, enabled=True, accelerator:Accelerator=None): +# self._scaler = torch.cuda.amp.GradScaler(enabled=enabled) +# self.accelerator = accelerator + +# def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True): +# # self.accelerator.backward(loss, create_graph=create_graph) #.backward(create_graph=create_graph) +# self._scaler.scale(loss).backward(create_graph=create_graph) +# if update_grad: +# if clip_grad is not None: +# assert parameters is not None +# # #self._scaler.unscale_(optimizer) # unscale the gradients of optimizer's assigned params in-place +# # norm = self.accelerator.clip_grad_norm_(parameters, clip_grad) +# self._scaler.unscale_(optimizer) # unscale the gradients of optimizer's assigned params in-place +# norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad) +# else: +# # if self.accelerator.scaler is not None: +# # self.accelerator.unscale_gradients() +# # norm = get_grad_norm_(parameters) +# self._scaler.unscale_(optimizer) +# norm = get_grad_norm_(parameters) +# # optimizer.step() +# self._scaler.step(optimizer) +# self._scaler.update() +# else: +# norm = None +# return norm + +# # def state_dict(self): +# # if self.accelerator.scaler is not None: +# # return self.accelerator.scaler.state_dict() +# # else: +# # return {} + +# # def load_state_dict(self, state_dict): +# # if self.accelerator.scaler is not None: +# # self.accelerator.scaler.load_state_dict(state_dict) + +# def state_dict(self): +# return self._scaler.state_dict() + +# def load_state_dict(self, state_dict): +# self._scaler.load_state_dict(state_dict) + + +def get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor: + if isinstance(parameters, torch.Tensor): + parameters = [parameters] + parameters = [p for p in parameters if p.grad is not None] + norm_type = float(norm_type) + if len(parameters) == 0: + return torch.tensor(0.0) + device = parameters[0].grad.device + if norm_type == inf: + total_norm = max(p.grad.detach().abs().max().to(device) for p in parameters) + else: + total_norm = torch.norm( + torch.stack( + [torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters] + ), + norm_type, + ) + return total_norm + + +def save_model( + accelerator, + args, + epoch, + model_without_ddp, + optimizer, + loss_scaler, + step, + fname=None, + best_so_far=None, +): + if accelerator.is_main_process: + output_dir = Path(args.output_dir) + if fname is None: + fname = str(epoch) + checkpoint_path = output_dir / ("checkpoint-%s.pth" % fname) + to_save = { + "model": model_without_ddp.state_dict(), + "optimizer": optimizer.state_dict(), + "scaler": loss_scaler.state_dict(), + "args": args, + "epoch": epoch, + "step": step, + } + if best_so_far is not None: + to_save["best_so_far"] = best_so_far + print(f">> Saving model to {checkpoint_path} ...") + save_on_master(accelerator, to_save, checkpoint_path) + + to_save = { + "model": model_without_ddp.state_dict(), + } + checkpoint_path = output_dir / ("model.pth") + save_on_master(accelerator, to_save, checkpoint_path) + + +def load_model(args, model_without_ddp, optimizer, loss_scaler): + args.start_epoch = 0 + args.start_step = 0 + best_so_far = None + if args.resume is not None: + if args.resume.startswith("https"): + checkpoint = torch.hub.load_state_dict_from_url( + args.resume, map_location="cpu", check_hash=True + ) + else: + checkpoint = torch.load(args.resume, map_location="cuda", weights_only=False) + printer.info("Resume checkpoint %s" % args.resume) + state_dict = checkpoint["model"] + new_state_dict = {k.replace('module.', ''): v for k, v in state_dict.items()} + model_without_ddp.load_state_dict(new_state_dict, strict=True) + args.start_epoch = checkpoint["epoch"] + 1 + if "step" in checkpoint: + args.start_step = checkpoint["step"] + device = next(model_without_ddp.parameters()).device + printer.info(f"Moving optimizer state to device: {device}") + + if "optimizer" in checkpoint: + for state in checkpoint["optimizer"]["state"].values(): + for k, v in state.items(): + if isinstance(v, torch.Tensor): + state[k] = v.to(device) + + optimizer.load_state_dict(checkpoint["optimizer"]) + + if "scaler" in checkpoint: + loss_scaler.load_state_dict(checkpoint["scaler"]) + if "best_so_far" in checkpoint: + best_so_far = checkpoint["best_so_far"] + printer.info(" & best_so_far={:g}".format(best_so_far)) + else: + printer.info("") + printer.info("With optim & sched! start_epoch={:d}".format(args.start_epoch)) + return best_so_far + + +def all_reduce_mean(x, accelerator): + """Use accelerator to all-reduce and compute mean.""" + if accelerator.state.num_processes > 1: + x_reduce = torch.tensor(x).cuda() + accelerator.reduce(x_reduce, reduce_op="SUM") + x_reduce /= accelerator.state.num_processes + return x_reduce.item() + else: + return x + + +def _replace(text, src, tgt, rm=""): + """Advanced string replacement. + Given a text: + - replace all elements in src by the corresponding element in tgt + - remove all elements in rm + """ + if len(tgt) == 1: + tgt = tgt * len(src) + assert len(src) == len(tgt), f"'{src}' and '{tgt}' should have the same len" + for s, t in zip(src, tgt): + text = text.replace(s, t) + for c in rm: + text = text.replace(c, "") + return text + + +def filename(obj): + """transform a python obj or cmd into a proper filename. + - \1 gets replaced by slash '/' + - \2 gets replaced by comma ',' + """ + if not isinstance(obj, str): + obj = repr(obj) + obj = str(obj).replace("()", "") + obj = _replace(obj, "_,(*/\1\2", "-__x%/,", rm=" )'\"") + assert all(len(s) < 256 for s in obj.split(os.sep)), ( + "filename too long (>256 characters):\n" + obj + ) + return obj + + +def _get_num_layer_for_vit(var_name, enc_depth, dec_depth): + if var_name in ("cls_token", "mask_token", "pos_embed", "global_tokens"): + return 0 + elif var_name.startswith("patch_embed"): + return 0 + elif var_name.startswith("enc_blocks"): + layer_id = int(var_name.split(".")[1]) + return layer_id + 1 + elif var_name.startswith("decoder_embed") or var_name.startswith( + "enc_norm" + ): # part of the last black + return enc_depth + elif var_name.startswith("dec_blocks"): + layer_id = int(var_name.split(".")[1]) + return enc_depth + layer_id + 1 + elif var_name.startswith("dec_norm"): # part of the last block + return enc_depth + dec_depth + elif any(var_name.startswith(k) for k in ["head", "prediction_head"]): + return enc_depth + dec_depth + 1 + else: + raise NotImplementedError(var_name) + + +def get_parameter_groups( + model, weight_decay, layer_decay=1.0, skip_list=(), no_lr_scale_list=[] +): + parameter_group_names = {} + parameter_group_vars = {} + enc_depth, dec_depth = None, None + # prepare layer decay values + assert layer_decay == 1.0 or 0.0 < layer_decay < 1.0 + if layer_decay < 1.0: + enc_depth = model.enc_depth + dec_depth = model.dec_depth if hasattr(model, "dec_blocks") else 0 + num_layers = enc_depth + dec_depth + layer_decay_values = list( + layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2) + ) + + for name, param in model.named_parameters(): + if not param.requires_grad: + continue # frozen weights + + # Assign weight decay values + if len(param.shape) == 1 or name.endswith(".bias") or name in skip_list: + if "enc_blocks" in name: + group_name = "no_decay_enc_blocks" + else: + group_name = "no_decay" + this_weight_decay = 0.0 + else: + if "enc_blocks" in name: + group_name = "decay_enc_blocks" + else: + group_name = "decay" + this_weight_decay = weight_decay + + # Assign layer ID for LR scaling + if layer_decay < 1.0: + skip_scale = False + layer_id = _get_num_layer_for_vit(name, enc_depth, dec_depth) + group_name = "layer_%d_%s" % (layer_id, group_name) + if name in no_lr_scale_list: + skip_scale = True + group_name = f"{group_name}_no_lr_scale" + else: + layer_id = 0 + skip_scale = True + + if group_name not in parameter_group_names: + if not skip_scale: + scale = layer_decay_values[layer_id] + else: + scale = 1.0 + + if "enc_blocks" in group_name: + scale *= 1.0 + parameter_group_names[group_name] = { + "weight_decay": this_weight_decay, + "params": [], + "lr_scale": scale, + } + parameter_group_vars[group_name] = { + "weight_decay": this_weight_decay, + "params": [], + "lr_scale": scale, + } + + parameter_group_vars[group_name]["params"].append(param) + parameter_group_names[group_name]["params"].append(name) + printer.info("Param groups = %s" % json.dumps(parameter_group_names, indent=2)) + return list(parameter_group_vars.values()) + + +def adjust_learning_rate(optimizer, epoch, args): + """Decay the learning rate with half-cycle cosine after warmup""" + + if epoch < args.warmup_epochs: + lr = args.lr * epoch / args.warmup_epochs + else: + # lr = args.lr + lr = args.min_lr + (args.lr - args.min_lr) * 0.5 * ( + 1.0 + + math.cos( + math.pi + * (epoch - args.warmup_epochs) + / (args.epochs - args.warmup_epochs) + ) + ) + + for param_group in optimizer.param_groups: + if "lr_scale" in param_group: + param_group["lr"] = lr * param_group["lr_scale"] + else: + param_group["lr"] = lr + + return lr diff --git a/dust3r/__init__.py b/dust3r/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/dust3r/blocks.py b/dust3r/blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..5ee03a37aa5173a23969a1096208c8442a66a043 --- /dev/null +++ b/dust3r/blocks.py @@ -0,0 +1,531 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import torch +import torch.nn as nn + +from itertools import repeat +import collections.abc +from torch.nn.functional import scaled_dot_product_attention +from functools import partial + + +def _ntuple(n): + def parse(x): + if isinstance(x, collections.abc.Iterable) and not isinstance(x, str): + return x + return tuple(repeat(x, n)) + + return parse + + +to_2tuple = _ntuple(2) + + +def drop_path( + x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True +): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).""" + if drop_prob == 0.0 or not training: + return x + keep_prob = 1 - drop_prob + shape = (x.shape[0],) + (1,) * ( + x.ndim - 1 + ) # work with diff dim tensors, not just 2D ConvNets + random_tensor = x.new_empty(shape).bernoulli_(keep_prob) + if keep_prob > 0.0 and scale_by_keep: + random_tensor.div_(keep_prob) + return x * random_tensor + + +class DropPath(nn.Module): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).""" + + def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True): + super(DropPath, self).__init__() + self.drop_prob = drop_prob + self.scale_by_keep = scale_by_keep + + def forward(self, x): + return drop_path(x, self.drop_prob, self.training, self.scale_by_keep) + + def extra_repr(self): + return f"drop_prob={round(self.drop_prob,3):0.3f}" + + +class Mlp(nn.Module): + """MLP as used in Vision Transformer, MLP-Mixer and related networks""" + + def __init__( + self, + in_features, + hidden_features=None, + out_features=None, + act_layer=nn.GELU, + bias=True, + drop=0.0, + ): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + bias = to_2tuple(bias) + drop_probs = to_2tuple(drop) + + self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0]) + self.act = act_layer() + self.drop1 = nn.Dropout(drop_probs[0]) + self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1]) + self.drop2 = nn.Dropout(drop_probs[1]) + + def forward(self, x): + return self.drop2(self.fc2(self.drop1(self.act(self.fc1(x))))) + + +class Attention(nn.Module): + + def __init__( + self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0 + ): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = head_dim**-0.5 + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + self.rope = rope.float() if rope is not None else None + + def forward(self, x, xpos): + B, N, C = x.shape + + qkv = ( + self.qkv(x) + .reshape(B, N, 3, self.num_heads, C // self.num_heads) + .transpose(1, 3) + ) + q, k, v = [qkv[:, :, i] for i in range(3)] + + q_type = q.dtype + k_type = k.dtype + if self.rope is not None: + q = q.float() + k = k.float() + with torch.autocast(device_type="cuda", enabled=False): + q = self.rope(q, xpos) + k = self.rope(k, xpos) + q = q.to(q_type) + k = k.to(k_type) + + x = ( + scaled_dot_product_attention( + query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale + ) + .transpose(1, 2) + .reshape(B, N, C) + ) + + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class Block(nn.Module): + + def __init__( + self, + dim, + num_heads, + mlp_ratio=4.0, + qkv_bias=False, + drop=0.0, + attn_drop=0.0, + drop_path=0.0, + act_layer=nn.GELU, + norm_layer=nn.LayerNorm, + rope=None, + ): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp( + in_features=dim, + hidden_features=mlp_hidden_dim, + act_layer=act_layer, + drop=drop, + ) + + def forward(self, x, xpos): + x = x + self.drop_path(self.attn(self.norm1(x), xpos)) + x = x + self.drop_path(self.mlp(self.norm2(x))) + return x + + +class CrossAttention(nn.Module): + + def __init__( + self, dim, rope=None, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0 + ): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = head_dim**-0.5 + + self.projq = nn.Linear(dim, dim, bias=qkv_bias) + self.projk = nn.Linear(dim, dim, bias=qkv_bias) + self.projv = nn.Linear(dim, dim, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + self.rope = rope.float() if rope is not None else None + + def forward(self, query, key, value, qpos, kpos): + B, Nq, C = query.shape + Nk = key.shape[1] + Nv = value.shape[1] + + q = ( + self.projq(query) + .reshape(B, Nq, self.num_heads, C // self.num_heads) + .permute(0, 2, 1, 3) + ) + k = ( + self.projk(key) + .reshape(B, Nk, self.num_heads, C // self.num_heads) + .permute(0, 2, 1, 3) + ) + v = ( + self.projv(value) + .reshape(B, Nv, self.num_heads, C // self.num_heads) + .permute(0, 2, 1, 3) + ) + + q_type = q.dtype + k_type = k.dtype + if self.rope is not None: + if qpos is not None: + q = q.float() + with torch.autocast(device_type="cuda", enabled=False): + q = self.rope(q, qpos) + q = q.to(q_type) + + if kpos is not None: + k = k.float() + with torch.autocast(device_type="cuda", enabled=False): + k = self.rope(k, kpos) + k = k.to(k_type) + + x = ( + scaled_dot_product_attention( + query=q, key=k, value=v, dropout_p=self.attn_drop.p, scale=self.scale + ) + .transpose(1, 2) + .reshape(B, Nq, C) + ) + + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class DecoderBlock(nn.Module): + + def __init__( + self, + dim, + num_heads, + mlp_ratio=4.0, + qkv_bias=False, + drop=0.0, + attn_drop=0.0, + drop_path=0.0, + act_layer=nn.GELU, + norm_layer=nn.LayerNorm, + norm_mem=True, + rope=None, + ): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + self.cross_attn = CrossAttention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.norm2 = norm_layer(dim) + self.norm3 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp( + in_features=dim, + hidden_features=mlp_hidden_dim, + act_layer=act_layer, + drop=drop, + ) + self.norm_y = norm_layer(dim) if norm_mem else nn.Identity() + + def forward(self, x, y, xpos, ypos): + x = x + self.drop_path(self.attn(self.norm1(x), xpos)) + y_ = self.norm_y(y) + x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, y_, xpos, ypos)) + x = x + self.drop_path(self.mlp(self.norm3(x))) + return x, y + + +class CustomDecoderBlock(nn.Module): + + def __init__( + self, + dim, + num_heads, + mlp_ratio=4.0, + qkv_bias=False, + drop=0.0, + attn_drop=0.0, + drop_path=0.0, + act_layer=nn.GELU, + norm_layer=nn.LayerNorm, + norm_mem=True, + rope=None, + ): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + self.cross_attn = CrossAttention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.norm2 = norm_layer(dim) + self.norm3 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp( + in_features=dim, + hidden_features=mlp_hidden_dim, + act_layer=act_layer, + drop=drop, + ) + self.norm_y = norm_layer(dim) if norm_mem else nn.Identity() + self.norm_z = norm_layer(dim) if norm_mem else nn.Identity() + + def forward(self, x, y, z, xpos, ypos): + x = x + self.drop_path(self.attn(self.norm1(x), xpos)) + y_ = self.norm_y(y) + z_ = self.norm_z(z) + x = x + self.drop_path(self.cross_attn(self.norm2(x), y_, z_, xpos, ypos)) + x = x + self.drop_path(self.mlp(self.norm3(x))) + return x, y + + +class ModLN(nn.Module): + """ + Modulation with adaLN. + + References: + DiT: https://github.com/facebookresearch/DiT/blob/main/models.py#L101 + """ + + def __init__(self, inner_dim: int, mod_dim: int, eps: float): + super().__init__() + self.norm = nn.LayerNorm(inner_dim, eps=eps) + self.mlp = nn.Sequential( + nn.SiLU(), + nn.Linear(mod_dim, inner_dim * 2), + ) + + @staticmethod + def modulate(x, shift, scale): + + return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) + + def forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor: + shift, scale = self.mlp(mod).chunk(2, dim=-1) # [N, D] + return self.modulate(self.norm(x), shift, scale) # [N, L, D] + + +class ConditionModulationBlock(nn.Module): + + def __init__( + self, + dim, + num_heads, + mlp_ratio=4.0, + qkv_bias=False, + drop=0.0, + attn_drop=0.0, + drop_path=0.0, + act_layer=nn.GELU, + norm_layer=partial(ModLN, eps=1e-6), + rope=None, + ): + super().__init__() + self.norm1 = norm_layer(dim, dim) + self.attn = Attention( + dim, + rope=rope, + num_heads=num_heads, + qkv_bias=qkv_bias, + attn_drop=attn_drop, + proj_drop=drop, + ) + self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + self.norm2 = norm_layer(dim, dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp( + in_features=dim, + hidden_features=mlp_hidden_dim, + act_layer=act_layer, + drop=drop, + ) + + def forward(self, x, mod, xpos): + x = x + self.drop_path(self.attn(self.norm1(x, mod), xpos)) + x = x + self.drop_path(self.mlp(self.norm2(x, mod))) + return x + + +class PositionGetter(object): + """return positions of patches""" + + def __init__(self): + self.cache_positions = {} + + def __call__(self, b, h, w, device): + if not (h, w) in self.cache_positions: + x = torch.arange(w, device=device) + y = torch.arange(h, device=device) + self.cache_positions[h, w] = torch.cartesian_prod(y, x) # (h, w, 2) + pos = self.cache_positions[h, w].view(1, h * w, 2).expand(b, -1, 2).clone() + return pos + + +class PatchEmbed(nn.Module): + """just adding _init_weights + position getter compared to timm.models.layers.patch_embed.PatchEmbed""" + + def __init__( + self, + img_size=224, + patch_size=16, + in_chans=3, + embed_dim=768, + norm_layer=None, + flatten=True, + ): + super().__init__() + img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + self.img_size = img_size + self.patch_size = patch_size + self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1]) + self.num_patches = self.grid_size[0] * self.grid_size[1] + self.flatten = flatten + + self.proj = nn.Conv2d( + in_chans, embed_dim, kernel_size=patch_size, stride=patch_size + ) + self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity() + + self.position_getter = PositionGetter() + + def forward(self, x): + B, C, H, W = x.shape + torch._assert( + H == self.img_size[0], + f"Input image height ({H}) doesn't match model ({self.img_size[0]}).", + ) + torch._assert( + W == self.img_size[1], + f"Input image width ({W}) doesn't match model ({self.img_size[1]}).", + ) + x = self.proj(x) + pos = self.position_getter(B, x.size(2), x.size(3), x.device) + if self.flatten: + x = x.flatten(2).transpose(1, 2) # BCHW -> BNC + x = self.norm(x) + return x, pos + + def _init_weights(self): + w = self.proj.weight.data + torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1])) + + +if __name__ == "__main__": + import os + import sys + + sys.path.append(os.path.dirname(os.path.dirname(__file__))) + import dust3r.utils.path_to_croco + from models.pos_embed import get_2d_sincos_pos_embed, RoPE2D + from functools import partial + from torch.utils.checkpoint import checkpoint + + torch.manual_seed(0) + + enc_blocks_ray_map = ( + nn.ModuleList( + [ + Block( + 768, + 16, + 4, + qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), + rope=RoPE2D(100), + ) + for _ in range(2) + ] + ) + .cuda() + .train() + ) + + x = torch.randn(2, 196, 768, requires_grad=True).cuda() + xpos = torch.arange(0, 196).unsqueeze(0).unsqueeze(-1).repeat(2, 1, 2).cuda().long() + enc_blocks_ray_map.zero_grad() + for blk in enc_blocks_ray_map: + + x = checkpoint(blk, x, xpos) + enc_blocks_ray_map.zero_grad() + x.sum().backward() + + grad_not_checkpointed = {} + for name, param in enc_blocks_ray_map.named_parameters(): + grad_not_checkpointed[name] = param.grad.data.clone() + print(name, grad_not_checkpointed[name]) + break diff --git a/dust3r/datasets/__init__.py b/dust3r/datasets/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..99239c9599f760d748dc2e41b3da519456484e2f --- /dev/null +++ b/dust3r/datasets/__init__.py @@ -0,0 +1,86 @@ +from .utils.transforms import * +from .base.batched_sampler import BatchedRandomSampler # noqa +from .arkitscenes import ARKitScenes_Multi # noqa +from .arkitscenes_highres import ARKitScenesHighRes_Multi +from .bedlam import BEDLAM_Multi +from .blendedmvs import BlendedMVS_Multi # noqa +from .co3d import Co3d_Multi # noqa +from .cop3d import Cop3D_Multi +from .dl3dv import DL3DV_Multi +from .dynamic_replica import DynamicReplica +from .eden import EDEN_Multi +from .hypersim import HyperSim_Multi +from .hoi4d import HOI4D_Multi +from .irs import IRS +from .mapfree import MapFree_Multi +from .megadepth import MegaDepth_Multi # noqa +from .mp3d import MP3D_Multi +from .mvimgnet import MVImgNet_Multi +from .mvs_synth import MVS_Synth_Multi +from .omniobject3d import OmniObject3D_Multi +from .pointodyssey import PointOdyssey_Multi +from .realestate10k import RE10K_Multi +from .scannet import ScanNet_Multi +from .scannetpp import ScanNetpp_Multi # noqa +from .smartportraits import SmartPortraits_Multi +from .spring import Spring +from .synscapes import SynScapes +from .tartanair import TartanAir_Multi +from .threedkb import ThreeDKenBurns +from .uasol import UASOL_Multi +from .urbansyn import UrbanSyn +from .unreal4k import UnReal4K_Multi +from .vkitti2 import VirtualKITTI2_Multi # noqa +from .waymo import Waymo_Multi # noqa +from .wildrgbd import WildRGBD_Multi # noqa + + +from accelerate import Accelerator + + +def get_data_loader( + dataset, + batch_size, + num_workers=8, + shuffle=True, + drop_last=True, + pin_mem=True, + accelerator: Accelerator = None, + fixed_length=False, +): + import torch + + # pytorch dataset + if isinstance(dataset, str): + dataset = eval(dataset) + + try: + sampler = dataset.make_sampler( + batch_size, + shuffle=shuffle, + drop_last=drop_last, + world_size=accelerator.num_processes, + fixed_length=fixed_length, + ) + shuffle = False + + data_loader = torch.utils.data.DataLoader( + dataset, + batch_sampler=sampler, + num_workers=num_workers, + pin_memory=pin_mem, + ) + + except (AttributeError, NotImplementedError): + sampler = None + + data_loader = torch.utils.data.DataLoader( + dataset, + batch_size=batch_size, + shuffle=shuffle, + num_workers=num_workers, + pin_memory=pin_mem, + drop_last=drop_last, + ) + + return data_loader diff --git a/dust3r/datasets/arkitscenes.py b/dust3r/datasets/arkitscenes.py new file mode 100644 index 0000000000000000000000000000000000000000..214ee4f9b5d0238c7a79ce48972b51cc4d1c2ab3 --- /dev/null +++ b/dust3r/datasets/arkitscenes.py @@ -0,0 +1,242 @@ +import os.path as osp +import os +import sys +import itertools + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +def stratified_sampling(indices, num_samples, rng=None): + if num_samples > len(indices): + raise ValueError("num_samples cannot exceed the number of available indices.") + elif num_samples == len(indices): + return indices + + sorted_indices = sorted(indices) + stride = len(sorted_indices) / num_samples + sampled_indices = [] + if rng is None: + rng = np.random.default_rng() + + for i in range(num_samples): + start = int(i * stride) + end = int((i + 1) * stride) + # Ensure end does not exceed the list + end = min(end, len(sorted_indices)) + if start < end: + # Randomly select within the current stratum + rand_idx = rng.integers(start, end) + sampled_indices.append(sorted_indices[rand_idx]) + else: + # In case of any rounding issues, select the last index + sampled_indices.append(sorted_indices[-1]) + + return rng.permutation(sampled_indices) + + +class ARKitScenes_Multi(BaseMultiViewDataset): + def __init__(self, *args, split, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 8 + super().__init__(*args, **kwargs) + if split == "train": + self.split = "Training" + elif split == "test": + self.split = "Test" + else: + raise ValueError("") + + self.loaded_data = self._load_data(self.split) + + def _load_data(self, split): + with np.load(osp.join(self.ROOT, split, "all_metadata.npz")) as data: + self.scenes: np.ndarray = data["scenes"] + high_res_list = np.array( + [ + d + for d in os.listdir( + os.path.join( + self.ROOT.rstrip("/") + "_highres", + split if split == "Training" else "Validation", + ) + ) + if os.path.join(self.ROOT + "_highres", split, d) + ] + ) + self.scenes = np.setdiff1d(self.scenes, high_res_list) + offset = 0 + counts = [] + scenes = [] + sceneids = [] + images = [] + intrinsics = [] + trajectories = [] + groups = [] + id_ranges = [] + j = 0 + for scene_idx, scene in enumerate(self.scenes): + scene_dir = osp.join(self.ROOT, self.split, scene) + with np.load( + osp.join(scene_dir, "new_scene_metadata.npz"), allow_pickle=True + ) as data: + imgs = data["images"] + intrins = data["intrinsics"] + traj = data["trajectories"] + min_seq_len = ( + self.num_views + if not self.allow_repeat + else max(self.num_views // 3, 3) + ) + if len(imgs) < min_seq_len: + print(f"Skipping {scene}") + continue + + collections = {} + assert "image_collection" in data, "Image collection not found" + collections["image"] = data["image_collection"] + + num_imgs = imgs.shape[0] + img_groups = [] + min_group_len = ( + self.num_views + if not self.allow_repeat + else max(self.num_views // 3, 3) + ) + for ref_id, group in collections["image"].item().items(): + if len(group) + 1 < min_group_len: + continue + + # groups are (idx, score)s + group.insert(0, (ref_id, 1.0)) + group = [int(x[0] + offset) for x in group] + img_groups.append(sorted(group)) + + if len(img_groups) == 0: + print(f"Skipping {scene}") + continue + + scenes.append(scene) + sceneids.extend([j] * num_imgs) + id_ranges.extend([(offset, offset + num_imgs) for _ in range(num_imgs)]) + images.extend(imgs) + K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0) + + K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins] + K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins] + K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins] + K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins] + intrinsics.extend(list(K)) + trajectories.extend(list(traj)) + + # offset groups + groups.extend(img_groups) + counts.append(offset) + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.id_ranges = id_ranges + self.images = images + self.intrinsics = intrinsics + self.trajectories = trajectories + self.groups = groups + + def __len__(self): + return len(self.groups) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + + if rng.choice([True, False]): + image_idxs = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1]) + cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3) + start_image_idxs = image_idxs[: len(image_idxs) - cut_off + 1] + start_id = rng.choice(start_image_idxs) + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + image_idxs.tolist(), + rng, + max_interval=self.max_interval, + video_prob=0.8, + fix_interval_prob=0.5, + block_shuffle=16, + ) + image_idxs = np.array(image_idxs)[pos] + else: + ordered_video = False + image_idxs = self.groups[idx] + image_idxs = rng.permutation(image_idxs) + if len(image_idxs) > num_views: + image_idxs = image_idxs[:num_views] + else: + if rng.random() < 0.8: + image_idxs = rng.choice(image_idxs, size=num_views, replace=True) + else: + repeat_num = num_views // len(image_idxs) + 1 + image_idxs = np.tile(image_idxs, repeat_num)[:num_views] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id]) + + intrinsics = self.intrinsics[view_idx] + camera_pose = self.trajectories[view_idx] + basename = self.images[view_idx] + assert ( + basename[:8] == self.scenes[scene_id] + ), f"{basename}, {self.scenes[scene_id]}" + # print(scene_dir, basename) + # Load RGB image + rgb_image = imread_cv2( + osp.join(scene_dir, "vga_wide", basename.replace(".png", ".jpg")) + ) + # Load depthmap + depthmap = imread_cv2( + osp.join(scene_dir, "lowres_depth", basename), cv2.IMREAD_UNCHANGED + ) + depthmap = depthmap.astype(np.float32) / 1000.0 + depthmap[~np.isfinite(depthmap)] = 0 # invalid + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="arkitscenes", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.98, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/arkitscenes_highres.py b/dust3r/datasets/arkitscenes_highres.py new file mode 100644 index 0000000000000000000000000000000000000000..92826e1c46a067ed93ffc30d0470685085377bf6 --- /dev/null +++ b/dust3r/datasets/arkitscenes_highres.py @@ -0,0 +1,175 @@ +import os.path as osp +import os +import sys +import itertools + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np +import h5py +import math +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class ARKitScenesHighRes_Multi(BaseMultiViewDataset): + def __init__(self, *args, split, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.max_interval = 8 + self.is_metric = True + super().__init__(*args, **kwargs) + if split == "train": + self.split = "Training" + elif split == "test": + self.split = "Validation" + else: + raise ValueError("") + + self.loaded_data = self._load_data(self.split) + + def _load_data(self, split): + all_scenes = sorted( + [ + d + for d in os.listdir(osp.join(self.ROOT, split)) + if osp.isdir(osp.join(self.ROOT, split, d)) + ] + ) + offset = 0 + scenes = [] + sceneids = [] + images = [] + start_img_ids = [] + scene_img_list = [] + timestamps = [] + intrinsics = [] + trajectories = [] + scene_id = 0 + for scene in all_scenes: + scene_dir = osp.join(self.ROOT, self.split, scene) + with np.load(osp.join(scene_dir, "scene_metadata.npz")) as data: + imgs_with_indices = sorted( + enumerate(data["images"]), key=lambda x: x[1] + ) + imgs = [x[1] for x in imgs_with_indices] + cut_off = ( + self.num_views + if not self.allow_repeat + else max(self.num_views // 3, 3) + ) + if len(imgs) < cut_off: + print(f"Skipping {scene}") + continue + indices = [x[0] for x in imgs_with_indices] + tsps = np.array( + [float(img_name.split("_")[1][:-4]) for img_name in imgs] + ) + assert [img[:8] == scene for img in imgs], f"{scene}, {imgs}" + num_imgs = data["images"].shape[0] + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + scenes.append(scene) + scene_img_list.append(img_ids) + sceneids.extend([scene_id] * num_imgs) + images.extend(imgs) + start_img_ids.extend(start_img_ids_) + timestamps.extend(tsps) + + K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0) + intrins = data["intrinsics"][indices] + K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins] + K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins] + K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins] + K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins] + intrinsics.extend(list(K)) + trajectories.extend(list(data["trajectories"][indices])) + + # offset groups + offset += num_imgs + scene_id += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.scene_img_list = scene_img_list + self.intrinsics = intrinsics + self.trajectories = trajectories + self.start_img_ids = start_img_ids + assert len(self.images) == len(self.intrinsics) == len(self.trajectories) + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + block_shuffle=16, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id]) + + intrinsics = self.intrinsics[view_idx] + camera_pose = self.trajectories[view_idx] + basename = self.images[view_idx] + assert ( + basename[:8] == self.scenes[scene_id] + ), f"{basename}, {self.scenes[scene_id]}" + # print(scene_dir, basename) + # Load RGB image + rgb_image = imread_cv2( + osp.join(scene_dir, "vga_wide", basename.replace(".png", ".jpg")) + ) + # Load depthmap + depthmap = imread_cv2( + osp.join(scene_dir, "highres_depth", basename), cv2.IMREAD_UNCHANGED + ) + depthmap = depthmap.astype(np.float32) / 1000.0 + depthmap[~np.isfinite(depthmap)] = 0 # invalid + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.7, 0.25, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="arkitscenes_highres", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.99, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/base/__init__.py b/dust3r/datasets/base/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/dust3r/datasets/base/base_multiview_dataset.py b/dust3r/datasets/base/base_multiview_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..dc04725a6d22e4c7685cee3e1fb75856803df7e7 --- /dev/null +++ b/dust3r/datasets/base/base_multiview_dataset.py @@ -0,0 +1,546 @@ +import PIL +import numpy as np +import torch +import random +import itertools +from dust3r.datasets.base.easy_dataset import EasyDataset +from dust3r.datasets.utils.transforms import ImgNorm, SeqColorJitter +from dust3r.utils.geometry import depthmap_to_absolute_camera_coordinates +import dust3r.datasets.utils.cropping as cropping +from dust3r.datasets.utils.corr import extract_correspondences_from_pts3d + + +def get_ray_map(c2w1, c2w2, intrinsics, h, w): + c2w = np.linalg.inv(c2w1) @ c2w2 + i, j = np.meshgrid(np.arange(w), np.arange(h), indexing="xy") + grid = np.stack([i, j, np.ones_like(i)], axis=-1) + ro = c2w[:3, 3] + rd = np.linalg.inv(intrinsics) @ grid.reshape(-1, 3).T + rd = (c2w @ np.vstack([rd, np.ones_like(rd[0])])).T[:, :3].reshape(h, w, 3) + rd = rd / np.linalg.norm(rd, axis=-1, keepdims=True) + ro = np.broadcast_to(ro, (h, w, 3)) + ray_map = np.concatenate([ro, rd], axis=-1) + return ray_map + + +class BaseMultiViewDataset(EasyDataset): + """Define all basic options. + + Usage: + class MyDataset (BaseMultiViewDataset): + def _get_views(self, idx, rng): + # overload here + views = [] + views.append(dict(img=, ...)) + return views + """ + + def __init__( + self, + *, # only keyword arguments + num_views=None, + split=None, + resolution=None, # square_size or (width, height) or list of [(width,height), ...] + transform=ImgNorm, + aug_crop=False, + n_corres=0, + nneg=0, + seed=None, + allow_repeat=False, + seq_aug_crop=False, + ): + assert num_views is not None, "undefined num_views" + self.num_views = num_views + self.split = split + self._set_resolutions(resolution) + + self.n_corres = n_corres + self.nneg = nneg + assert ( + self.n_corres == "all" + or isinstance(self.n_corres, int) + or ( + isinstance(self.n_corres, list) and len(self.n_corres) == self.num_views + ) + ), f"Error, n_corres should either be 'all', a single integer or a list of length {self.num_views}" + assert ( + self.nneg == 0 or self.n_corres != "all" + ), "nneg should be 0 if n_corres is all" + + self.is_seq_color_jitter = False + if isinstance(transform, str): + transform = eval(transform) + if transform == SeqColorJitter: + transform = SeqColorJitter() + self.is_seq_color_jitter = True + self.transform = transform + + self.aug_crop = aug_crop + self.seed = seed + self.allow_repeat = allow_repeat + self.seq_aug_crop = seq_aug_crop + + def __len__(self): + return len(self.scenes) + + @staticmethod + def efficient_random_intervals( + start, + num_elements, + interval_range, + fixed_interval_prob=0.8, + weights=None, + seed=42, + ): + if random.random() < fixed_interval_prob: + intervals = random.choices(interval_range, weights=weights) * ( + num_elements - 1 + ) + else: + intervals = [ + random.choices(interval_range, weights=weights)[0] + for _ in range(num_elements - 1) + ] + return list(itertools.accumulate([start] + intervals)) + + def sample_based_on_timestamps(self, i, timestamps, num_views, interval=1): + time_diffs = np.abs(timestamps - timestamps[i]) + ids_candidate = np.where(time_diffs < interval)[0] + ids_candidate = np.sort(ids_candidate) + if (self.allow_repeat and len(ids_candidate) < num_views // 3) or ( + len(ids_candidate) < num_views + ): + return [] + ids_sel_list = [] + ids_candidate_left = ids_candidate.copy() + while len(ids_candidate_left) >= num_views: + ids_sel = np.random.choice(ids_candidate_left, num_views, replace=False) + ids_sel_list.append(sorted(ids_sel)) + ids_candidate_left = np.setdiff1d(ids_candidate_left, ids_sel) + + if len(ids_candidate_left) > 0 and len(ids_candidate) >= num_views: + ids_sel = np.concatenate( + [ + ids_candidate_left, + np.random.choice( + np.setdiff1d(ids_candidate, ids_candidate_left), + num_views - len(ids_candidate_left), + replace=False, + ), + ] + ) + ids_sel_list.append(sorted(ids_sel)) + + if self.allow_repeat: + ids_sel_list.append( + sorted(np.random.choice(ids_candidate, num_views, replace=True)) + ) + + # add sequences with fixed intervals (all possible intervals) + pos_i = np.where(ids_candidate == i)[0][0] + curr_interval = 1 + stop = len(ids_candidate) < num_views + while not stop: + pos_sel = [pos_i] + count = 0 + while len(pos_sel) < num_views: + if count % 2 == 0: + curr_pos_i = pos_sel[-1] + curr_interval + if curr_pos_i >= len(ids_candidate): + stop = True + break + pos_sel.append(curr_pos_i) + else: + curr_pos_i = pos_sel[0] - curr_interval + if curr_pos_i < 0: + stop = True + break + pos_sel.insert(0, curr_pos_i) + count += 1 + if not stop and len(pos_sel) == num_views: + ids_sel = sorted([ids_candidate[pos] for pos in pos_sel]) + if ids_sel not in ids_sel_list: + ids_sel_list.append(ids_sel) + curr_interval += 1 + return ids_sel_list + + @staticmethod + def blockwise_shuffle(x, rng, block_shuffle): + if block_shuffle is None: + return rng.permutation(x).tolist() + else: + assert block_shuffle > 0 + blocks = [x[i : i + block_shuffle] for i in range(0, len(x), block_shuffle)] + shuffled_blocks = [rng.permutation(block).tolist() for block in blocks] + shuffled_list = [item for block in shuffled_blocks for item in block] + return shuffled_list + + def get_seq_from_start_id( + self, + num_views, + id_ref, + ids_all, + rng, + min_interval=1, + max_interval=25, + video_prob=0.5, + fix_interval_prob=0.5, + block_shuffle=None, + ): + """ + args: + num_views: number of views to return + id_ref: the reference id (first id) + ids_all: all the ids + rng: random number generator + max_interval: maximum interval between two views + returns: + pos: list of positions of the views in ids_all, i.e., index for ids_all + is_video: True if the views are consecutive + """ + assert min_interval > 0, f"min_interval should be > 0, got {min_interval}" + assert ( + min_interval <= max_interval + ), f"min_interval should be <= max_interval, got {min_interval} and {max_interval}" + assert id_ref in ids_all + pos_ref = ids_all.index(id_ref) + all_possible_pos = np.arange(pos_ref, len(ids_all)) + + remaining_sum = len(ids_all) - 1 - pos_ref + + if remaining_sum >= num_views - 1: + if remaining_sum == num_views - 1: + assert ids_all[-num_views] == id_ref + return [pos_ref + i for i in range(num_views)], True + max_interval = min(max_interval, 2 * remaining_sum // (num_views - 1)) + intervals = [ + rng.choice(range(min_interval, max_interval + 1)) + for _ in range(num_views - 1) + ] + + # if video or collection + if rng.random() < video_prob: + # if fixed interval or random + if rng.random() < fix_interval_prob: + # regular interval + fixed_interval = rng.choice( + range( + 1, + min(remaining_sum // (num_views - 1) + 1, max_interval + 1), + ) + ) + intervals = [fixed_interval for _ in range(num_views - 1)] + is_video = True + else: + is_video = False + + pos = list(itertools.accumulate([pos_ref] + intervals)) + pos = [p for p in pos if p < len(ids_all)] + pos_candidates = [p for p in all_possible_pos if p not in pos] + pos = ( + pos + + rng.choice( + pos_candidates, num_views - len(pos), replace=False + ).tolist() + ) + + pos = ( + sorted(pos) + if is_video + else self.blockwise_shuffle(pos, rng, block_shuffle) + ) + else: + # assert self.allow_repeat + uniq_num = remaining_sum + new_pos_ref = rng.choice(np.arange(pos_ref + 1)) + new_remaining_sum = len(ids_all) - 1 - new_pos_ref + new_max_interval = min(max_interval, new_remaining_sum // (uniq_num - 1)) + new_intervals = [ + rng.choice(range(1, new_max_interval + 1)) for _ in range(uniq_num - 1) + ] + + revisit_random = rng.random() + video_random = rng.random() + + if rng.random() < fix_interval_prob and video_random < video_prob: + # regular interval + fixed_interval = rng.choice(range(1, new_max_interval + 1)) + new_intervals = [fixed_interval for _ in range(uniq_num - 1)] + pos = list(itertools.accumulate([new_pos_ref] + new_intervals)) + + is_video = False + if revisit_random < 0.5 or video_prob == 1.0: # revisit, video / collection + is_video = video_random < video_prob + pos = ( + self.blockwise_shuffle(pos, rng, block_shuffle) + if not is_video + else pos + ) + num_full_repeat = num_views // uniq_num + pos = ( + pos * num_full_repeat + + pos[: num_views - len(pos) * num_full_repeat] + ) + elif revisit_random < 0.9: # random + pos = rng.choice(pos, num_views, replace=True) + else: # ordered + pos = sorted(rng.choice(pos, num_views, replace=True)) + assert len(pos) == num_views + return pos, is_video + + def get_img_and_ray_masks(self, is_metric, v, rng, p=[0.8, 0.15, 0.05]): + # generate img mask and raymap mask + if v == 0 or (not is_metric): + img_mask = True + raymap_mask = False + else: + rand_val = rng.random() + if rand_val < p[0]: + img_mask = True + raymap_mask = False + elif rand_val < p[0] + p[1]: + img_mask = False + raymap_mask = True + else: + img_mask = True + raymap_mask = True + return img_mask, raymap_mask + + def get_stats(self): + return f"{len(self)} groups of views" + + def __repr__(self): + resolutions_str = "[" + ";".join(f"{w}x{h}" for w, h in self._resolutions) + "]" + return ( + f"""{type(self).__name__}({self.get_stats()}, + {self.num_views=}, + {self.split=}, + {self.seed=}, + resolutions={resolutions_str}, + {self.transform=})""".replace( + "self.", "" + ) + .replace("\n", "") + .replace(" ", "") + ) + + def _get_views(self, idx, resolution, rng, num_views): + raise NotImplementedError() + + def __getitem__(self, idx): + # print("Receiving:" , idx) + if isinstance(idx, (tuple, list, np.ndarray)): + # the idx is specifying the aspect-ratio + idx, ar_idx, nview = idx + else: + assert len(self._resolutions) == 1 + ar_idx = 0 + nview = self.num_views + + assert nview >= 1 and nview <= self.num_views + # set-up the rng + if self.seed: # reseed for each __getitem__ + self._rng = np.random.default_rng(seed=self.seed + idx) + elif not hasattr(self, "_rng"): + seed = torch.randint(0, 2**32, (1,)).item() + self._rng = np.random.default_rng(seed=seed) + + if self.aug_crop > 1 and self.seq_aug_crop: + self.delta_target_resolution = self._rng.integers(0, self.aug_crop) + + # over-loaded code + resolution = self._resolutions[ + ar_idx + ] # DO NOT CHANGE THIS (compatible with BatchedRandomSampler) + views = self._get_views(idx, resolution, self._rng, nview) + assert len(views) == nview + + if "camera_pose" not in views[0]: + views[0]["camera_pose"] = np.ones((4, 4), dtype=np.float32) + first_view_camera_pose = views[0]["camera_pose"] + transform = SeqColorJitter() if self.is_seq_color_jitter else self.transform + + for v, view in enumerate(views): + assert ( + "pts3d" not in view + ), f"pts3d should not be there, they will be computed afterwards based on intrinsics+depthmap for view {view_name(view)}" + view["idx"] = (idx, ar_idx, v) + + # encode the image + width, height = view["img"].size + + view["true_shape"] = np.int32((height, width)) + view["img"] = transform(view["img"]) + view["sky_mask"] = view["depthmap"] < 0 + + assert "camera_intrinsics" in view + if "camera_pose" not in view: + view["camera_pose"] = np.full((4, 4), np.nan, dtype=np.float32) + else: + assert np.isfinite( + view["camera_pose"] + ).all(), f"NaN in camera pose for view {view_name(view)}" + + ray_map = get_ray_map( + first_view_camera_pose, + view["camera_pose"], + view["camera_intrinsics"], + height, + width, + ) + view["ray_map"] = ray_map.astype(np.float32) + + assert "pts3d" not in view + assert "valid_mask" not in view + assert np.isfinite( + view["depthmap"] + ).all(), f"NaN in depthmap for view {view_name(view)}" + pts3d, valid_mask = depthmap_to_absolute_camera_coordinates(**view) + + view["pts3d"] = pts3d + view["valid_mask"] = valid_mask & np.isfinite(pts3d).all(axis=-1) + + # check all datatypes + for key, val in view.items(): + res, err_msg = is_good_type(key, val) + assert res, f"{err_msg} with {key}={val} for view {view_name(view)}" + K = view["camera_intrinsics"] + + if self.n_corres > 0: + ref_view = views[0] + for view in views: + corres1, corres2, valid = extract_correspondences_from_pts3d( + ref_view, view, self.n_corres, self._rng, nneg=self.nneg + ) + view["corres"] = (corres1, corres2) + view["valid_corres"] = valid + + # last thing done! + for view in views: + view["rng"] = int.from_bytes(self._rng.bytes(4), "big") + return views + + def _set_resolutions(self, resolutions): + assert resolutions is not None, "undefined resolution" + + if not isinstance(resolutions, list): + resolutions = [resolutions] + + self._resolutions = [] + for resolution in resolutions: + if isinstance(resolution, int): + width = height = resolution + else: + width, height = resolution + assert isinstance( + width, int + ), f"Bad type for {width=} {type(width)=}, should be int" + assert isinstance( + height, int + ), f"Bad type for {height=} {type(height)=}, should be int" + self._resolutions.append((width, height)) + + def _crop_resize_if_necessary( + self, image, depthmap, intrinsics, resolution, rng=None, info=None + ): + """This function: + - first downsizes the image with LANCZOS inteprolation, + which is better than bilinear interpolation in + """ + if not isinstance(image, PIL.Image.Image): + image = PIL.Image.fromarray(image) + + # downscale with lanczos interpolation so that image.size == resolution + # cropping centered on the principal point + W, H = image.size + cx, cy = intrinsics[:2, 2].round().astype(int) + min_margin_x = min(cx, W - cx) + min_margin_y = min(cy, H - cy) + assert min_margin_x > W / 5, f"Bad principal point in view={info}" + assert min_margin_y > H / 5, f"Bad principal point in view={info}" + # the new window will be a rectangle of size (2*min_margin_x, 2*min_margin_y) centered on (cx,cy) + l, t = cx - min_margin_x, cy - min_margin_y + r, b = cx + min_margin_x, cy + min_margin_y + crop_bbox = (l, t, r, b) + image, depthmap, intrinsics = cropping.crop_image_depthmap( + image, depthmap, intrinsics, crop_bbox + ) + + # transpose the resolution if necessary + W, H = image.size # new size + + # high-quality Lanczos down-scaling + target_resolution = np.array(resolution) + if self.aug_crop > 1: + target_resolution += ( + rng.integers(0, self.aug_crop) + if not self.seq_aug_crop + else self.delta_target_resolution + ) + image, depthmap, intrinsics = cropping.rescale_image_depthmap( + image, depthmap, intrinsics, target_resolution + ) + + # actual cropping (if necessary) with bilinear interpolation + intrinsics2 = cropping.camera_matrix_of_crop( + intrinsics, image.size, resolution, offset_factor=0.5 + ) + crop_bbox = cropping.bbox_from_intrinsics_in_out( + intrinsics, intrinsics2, resolution + ) + image, depthmap, intrinsics2 = cropping.crop_image_depthmap( + image, depthmap, intrinsics, crop_bbox + ) + + return image, depthmap, intrinsics2 + + +def is_good_type(key, v): + """returns (is_good, err_msg)""" + if isinstance(v, (str, int, tuple)): + return True, None + if v.dtype not in (np.float32, torch.float32, bool, np.int32, np.int64, np.uint8): + return False, f"bad {v.dtype=}" + return True, None + + +def view_name(view, batch_index=None): + def sel(x): + return x[batch_index] if batch_index not in (None, slice(None)) else x + + db = sel(view["dataset"]) + label = sel(view["label"]) + instance = sel(view["instance"]) + return f"{db}/{label}/{instance}" + + +def transpose_to_landscape(view): + height, width = view["true_shape"] + + if width < height: + # rectify portrait to landscape + assert view["img"].shape == (3, height, width) + view["img"] = view["img"].swapaxes(1, 2) + + assert view["valid_mask"].shape == (height, width) + view["valid_mask"] = view["valid_mask"].swapaxes(0, 1) + + assert view["depthmap"].shape == (height, width) + view["depthmap"] = view["depthmap"].swapaxes(0, 1) + + assert view["pts3d"].shape == (height, width, 3) + view["pts3d"] = view["pts3d"].swapaxes(0, 1) + + # transpose x and y pixels + view["camera_intrinsics"] = view["camera_intrinsics"][[1, 0, 2]] + + assert view["ray_map"].shape == (height, width, 6) + view["ray_map"] = view["ray_map"].swapaxes(0, 1) + + assert view["sky_mask"].shape == (height, width) + view["sky_mask"] = view["sky_mask"].swapaxes(0, 1) + + if "corres" in view: + # transpose correspondences x and y + view["corres"][0] = view["corres"][0][:, [1, 0]] + view["corres"][1] = view["corres"][1][:, [1, 0]] diff --git a/dust3r/datasets/base/batched_sampler.py b/dust3r/datasets/base/batched_sampler.py new file mode 100644 index 0000000000000000000000000000000000000000..b556e913c55791eea3323057402e9637abc9888a --- /dev/null +++ b/dust3r/datasets/base/batched_sampler.py @@ -0,0 +1,93 @@ +import numpy as np +import torch +from accelerate import Accelerator +import torch.utils +from torch.utils.data import BatchSampler, Sampler +import torch.utils.data + + +class CustomRandomSampler(Sampler): + """Random sampling under a constraint: each sample in the batch has the same feature, + which is chosen randomly from a known pool of 'features' for each batch. + + For instance, the 'feature' could be the image aspect-ratio. + + The index returned is a tuple (sample_idx, feat_idx). + This sampler ensures that each series of `batch_size` indices has the same `feat_idx`. + """ + + def __init__( + self, + dataset, + batch_size, + pool_size, + min_view_size, + max_view_size, + world_size, + warmup=1, + drop_last=True, + ): + self.batch_size = batch_size + self.pool_size = pool_size + self.min_view_size = min_view_size + self.max_view_size = max_view_size + self.drop_last = drop_last + self.len_dataset = N = len(dataset) + self.total_size = N + self.epoch = None + self.epochf = 0.0 + + def __len__(self): + return self.total_size + + def set_epoch(self, epoch): + self.epoch = epoch + + def __iter__(self): + if self.epoch is None: + raise ValueError( + "Epoch number not set. Please call 'set_epoch(epoch)' before iterating." + ) + + seed = self.epoch + 788 + rng = np.random.default_rng(seed=seed) + # random indices (will restart from 0 if not drop_last) + sample_idxs = np.arange(self.total_size) + rng.shuffle(sample_idxs) + # random feat_idxs (same across each batch) + n_batches = (self.total_size + self.batch_size - 1) // self.batch_size + if self.pool_size > 1: + p = np.ones(self.pool_size) + p[: self.pool_size // 2] *= 2 + p = p / p.sum() + _feat_idxs = rng.choice(self.pool_size, size=n_batches, p=p) + else: + _feat_idxs = rng.integers(self.pool_size, size=n_batches) + _feat_idxs = np.broadcast_to(_feat_idxs[:, None], (n_batches, self.batch_size)) + _feat_idxs = _feat_idxs.ravel()[: self.total_size] + _view_idxs = rng.integers( + self.min_view_size, self.max_view_size + 1, size=n_batches + ) + _view_idxs = np.broadcast_to(_view_idxs[:, None], (n_batches, self.batch_size)) + _view_idxs = _view_idxs.ravel()[: self.total_size] + + idxs = np.c_[sample_idxs, _feat_idxs, _view_idxs] + yield from (tuple(idx) for idx in idxs) + + +class BatchedRandomSampler(BatchSampler): + """Batch sampler that groups indices from RandomSampler into batches.""" + + def __init__(self, sampler: CustomRandomSampler, batch_size, drop_last=True): + self.sampler = sampler # An instance of RandomSampler + self.batch_size = batch_size + self.drop_last = drop_last + + def set_epoch(self, epoch): + self.sampler.set_epoch(epoch) + + +def round_by(total, multiple, up=False): + if up: + total = total + multiple - 1 + return (total // multiple) * multiple diff --git a/dust3r/datasets/base/easy_dataset.py b/dust3r/datasets/base/easy_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..67a57f57e89337727b732db348abafaa9a1a4335 --- /dev/null +++ b/dust3r/datasets/base/easy_dataset.py @@ -0,0 +1,198 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import numpy as np +from dust3r.datasets.base.batched_sampler import ( + BatchedRandomSampler, + CustomRandomSampler, +) +import torch + + +class EasyDataset: + """a dataset that you can easily resize and combine. + Examples: + --------- + 2 * dataset ==> duplicate each element 2x + + 10 @ dataset ==> set the size to 10 (random sampling, duplicates if necessary) + + dataset1 + dataset2 ==> concatenate datasets + """ + + def __add__(self, other): + return CatDataset([self, other]) + + def __rmul__(self, factor): + return MulDataset(factor, self) + + def __rmatmul__(self, factor): + return ResizedDataset(factor, self) + + def set_epoch(self, epoch): + pass # nothing to do by default + + def make_sampler( + self, batch_size, shuffle=True, drop_last=True, world_size=1, rank=0, fixed_length=False + ): + if not (shuffle): + raise NotImplementedError() # cannot deal yet + num_of_aspect_ratios = len(self._resolutions) + num_of_views = self.num_views + sampler = CustomRandomSampler( + self, + batch_size, + num_of_aspect_ratios, + 4 if not fixed_length else num_of_views, + num_of_views, + world_size, + warmup=1, + drop_last=drop_last, + ) + return BatchedRandomSampler(sampler, batch_size, drop_last) + + +class MulDataset(EasyDataset): + """Artifically augmenting the size of a dataset.""" + + multiplicator: int + + def __init__(self, multiplicator, dataset): + assert isinstance(multiplicator, int) and multiplicator > 0 + self.multiplicator = multiplicator + self.dataset = dataset + + def __len__(self): + return self.multiplicator * len(self.dataset) + + def __repr__(self): + return f"{self.multiplicator}*{repr(self.dataset)}" + + def __getitem__(self, idx): + if isinstance(idx, tuple): + idx, other, another = idx + return self.dataset[idx // self.multiplicator, other, another] + else: + return self.dataset[idx // self.multiplicator] + + @property + def _resolutions(self): + return self.dataset._resolutions + + @property + def num_views(self): + return self.dataset.num_views + + +class ResizedDataset(EasyDataset): + """Artifically changing the size of a dataset.""" + + new_size: int + + def __init__(self, new_size, dataset): + assert isinstance(new_size, int) and new_size > 0 + self.new_size = new_size + self.dataset = dataset + + def __len__(self): + return self.new_size + + def __repr__(self): + size_str = str(self.new_size) + for i in range((len(size_str) - 1) // 3): + sep = -4 * i - 3 + size_str = size_str[:sep] + "_" + size_str[sep:] + return f"{size_str} @ {repr(self.dataset)}" + + def set_epoch(self, epoch): + # this random shuffle only depends on the epoch + rng = np.random.default_rng(seed=epoch + 777) + + # shuffle all indices + perm = rng.permutation(len(self.dataset)) + + # rotary extension until target size is met + shuffled_idxs = np.concatenate( + [perm] * (1 + (len(self) - 1) // len(self.dataset)) + ) + self._idxs_mapping = shuffled_idxs[: self.new_size] + + assert len(self._idxs_mapping) == self.new_size + + def __getitem__(self, idx): + assert hasattr( + self, "_idxs_mapping" + ), "You need to call dataset.set_epoch() to use ResizedDataset.__getitem__()" + if isinstance(idx, tuple): + idx, other, another = idx + return self.dataset[self._idxs_mapping[idx], other, another] + else: + return self.dataset[self._idxs_mapping[idx]] + + @property + def _resolutions(self): + return self.dataset._resolutions + + @property + def num_views(self): + return self.dataset.num_views + + +class CatDataset(EasyDataset): + """Concatenation of several datasets""" + + def __init__(self, datasets): + for dataset in datasets: + assert isinstance(dataset, EasyDataset) + self.datasets = datasets + self._cum_sizes = np.cumsum([len(dataset) for dataset in datasets]) + + def __len__(self): + return self._cum_sizes[-1] + + def __repr__(self): + # remove uselessly long transform + return " + ".join( + repr(dataset).replace( + ",transform=Compose( ToTensor() Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)))", + "", + ) + for dataset in self.datasets + ) + + def set_epoch(self, epoch): + for dataset in self.datasets: + dataset.set_epoch(epoch) + + def __getitem__(self, idx): + other = None + if isinstance(idx, tuple): + idx, other, another = idx + + if not (0 <= idx < len(self)): + raise IndexError() + + db_idx = np.searchsorted(self._cum_sizes, idx, "right") + dataset = self.datasets[db_idx] + new_idx = idx - (self._cum_sizes[db_idx - 1] if db_idx > 0 else 0) + + if other is not None and another is not None: + new_idx = (new_idx, other, another) + return dataset[new_idx] + + @property + def _resolutions(self): + resolutions = self.datasets[0]._resolutions + for dataset in self.datasets[1:]: + assert tuple(dataset._resolutions) == tuple(resolutions) + return resolutions + + @property + def num_views(self): + num_views = self.datasets[0].num_views + for dataset in self.datasets[1:]: + assert dataset.num_views == num_views + return num_views diff --git a/dust3r/datasets/bedlam.py b/dust3r/datasets/bedlam.py new file mode 100644 index 0000000000000000000000000000000000000000..f680a29fd8b446d30db51d531a939de5abf9e521 --- /dev/null +++ b/dust3r/datasets/bedlam.py @@ -0,0 +1,297 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + +invalid_seqs = [ + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000042", + "20221024_10_100_batch01handhair_zoom_suburb_d_seq_000059", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000079", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000978", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000081", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000268", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000089", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000189", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000034", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000889", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000293", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000067", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000904", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000434", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000044", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000013", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000396", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000012", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000082", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000120", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000324", + "20221013_3_250_batch01hand_static_bigOffice_seq_000038", + "20221012_3-10_500_batch01hand_zoom_highSchoolGym_seq_000486", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000421", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000226", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000012", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000149", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000311", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000080", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000122", + "20221012_3-10_500_batch01hand_zoom_highSchoolGym_seq_000079", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000077", + "20221014_3_250_batch01hand_orbit_archVizUI3_time15_seq_000095", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000062", + "20221013_3_250_batch01hand_static_bigOffice_seq_000015", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000095", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000119", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000297", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000011", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000196", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000316", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000283", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000085", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000287", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000163", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000804", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000842", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000027", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000182", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000982", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000029", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000031", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000025", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000250", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000785", + "20221024_10_100_batch01handhair_zoom_suburb_d_seq_000069", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000122", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000246", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000352", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000425", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000192", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000900", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000043", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000063", + "20221014_3_250_batch01hand_orbit_archVizUI3_time15_seq_000096", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000091", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000013", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000309", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000114", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000969", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000361", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000267", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000083", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000383", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000890", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000003", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000045", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000317", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000076", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000082", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000907", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000279", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000076", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000004", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000061", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000811", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000800", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000841", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000794", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000308", + "20221024_10_100_batch01handhair_zoom_suburb_d_seq_000064", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000284", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000752", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000269", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000036", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000419", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000290", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000322", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000818", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000327", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000326", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000002", + "20221024_10_100_batch01handhair_zoom_suburb_d_seq_000060", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000348", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000059", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000016", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000817", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000332", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000094", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000193", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000779", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000177", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000368", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000023", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000024", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000310", + "20221014_3_250_batch01hand_orbit_archVizUI3_time15_seq_000086", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000038", + "20221024_10_100_batch01handhair_zoom_suburb_d_seq_000071", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000768", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000017", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000053", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000097", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000856", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000827", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000161", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000084", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000106", + "20221013_3_250_batch01hand_orbit_bigOffice_seq_000207", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000007", + "20221024_3-10_100_batch01handhair_static_highSchoolGym_seq_000013", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000251", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000796", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000105", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000251", + "20221019_3-8_250_highbmihand_orbit_stadium_seq_000046", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000334", + "20221019_3-8_1000_highbmihand_static_suburb_d_seq_000453", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000373", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000283", + "20221010_3-10_500_batch01hand_zoom_suburb_d_seq_000249", +] +hdri_scenes = [ + "20221010_3_1000_batch01hand", + "20221017_3_1000_batch01hand", + "20221018_3-8_250_batch01hand", + "20221019_3_250_highbmihand", +] + + +class BEDLAM_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.pose_root = os.path.join( + os.path.dirname(ROOT), f"{os.path.basename(ROOT)}_pose" + ) + assert os.path.exists(self.pose_root) + self.video = True + self.is_metric = True + self.max_interval = 4 + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = os.listdir(self.ROOT) + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + if scene in invalid_seqs: + continue + if any([scene.startswith(x) for x in hdri_scenes]): + continue + if "closeup" in scene: + continue + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")] + ) + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + assert len(set(self.scenes) - set(os.listdir(self.pose_root))) == 0 + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=1.0, + fix_interval_prob=1.0, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(osp.join(self.pose_root, self.scenes[scene_id]), "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".png")) + # Load depthmap + depthmap = np.load(osp.join(depth_dir, basename + ".npy")) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + depthmap[depthmap > 200.0] = 0.0 + + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.85, 0.10, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="BEDLAM", + label=self.scenes[scene_id] + "_" + basename, + instance=osp.join(rgb_dir, basename + ".png"), + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(1, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/blendedmvs.py b/dust3r/datasets/blendedmvs.py new file mode 100644 index 0000000000000000000000000000000000000000..e9c290bbc1a6c0535e2267dc4e5eb0ecc62b6019 --- /dev/null +++ b/dust3r/datasets/blendedmvs.py @@ -0,0 +1,305 @@ +import os.path as osp +import numpy as np +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 +import h5py +from tqdm import tqdm + + +class BlendedMVS_Multi(BaseMultiViewDataset): + """Dataset of outdoor street scenes, 5 images each time""" + + def __init__(self, *args, ROOT, split=None, **kwargs): + self.ROOT = ROOT + self.video = False + self.is_metric = False + super().__init__(*args, **kwargs) + # assert split is None + self._load_data() + + def _load_data(self): + self.data_dict = self.read_h5_file(os.path.join(self.ROOT, "new_overlap.h5")) + self.num_imgs = sum( + [len(self.data_dict[s]["basenames"]) for s in self.data_dict.keys()] + ) + self.num_scenes = len(self.data_dict.keys()) + self.invalid_scenes = [] + self.is_reachable_cache = {scene: {} for scene in self.data_dict.keys()} + + def read_h5_file(self, h5_file_path): + data_dict = {} + self.all_ref_imgs = [] + with h5py.File(h5_file_path, "r") as f: + for scene_dir in tqdm(f.keys()): + group = f[scene_dir] + basenames = group["basenames"][:] + indices = group["indices"][:] + values = group["values"][:] + shape = group.attrs["shape"] + # Reconstruct the sparse matrix + score_matrix = np.zeros(shape, dtype=np.float32) + score_matrix[indices[0], indices[1]] = values + data_dict[scene_dir] = { + "basenames": basenames, + "score_matrix": self.build_adjacency_list(score_matrix), + } + self.all_ref_imgs.extend( + [(scene_dir, b) for b in range(len(basenames))] + ) + return data_dict + + @staticmethod + def build_adjacency_list(S, thresh=0.2): + adjacency_list = [[] for _ in range(len(S))] + S = S - thresh + S[S < 0] = 0 + rows, cols = np.nonzero(S) + for i, j in zip(rows, cols): + adjacency_list[i].append((j, S[i][j])) + return adjacency_list + + @staticmethod + def is_reachable(adjacency_list, start_index, k): + visited = set() + stack = [start_index] + while stack and len(visited) < k: + node = stack.pop() + if node not in visited: + visited.add(node) + for neighbor in adjacency_list[node]: + if neighbor[0] not in visited: + stack.append(neighbor[0]) + return len(visited) >= k + + @staticmethod + def random_sequence_no_revisit_with_backtracking( + adjacency_list, k, start_index, rng: np.random.Generator + ): + path = [start_index] + visited = set([start_index]) + + neighbor_iterators = [] + # Initialize the iterator for the start index + neighbors = adjacency_list[start_index] + neighbor_idxs = [n[0] for n in neighbors] + neighbor_weights = [n[1] for n in neighbors] + neighbor_idxs = rng.choice( + neighbor_idxs, + size=len(neighbor_idxs), + replace=False, + p=np.array(neighbor_weights) / np.sum(neighbor_weights), + ).tolist() + neighbor_iterators.append(iter(neighbor_idxs)) + + while len(path) < k: + if not neighbor_iterators: + # No possible sequence + return None + current_iterator = neighbor_iterators[-1] + try: + next_index = next(current_iterator) + if next_index not in visited: + path.append(next_index) + visited.add(next_index) + + # Prepare iterator for the next node + neighbors = adjacency_list[next_index] + neighbor_idxs = [n[0] for n in neighbors] + neighbor_weights = [n[1] for n in neighbors] + neighbor_idxs = rng.choice( + neighbor_idxs, + size=len(neighbor_idxs), + replace=False, + p=np.array(neighbor_weights) / np.sum(neighbor_weights), + ).tolist() + neighbor_iterators.append(iter(neighbor_idxs)) + except StopIteration: + # No more neighbors to try at this node, backtrack + neighbor_iterators.pop() + visited.remove(path.pop()) + return path + + @staticmethod + def random_sequence_with_optional_repeats( + adjacency_list, + k, + start_index, + rng: np.random.Generator, + max_k=None, + max_attempts=100, + ): + if max_k is None: + max_k = k + path = [start_index] + visited = set([start_index]) + current_index = start_index + attempts = 0 + + while len(path) < max_k and attempts < max_attempts: + attempts += 1 + neighbors = adjacency_list[current_index] + neighbor_idxs = [n[0] for n in neighbors] + neighbor_weights = [n[1] for n in neighbors] + + if not neighbor_idxs: + # No neighbors, cannot proceed further + break + + # Try to find unvisited neighbors + unvisited_neighbors = [ + (idx, wgt) + for idx, wgt in zip(neighbor_idxs, neighbor_weights) + if idx not in visited + ] + if unvisited_neighbors: + # Select among unvisited neighbors + unvisited_idxs = [idx for idx, _ in unvisited_neighbors] + unvisited_weights = [wgt for _, wgt in unvisited_neighbors] + probabilities = np.array(unvisited_weights) / np.sum(unvisited_weights) + next_index = rng.choice(unvisited_idxs, p=probabilities) + visited.add(next_index) + else: + # All neighbors visited, but we need to reach length max_k + # So we can revisit nodes + probabilities = np.array(neighbor_weights) / np.sum(neighbor_weights) + next_index = rng.choice(neighbor_idxs, p=probabilities) + + path.append(next_index) + current_index = next_index + + if len(set(path)) >= k: + # If path is shorter than max_k, extend it by repeating existing elements + while len(path) < max_k: + # Randomly select nodes from the existing path to repeat + next_index = rng.choice(path) + path.append(next_index) + return path + else: + # Could not reach k unique nodes + return None + + def __len__(self): + return len(self.all_ref_imgs) + + def get_image_num(self): + return self.num_imgs + + def get_stats(self): + return f"{len(self)} imgs from {self.num_scenes} scenes" + + def generate_sequence( + self, scene, adj_list, num_views, start_index, rng, allow_repeat=False + ): + cutoff = num_views if not allow_repeat else max(num_views // 5, 3) + if start_index in self.is_reachable_cache[scene]: + if not self.is_reachable_cache[scene][start_index]: + print( + f"Cannot reach {num_views} unique elements from index {start_index}." + ) + return None + else: + self.is_reachable_cache[scene][start_index] = self.is_reachable( + adj_list, start_index, cutoff + ) + if not self.is_reachable_cache[scene][start_index]: + print( + f"Cannot reach {num_views} unique elements from index {start_index}." + ) + return None + if not allow_repeat: + sequence = self.random_sequence_no_revisit_with_backtracking( + adj_list, cutoff, start_index, rng + ) + else: + sequence = self.random_sequence_with_optional_repeats( + adj_list, cutoff, start_index, rng, max_k=num_views + ) + if not sequence: + self.is_reachable_cache[scene][start_index] = False + print("Failed to generate a sequence without revisiting.") + return sequence + + def _get_views(self, idx, resolution, rng: np.random.Generator, num_views): + scene_info, ref_img_idx = self.all_ref_imgs[idx] + invalid_seq = True + ordered_video = False + + while invalid_seq: + basenames = self.data_dict[scene_info]["basenames"] + if ( + sum( + [ + (1 - int(x)) + for x in list(self.is_reachable_cache[scene_info].values()) + ] + ) + > len(basenames) - self.num_views + ): + self.invalid_scenes.append(scene_info) + while scene_info in self.invalid_scenes: + idx = rng.integers(low=0, high=len(self.all_ref_imgs)) + scene_info, ref_img_idx = self.all_ref_imgs[idx] + basenames = self.data_dict[scene_info]["basenames"] + + score_matrix = self.data_dict[scene_info]["score_matrix"] + imgs_idxs = self.generate_sequence( + scene_info, score_matrix, num_views, ref_img_idx, rng, self.allow_repeat + ) + + if imgs_idxs is None: + random_direction = 2 * rng.choice(2) - 1 + for offset in range(1, len(basenames)): + tentative_im_idx = ( + ref_img_idx + (random_direction * offset) + ) % len(basenames) + if ( + tentative_im_idx not in self.is_reachable_cache[scene_info] + or self.is_reachable_cache[scene_info][tentative_im_idx] + ): + ref_img_idx = tentative_im_idx + break + else: + invalid_seq = False + views = [] + for view_idx in imgs_idxs: + scene_dir = osp.join(self.ROOT, scene_info) + impath = basenames[view_idx].decode("utf-8") + image = imread_cv2(osp.join(scene_dir, impath + ".jpg")) + depthmap = imread_cv2(osp.join(scene_dir, impath + ".exr")) + camera_params = np.load(osp.join(scene_dir, impath + ".npz")) + + intrinsics = np.float32(camera_params["intrinsics"]) + camera_pose = np.eye(4, dtype=np.float32) + camera_pose[:3, :3] = camera_params["R_cam2world"] + camera_pose[:3, 3] = camera_params["t_cam2world"] + + image, depthmap, intrinsics = self._crop_resize_if_necessary( + image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath) + ) + + views.append( + dict( + img=image, + depthmap=depthmap, + camera_pose=camera_pose, # cam2world + camera_intrinsics=intrinsics, + dataset="BlendedMVS", + label=osp.relpath(scene_dir, self.ROOT), + is_metric=self.is_metric, + is_video=ordered_video, + instance=osp.join(scene_dir, impath + ".jpg"), + quantile=np.array(0.97, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/co3d.py b/dust3r/datasets/co3d.py new file mode 100644 index 0000000000000000000000000000000000000000..98dcc820fcd70fd496396ef000c22aeb2adee35a --- /dev/null +++ b/dust3r/datasets/co3d.py @@ -0,0 +1,190 @@ +import os.path as osp +import json +import itertools +from collections import deque +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np +import time + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class Co3d_Multi(BaseMultiViewDataset): + def __init__(self, mask_bg="rand", *args, ROOT, **kwargs): + self.ROOT = ROOT + super().__init__(*args, **kwargs) + assert mask_bg in (True, False, "rand") + self.mask_bg = mask_bg + self.is_metric = False + self.dataset_label = "Co3d_v2" + + # load all scenes + with open(osp.join(self.ROOT, f"selected_seqs_{self.split}.json"), "r") as f: + self.scenes = json.load(f) + self.scenes = {k: v for k, v in self.scenes.items() if len(v) > 0} + self.scenes = { + (k, k2): v2 for k, v in self.scenes.items() for k2, v2 in v.items() + } + self.scene_list = list(self.scenes.keys()) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + self.cut_off = cut_off + self.all_ref_imgs = [ + (key, value) + for key, values in self.scenes.items() + for value in values[: len(values) - cut_off + 1] + ] + self.invalidate = {scene: {} for scene in self.scene_list} + self.invalid_scenes = {scene: False for scene in self.scene_list} + + def __len__(self): + return len(self.all_ref_imgs) + + def _get_metadatapath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "images", f"frame{view_idx:06n}.npz") + + def _get_impath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "images", f"frame{view_idx:06n}.jpg") + + def _get_depthpath(self, obj, instance, view_idx): + return osp.join( + self.ROOT, obj, instance, "depths", f"frame{view_idx:06n}.jpg.geometric.png" + ) + + def _get_maskpath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "masks", f"frame{view_idx:06n}.png") + + def _read_depthmap(self, depthpath, input_metadata): + depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED) + depthmap = (depthmap.astype(np.float32) / 65535) * np.nan_to_num( + input_metadata["maximum_depth"] + ) + return depthmap + + def _get_views(self, idx, resolution, rng, num_views): + invalid_seq = True + scene_info, ref_img_idx = self.all_ref_imgs[idx] + + while invalid_seq: + while self.invalid_scenes[scene_info]: + idx = rng.integers(low=0, high=len(self.all_ref_imgs)) + scene_info, ref_img_idx = self.all_ref_imgs[idx] + + obj, instance = scene_info + + image_pool = self.scenes[obj, instance] + if len(image_pool) < self.cut_off: + print("Invalid scene!") + self.invalid_scenes[scene_info] = True + continue + + imgs_idxs, ordered_video = self.get_seq_from_start_id( + num_views, ref_img_idx, image_pool, rng + ) + + if resolution not in self.invalidate[obj, instance]: # flag invalid images + self.invalidate[obj, instance][resolution] = [ + False for _ in range(len(image_pool)) + ] + # decide now if we mask the bg + mask_bg = (self.mask_bg == True) or ( + self.mask_bg == "rand" and rng.choice(2, p=[0.9, 0.1]) + ) + views = [] + + imgs_idxs = deque(imgs_idxs) + + while len(imgs_idxs) > 0: # some images (few) have zero depth + if ( + len(image_pool) - sum(self.invalidate[obj, instance][resolution]) + < self.cut_off + ): + print("Invalid scene!") + invalid_seq = True + self.invalid_scenes[scene_info] = True + break + + im_idx = imgs_idxs.pop() + if self.invalidate[obj, instance][resolution][im_idx]: + # search for a valid image + ordered_video = False + random_direction = 2 * rng.choice(2) - 1 + for offset in range(1, len(image_pool)): + tentative_im_idx = (im_idx + (random_direction * offset)) % len( + image_pool + ) + if not self.invalidate[obj, instance][resolution][ + tentative_im_idx + ]: + im_idx = tentative_im_idx + break + view_idx = image_pool[im_idx] + impath = self._get_impath(obj, instance, view_idx) + depthpath = self._get_depthpath(obj, instance, view_idx) + + # load camera params + metadata_path = self._get_metadatapath(obj, instance, view_idx) + input_metadata = np.load(metadata_path) + camera_pose = input_metadata["camera_pose"].astype(np.float32) + intrinsics = input_metadata["camera_intrinsics"].astype(np.float32) + + # load image and depth + rgb_image = imread_cv2(impath) + depthmap = self._read_depthmap(depthpath, input_metadata) + + if mask_bg: + # load object mask + maskpath = self._get_maskpath(obj, instance, view_idx) + maskmap = imread_cv2(maskpath, cv2.IMREAD_UNCHANGED).astype( + np.float32 + ) + maskmap = (maskmap / 255.0) > 0.1 + + # update the depthmap with mask + depthmap *= maskmap + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath + ) + num_valid = (depthmap > 0.0).sum() + if num_valid == 0: + # problem, invalidate image and retry + self.invalidate[obj, instance][resolution][im_idx] = True + imgs_idxs.append(im_idx) + continue + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, len(views), rng + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap, + camera_pose=camera_pose, + camera_intrinsics=intrinsics, + dataset=self.dataset_label, + label=osp.join(obj, instance), + instance=osp.split(impath)[1], + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.9, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + + if len(views) == num_views and not all( + [view["instance"] == views[0]["instance"] for view in views] + ): + invalid_seq = False + return views diff --git a/dust3r/datasets/cop3d.py b/dust3r/datasets/cop3d.py new file mode 100644 index 0000000000000000000000000000000000000000..aa93c7d109f80d70869250b8a44daf59cf202e0f --- /dev/null +++ b/dust3r/datasets/cop3d.py @@ -0,0 +1,110 @@ +import os.path as osp +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np + +from dust3r.datasets.co3d import Co3d_Multi +from dust3r.utils.image import imread_cv2 + + +class Cop3D_Multi(Co3d_Multi): + def __init__(self, mask_bg="rand", *args, ROOT, **kwargs): + super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs) + self.dataset_label = "Cop3D" + self.is_metric = False + + def _get_metadatapath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "images", f"frame{view_idx:06n}.npz") + + def _get_impath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "images", f"frame{view_idx:06n}.jpg") + + def _get_depthpath(self, obj, instance, view_idx): + # no depth, pseduo path just for getting the right resolution + return osp.join(self.ROOT, obj, instance, "images", f"frame{view_idx:06n}.jpg") + + def _get_maskpath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "masks", f"frame{view_idx:06n}.png") + + def _read_depthmap(self, impath, input_metadata): + # no depth, set to all ones + img = imread_cv2(impath, cv2.IMREAD_UNCHANGED) + depthmap = np.ones_like(img[..., 0], dtype=np.float32) + return depthmap + + def _get_views(self, idx, resolution, rng, num_views): + invalid_seq = True + scene_info, ref_img_idx = self.all_ref_imgs[idx] + + while invalid_seq: + while self.invalid_scenes[scene_info]: + idx = rng.integers(low=0, high=len(self.all_ref_imgs)) + scene_info, ref_img_idx = self.all_ref_imgs[idx] + + obj, instance = scene_info + + image_pool = self.scenes[obj, instance] + if len(image_pool) < self.num_views: + print("Invalid scene!") + self.invalid_scenes[scene_info] = True + continue + + imgs_idxs, ordered_video = self.get_seq_from_start_id( + num_views, + ref_img_idx, + image_pool, + rng, + max_interval=5, + video_prob=1.0, + fix_interval_prob=0.9, + ) + + views = [] + + for im_idx in imgs_idxs: + view_idx = image_pool[im_idx] + impath = self._get_impath(obj, instance, view_idx) + depthpath = self._get_depthpath(obj, instance, view_idx) + + # load camera params + metadata_path = self._get_metadatapath(obj, instance, view_idx) + input_metadata = np.load(metadata_path) + camera_pose = input_metadata["camera_pose"].astype(np.float32) + intrinsics = input_metadata["camera_intrinsics"].astype(np.float32) + + # load image and depth + rgb_image = imread_cv2(impath) + depthmap = self._read_depthmap(depthpath, input_metadata) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=impath + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap, + camera_pose=camera_pose, + camera_intrinsics=intrinsics, + dataset=self.dataset_label, + label=osp.join(obj, instance), + instance=osp.split(impath)[1], + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.96, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=True, + depth_only=False, + single_view=False, + reset=False, + ) + ) + + if len(views) == num_views and not all( + [view["instance"] == views[0]["instance"] for view in views] + ): + invalid_seq = False + return views diff --git a/dust3r/datasets/dl3dv.py b/dust3r/datasets/dl3dv.py new file mode 100644 index 0000000000000000000000000000000000000000..2650d573123b86f10c99bb663ec399372808fe37 --- /dev/null +++ b/dust3r/datasets/dl3dv.py @@ -0,0 +1,166 @@ +import os.path as osp +import os +import sys +import itertools + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class DL3DV_Multi(BaseMultiViewDataset): + def __init__(self, *args, split, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.max_interval = 20 + self.is_metric = False + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data() + + def _load_data(self): + self.all_scenes = sorted( + [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))] + ) + subscenes = [] + for scene in self.all_scenes: + # not empty + subscenes.extend( + [ + osp.join(scene, f) + for f in os.listdir(osp.join(self.ROOT, scene)) + if os.path.isdir(osp.join(self.ROOT, scene, f)) + and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0 + ] + ) + + offset = 0 + scenes = [] + sceneids = [] + images = [] + scene_img_list = [] + start_img_ids = [] + j = 0 + + for scene_idx, scene in enumerate(subscenes): + scene_dir = osp.join(self.ROOT, scene, "dense") + rgb_paths = sorted( + [ + f + for f in os.listdir(os.path.join(scene_dir, "rgb")) + if f.endswith(".png") + ] + ) + assert len(rgb_paths) > 0, f"{scene_dir} is empty." + num_imgs = len(rgb_paths) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + scenes.append(scene) + scene_img_list.append(img_ids) + sceneids.extend([j] * num_imgs) + images.extend(rgb_paths) + start_img_ids.extend(start_img_ids_) + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + scene_id = self.sceneids[start_id] + all_image_ids = self.scene_img_list[scene_id] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + block_shuffle=25, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for view_idx in image_idxs: + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id], "dense") + + rgb_path = self.images[view_idx] + basename = rgb_path[:-4] + + rgb_image = imread_cv2( + osp.join(scene_dir, "rgb", rgb_path), cv2.IMREAD_COLOR + ) + depthmap = np.load(osp.join(scene_dir, "depth", basename + ".npy")).astype( + np.float32 + ) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + cam_file = np.load(osp.join(scene_dir, "cam", basename + ".npz")) + sky_mask = ( + cv2.imread( + osp.join(scene_dir, "sky_mask", rgb_path), cv2.IMREAD_UNCHANGED + ) + >= 127 + ) + outlier_mask = cv2.imread( + osp.join(scene_dir, "outlier_mask", rgb_path), cv2.IMREAD_UNCHANGED + ) + depthmap[sky_mask] = -1.0 + depthmap[outlier_mask >= 127] = 0.0 + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + threshold = ( + np.percentile(depthmap[depthmap > 0], 98) + if depthmap[depthmap > 0].size > 0 + else 0 + ) + depthmap[depthmap > threshold] = 0.0 + + intrinsics = cam_file["intrinsic"].astype(np.float32) + camera_pose = cam_file["pose"].astype(np.float32) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="dl3dv", + label=self.scenes[scene_id] + "_" + rgb_path, + instance=osp.join(scene_dir, "rgb", rgb_path), + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.9, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + return views diff --git a/dust3r/datasets/dynamic_replica.py b/dust3r/datasets/dynamic_replica.py new file mode 100644 index 0000000000000000000000000000000000000000..1d816e58be6518e1274fa84fa8c6a7cae73741ca --- /dev/null +++ b/dust3r/datasets/dynamic_replica.py @@ -0,0 +1,137 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class DynamicReplica(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 16 + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data(self.split) + + def _load_data(self, split): + self.scenes = os.listdir(os.path.join(self.ROOT, split)) + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, self.split, scene, "left") + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")], + key=lambda x: float(x), + ) + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=1.0, + fix_interval_prob=1.0, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id], "left") + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".png")) + # Load depthmap + depthmap = np.load(osp.join(depth_dir, basename + ".npy")) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.85, 0.10, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="dynamic_replica", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/eden.py b/dust3r/datasets/eden.py new file mode 100644 index 0000000000000000000000000000000000000000..00af2fffc73535f436557929b1b0220737907b2b --- /dev/null +++ b/dust3r/datasets/eden.py @@ -0,0 +1,94 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class EDEN_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + scenes = os.listdir(self.ROOT) + img_names = [] + for scene in scenes: + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")] + ) + img_names.extend([(scene, basename) for basename in basenames]) + + self.img_names = img_names + + def __len__(self): + return len(self.img_names) + + def get_image_num(self): + return len(self.img_names) + + def _get_views(self, idx, resolution, rng, num_views): + new_seed = rng.integers(0, 2**32) + idx + new_rng = np.random.default_rng(new_seed) + img_names = new_rng.permutation(self.img_names) + + views = [] + i = 0 + while len(views) < num_views: + # Load RGB image + scene, img_name = img_names[i] + try: + rgb_image = imread_cv2( + osp.join(self.ROOT, scene, "rgb", f"{img_name}.png") + ) + depthmap = np.load( + osp.join(self.ROOT, scene, "depth", f"{img_name}.npy") + ) + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + + intrinsics = np.load( + osp.join(self.ROOT, scene, "cam", f"{img_name}.npz") + )["intrinsics"] + # camera pose is not provided, placeholder + camera_pose = np.eye(4) + except: + i += 1 + continue + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="EDEN", + label=img_name, + instance=osp.join(self.ROOT, scene, "rgb", f"{img_name}.png"), + is_metric=self.is_metric, + is_video=False, + quantile=np.array(1.0, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=True, + reset=True, + ) + ) + i += 1 + return views diff --git a/dust3r/datasets/hoi4d.py b/dust3r/datasets/hoi4d.py new file mode 100644 index 0000000000000000000000000000000000000000..b602df5d4dd1493d02377039379fd2ffb3b08ba2 --- /dev/null +++ b/dust3r/datasets/hoi4d.py @@ -0,0 +1,84 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys +sys.path.append(osp.join(osp.dirname(__file__), '..','..')) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class HOI4D_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + scenes = os.listdir(self.ROOT) + img_names = [] + for scene in scenes: + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, 'rgb') + basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith('.png')]) + img_names.extend([(scene, basename) for basename in basenames]) + + self.img_names = img_names + + def __len__(self): + return len(self.img_names) + + def get_image_num(self): + return len(self.img_names) + + def _get_views(self, idx, resolution, rng, num_views): + new_seed = rng.integers(0, 2**32) + idx + new_rng = np.random.default_rng(new_seed) + invalid_seq = True + while invalid_seq: + img_names = new_rng.choice(self.img_names, num_views, replace=False) + + views = [] + for v, img_name in enumerate(img_names): + # Load RGB image + scene, img_name = img_name + try: + rgb_image = imread_cv2(osp.join(self.ROOT, scene, "rgb", f"{img_name}.png")) + depthmap = np.load(osp.join(self.ROOT, scene, "depth", f"{img_name}.npy")) + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + + intrinsics = np.load(osp.join(self.ROOT, scene, "cam", f"{img_name}.npz"))["intrinsics"] + except: + print(f"Error loading {scene} {img_name}, skipping") + break + # camera pose is not provided, placeholder + camera_pose = np.eye(4) + + rgb_image, depthmap, intrinsics= self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name) + + views.append(dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset='HOI4D', + label=img_name, + instance=osp.join(self.ROOT, scene, "rgb", f"{img_name}.png"), + is_metric=self.is_metric, + is_video=False, + quantile=np.array(0.99, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=True, + reset=True, + )) + if len(views) == num_views: + invalid_seq = False + return views diff --git a/dust3r/datasets/hypersim.py b/dust3r/datasets/hypersim.py new file mode 100644 index 0000000000000000000000000000000000000000..c194df6db72525f2f164297dd4198a27085ce95c --- /dev/null +++ b/dust3r/datasets/hypersim.py @@ -0,0 +1,141 @@ +import os.path as osp +import os +import sys +import itertools + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class HyperSim_Multi(BaseMultiViewDataset): + def __init__(self, *args, split, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 4 + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data() + + def _load_data(self): + self.all_scenes = sorted( + [f for f in os.listdir(self.ROOT) if os.path.isdir(osp.join(self.ROOT, f))] + ) + subscenes = [] + for scene in self.all_scenes: + # not empty + subscenes.extend( + [ + osp.join(scene, f) + for f in os.listdir(osp.join(self.ROOT, scene)) + if os.path.isdir(osp.join(self.ROOT, scene, f)) + and len(os.listdir(osp.join(self.ROOT, scene, f))) > 0 + ] + ) + + offset = 0 + scenes = [] + sceneids = [] + images = [] + start_img_ids = [] + scene_img_list = [] + j = 0 + for scene_idx, scene in enumerate(subscenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_paths = sorted([f for f in os.listdir(scene_dir) if f.endswith(".png")]) + assert len(rgb_paths) > 0, f"{scene_dir} is empty." + num_imgs = len(rgb_paths) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + scenes.append(scene) + scene_img_list.append(img_ids) + sceneids.extend([j] * num_imgs) + images.extend(rgb_paths) + start_img_ids.extend(start_img_ids_) + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.scene_img_list = scene_img_list + self.start_img_ids = start_img_ids + + def __len__(self): + return len(self.start_img_ids) * 10 + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + idx = idx // 10 + start_id = self.start_img_ids[idx] + scene_id = self.sceneids[start_id] + all_image_ids = self.scene_img_list[scene_id] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + block_shuffle=16, + ) + image_idxs = np.array(all_image_ids)[pos] + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + + rgb_path = self.images[view_idx] + depth_path = rgb_path.replace("rgb.png", "depth.npy") + cam_path = rgb_path.replace("rgb.png", "cam.npz") + + rgb_image = imread_cv2(osp.join(scene_dir, rgb_path), cv2.IMREAD_COLOR) + depthmap = np.load(osp.join(scene_dir, depth_path)).astype(np.float32) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + cam_file = np.load(osp.join(scene_dir, cam_path)) + intrinsics = cam_file["intrinsics"].astype(np.float32) + camera_pose = cam_file["pose"].astype(np.float32) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="hypersim", + label=self.scenes[scene_id] + "_" + rgb_path, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/irs.py b/dust3r/datasets/irs.py new file mode 100644 index 0000000000000000000000000000000000000000..52baa76d6f6a952dc5fa69aeab6b45239cc6b549 --- /dev/null +++ b/dust3r/datasets/irs.py @@ -0,0 +1,86 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class IRS(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = False + self.is_metric = True + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + scenes = os.listdir(self.ROOT) + img_names = [] + for scene in scenes: + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")] + ) + img_names.extend([(scene, basename) for basename in basenames]) + + self.img_names = img_names + + def __len__(self): + return len(self.img_names) + + def get_image_num(self): + return len(self.img_names) + + def _get_views(self, idx, resolution, rng, num_views): + new_seed = rng.integers(0, 2**32) + idx + new_rng = np.random.default_rng(new_seed) + img_names = new_rng.choice(self.img_names, num_views, replace=False) + + views = [] + for v, img_name in enumerate(img_names): + # Load RGB image + scene, img_name = img_name + rgb_image = imread_cv2(osp.join(self.ROOT, scene, "rgb", f"{img_name}.png")) + depthmap = np.load(osp.join(self.ROOT, scene, "depth", f"{img_name}.npy")) + depthmap[depthmap > 200] = 0.0 + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + + intrinsics = np.load(osp.join(self.ROOT, scene, "cam", f"{img_name}.npz"))[ + "intrinsics" + ] + # camera pose is not provided, placeholder + camera_pose = np.eye(4) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="irs", + label=img_name, + instance=f"{str(idx)}_{img_name}", + is_metric=self.is_metric, + is_video=False, + quantile=np.array(1.0, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=True, + reset=True, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/mapfree.py b/dust3r/datasets/mapfree.py new file mode 100644 index 0000000000000000000000000000000000000000..58eef2f61642deeca4e7accb84429f3d471a5bd9 --- /dev/null +++ b/dust3r/datasets/mapfree.py @@ -0,0 +1,282 @@ +import os.path as osp +import numpy as np +import cv2 +import numpy as np +import itertools +import os +import sys +import pickle +import h5py +from tqdm import tqdm + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class MapFree_Multi(BaseMultiViewDataset): + + def __init__(self, ROOT, *args, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 30 + super().__init__(*args, **kwargs) + + self._load_data() + + def imgid2path(self, img_id, scene): + first_seq_id, first_frame_id = img_id + return os.path.join( + self.ROOT, + scene, + f"dense{first_seq_id}", + "rgb", + f"frame_{first_frame_id:05d}.jpg", + ) + + def path2imgid(self, subscene, filename): + first_seq_id = int(subscene[5:]) + first_frame_id = int(filename[6:-4]) + return [first_seq_id, first_frame_id] + + def _load_data(self): + cache_file = f"{self.ROOT}/cached_metadata_50_col_only.h5" + if os.path.exists(cache_file): + print(f"Loading cached metadata from {cache_file}") + with h5py.File(cache_file, "r") as hf: + self.scenes = list(map(lambda x: x.decode("utf-8"), hf["scenes"][:])) + self.sceneids = hf["sceneids"][:] + self.scope = hf["scope"][:] + self.video_flags = hf["video_flags"][:] + self.groups = hf["groups"][:] + self.id_ranges = hf["id_ranges"][:] + self.images = hf["images"][:] + else: + scene_dirs = sorted( + [ + d + for d in os.listdir(self.ROOT) + if os.path.isdir(os.path.join(self.ROOT, d)) + ] + ) + scenes = [] + sceneids = [] + groups = [] + scope = [] + images = [] + id_ranges = [] + is_video = [] + start = 0 + j = 0 + offset = 0 + + for scene in tqdm(scene_dirs): + scenes.append(scene) + # video sequences + subscenes = sorted( + [ + d + for d in os.listdir(os.path.join(self.ROOT, scene)) + if d.startswith("dense") + ] + ) + id_range_subscenes = [] + for subscene in subscenes: + rgb_paths = sorted( + [ + d + for d in os.listdir( + os.path.join(self.ROOT, scene, subscene, "rgb") + ) + if d.endswith(".jpg") + ] + ) + assert ( + len(rgb_paths) > 0 + ), f"{os.path.join(self.ROOT, scene, subscene)} is empty." + num_imgs = len(rgb_paths) + images.extend( + [self.path2imgid(subscene, rgb_path) for rgb_path in rgb_paths] + ) + id_range_subscenes.append((offset, offset + num_imgs)) + offset += num_imgs + + # image collections + metadata = pickle.load( + open(os.path.join(self.ROOT, scene, "metadata.pkl"), "rb") + ) + ref_imgs = list(metadata.keys()) + img_groups = [] + for ref_img in ref_imgs: + other_imgs = metadata[ref_img] + if len(other_imgs) + 1 < self.num_views: + continue + group = [(*other_img[0], other_img[1]) for other_img in other_imgs] + group.insert(0, (*ref_img, 1)) + img_groups.append(np.array(group)) + id_ranges.append(id_range_subscenes[ref_img[0]]) + scope.append(start) + start = start + len(group) + + num_groups = len(img_groups) + sceneids.extend([j] * num_groups) + groups.extend(img_groups) + is_video.extend([False] * num_groups) + j += 1 + + self.scenes = np.array(scenes) + self.sceneids = np.array(sceneids) + self.scope = np.array(scope) + self.video_flags = np.array(is_video) + self.groups = np.concatenate(groups, 0) + self.id_ranges = np.array(id_ranges) + self.images = np.array(images) + + data = dict( + scenes=self.scenes, + sceneids=self.sceneids, + scope=self.scope, + video_flags=self.video_flags, + groups=self.groups, + id_ranges=self.id_ranges, + images=self.images, + ) + + with h5py.File(cache_file, "w") as h5f: + h5f.create_dataset( + "scenes", + data=data["scenes"].astype(object), + dtype=h5py.string_dtype(encoding="utf-8"), + compression="lzf", + chunks=True, + ) + h5f.create_dataset( + "sceneids", data=data["sceneids"], compression="lzf", chunks=True + ) + h5f.create_dataset( + "scope", data=data["scope"], compression="lzf", chunks=True + ) + h5f.create_dataset( + "video_flags", + data=data["video_flags"], + compression="lzf", + chunks=True, + ) + h5f.create_dataset( + "groups", data=data["groups"], compression="lzf", chunks=True + ) + h5f.create_dataset( + "id_ranges", data=data["id_ranges"], compression="lzf", chunks=True + ) + h5f.create_dataset( + "images", data=data["images"], compression="lzf", chunks=True + ) + + def __len__(self): + return len(self.scope) + + def get_image_num(self): + return len(self.images) + + def get_stats(self): + return f"{len(self)} groups of views" + + def _get_views(self, idx, resolution, rng, num_views): + scene = self.scenes[self.sceneids[idx]] + if rng.random() < 0.6: + ids = np.arange(self.id_ranges[idx][0], self.id_ranges[idx][1]) + cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3) + start_ids = ids[: len(ids) - cut_off + 1] + start_id = rng.choice(start_ids) + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + ids.tolist(), + rng, + max_interval=self.max_interval, + video_prob=0.8, + fix_interval_prob=0.5, + block_shuffle=16, + ) + ids = np.array(ids)[pos] + image_idxs = self.images[ids] + else: + ordered_video = False + seq_start_index = self.scope[idx] + seq_end_index = self.scope[idx + 1] if idx < len(self.scope) - 1 else None + image_idxs = ( + self.groups[seq_start_index:seq_end_index] + if seq_end_index is not None + else self.groups[seq_start_index:] + ) + image_idxs, overlap_scores = image_idxs[:, :2], image_idxs[:, 2] + replace = ( + True + if self.allow_repeat + or len(overlap_scores[overlap_scores > 0]) < num_views + else False + ) + image_idxs = rng.choice( + image_idxs, + num_views, + replace=replace, + p=overlap_scores / np.sum(overlap_scores), + ) + image_idxs = image_idxs.astype(np.int64) + + views = [] + for v, view_idx in enumerate(image_idxs): + img_path = self.imgid2path(view_idx, scene) + depth_path = img_path.replace("rgb", "depth").replace(".jpg", ".npy") + cam_path = img_path.replace("rgb", "cam").replace(".jpg", ".npz") + sky_mask_path = img_path.replace("rgb", "sky_mask") + image = imread_cv2(img_path) + depthmap = np.load(depth_path) + camera_params = np.load(cam_path) + sky_mask = cv2.imread(sky_mask_path, cv2.IMREAD_UNCHANGED) >= 127 + + intrinsics = camera_params["intrinsic"].astype(np.float32) + camera_pose = camera_params["pose"].astype(np.float32) + + depthmap[sky_mask] = -1.0 + depthmap[depthmap > 400.0] = 0.0 + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + threshold = ( + np.percentile(depthmap[depthmap > 0], 98) + if depthmap[depthmap > 0].size > 0 + else 0 + ) + depthmap[depthmap > threshold] = 0.0 + + image, depthmap, intrinsics = self._crop_resize_if_necessary( + image, depthmap, intrinsics, resolution, rng, info=(img_path) + ) + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=image, + depthmap=depthmap, + camera_pose=camera_pose, # cam2world + camera_intrinsics=intrinsics, + dataset="MapFree", + label=img_path, + is_metric=self.is_metric, + instance=img_path, + is_video=ordered_video, + quantile=np.array(0.96, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/megadepth.py b/dust3r/datasets/megadepth.py new file mode 100644 index 0000000000000000000000000000000000000000..321500f9260513f81c009fa8155be0612a5a4ba7 --- /dev/null +++ b/dust3r/datasets/megadepth.py @@ -0,0 +1,98 @@ +import os.path as osp +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class MegaDepth_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + super().__init__(*args, **kwargs) + self._load_data(self.split) + self.is_metric = False + if self.split is None: + pass + elif self.split == "train": + self.select_scene(("0015", "0022"), opposite=True) + elif self.split == "val": + self.select_scene(("0015", "0022")) + else: + raise ValueError(f"bad {self.split=}") + + def _load_data(self, split): + with np.load( + osp.join(self.ROOT, "megadepth_sets_64.npz"), allow_pickle=True + ) as data: + self.all_scenes = data["scenes"] + self.all_images = data["images"] + self.sets = data["sets"] + + def __len__(self): + return len(self.sets) + + def get_image_num(self): + return len(self.all_images) + + def get_stats(self): + return f"{len(self)} groups from {len(self.all_scenes)} scenes" + + def select_scene(self, scene, *instances, opposite=False): + scenes = (scene,) if isinstance(scene, str) else tuple(scene) + scene_id = [s.startswith(scenes) for s in self.all_scenes] + assert any(scene_id), "no scene found" + valid = np.in1d(self.sets[:, 0], np.nonzero(scene_id)[0]) + if instances: + raise NotImplementedError("selecting instances not implemented") + if opposite: + valid = ~valid + assert valid.any() + self.sets = self.sets[valid] + + def _get_views(self, idx, resolution, rng, num_views): + scene_id = self.sets[idx][0] + image_idxs = self.sets[idx][1:65] + replace = False if not self.allow_repeat else True + image_idxs = rng.choice(image_idxs, num_views, replace=replace) + scene, subscene = self.all_scenes[scene_id].split() + seq_path = osp.join(self.ROOT, scene, subscene) + views = [] + for im_id in image_idxs: + img = self.all_images[im_id] + try: + image = imread_cv2(osp.join(seq_path, img + ".jpg")) + depthmap = imread_cv2(osp.join(seq_path, img + ".exr")) + camera_params = np.load(osp.join(seq_path, img + ".npz")) + except Exception as e: + raise OSError(f"cannot load {img}, got exception {e}") + intrinsics = np.float32(camera_params["intrinsics"]) + camera_pose = np.float32(camera_params["cam2world"]) + image, depthmap, intrinsics = self._crop_resize_if_necessary( + image, depthmap, intrinsics, resolution, rng, info=(seq_path, img) + ) + views.append( + dict( + img=image, + depthmap=depthmap, + camera_pose=camera_pose, # cam2world + camera_intrinsics=intrinsics, + dataset="MegaDepth", + label=osp.relpath(seq_path, self.ROOT), + is_metric=self.is_metric, + instance=img, + is_video=False, + quantile=np.array(0.96, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/mp3d.py b/dust3r/datasets/mp3d.py new file mode 100644 index 0000000000000000000000000000000000000000..f88d39e1c56907c4cad9e105ad8ffa505aa362d1 --- /dev/null +++ b/dust3r/datasets/mp3d.py @@ -0,0 +1,132 @@ +import os.path as osp +import os +import sys +import itertools + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class MP3D_Multi(BaseMultiViewDataset): + def __init__(self, *args, split, ROOT, **kwargs): + self.ROOT = ROOT + self.video = False + self.is_metric = True + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data() + + def _load_data(self): + scenes = os.listdir(self.ROOT) + offset = 0 + overlaps = {scene: [] for scene in scenes} + scene_img_list = {scene: [] for scene in scenes} + images = [] + + j = 0 + for scene in scenes: + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")] + ) + overlap = np.load(osp.join(scene_dir, "overlap.npy")) + overlaps[scene] = overlap + num_imgs = len(basenames) + + images.extend( + [(scene, i, basename) for i, basename in enumerate(basenames)] + ) + scene_img_list[scene] = np.arange(num_imgs) + offset + offset += num_imgs + j += 1 + + self.scenes = scenes + self.scene_img_list = scene_img_list + self.images = images + self.overlaps = overlaps + + def __len__(self): + return len(self.images) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + num_views_posible = 0 + num_unique = num_views if not self.allow_repeat else max(num_views // 3, 3) + while num_views_posible < num_unique - 1: + scene, img_idx, _ = self.images[idx] + overlap = self.overlaps[scene] + sel_img_idx = np.where(overlap[:, 0] == img_idx)[0] + overlap_sel = overlap[sel_img_idx] + overlap_sel = overlap_sel[ + (overlap_sel[:, 2] > 0.01) * (overlap_sel[:, 2] < 1) + ] + num_views_posible = len(overlap_sel) + if num_views_posible >= num_unique - 1: + break + idx = rng.choice(len(self.images)) + + ref_id = self.scene_img_list[scene][img_idx] + ids = self.scene_img_list[scene][overlap_sel[:, 1].astype(np.int64)] + replace = False if not self.allow_repeat else True + image_idxs = rng.choice( + ids, + num_views - 1, + replace=replace, + p=overlap_sel[:, 2] / np.sum(overlap_sel[:, 2]), + ) + image_idxs = np.concatenate([[ref_id], image_idxs]) + + ordered_video = False + views = [] + for v, view_idx in enumerate(image_idxs): + scene, _, basename = self.images[view_idx] + scene_dir = osp.join(self.ROOT, scene) + rgb_path = osp.join(scene_dir, "rgb", basename + ".png") + depth_path = osp.join(scene_dir, "depth", basename + ".npy") + cam_path = osp.join(scene_dir, "cam", basename + ".npz") + + rgb_image = imread_cv2(rgb_path, cv2.IMREAD_COLOR) + depthmap = np.load(depth_path).astype(np.float32) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + cam_file = np.load(cam_path) + intrinsics = cam_file["intrinsics"] + camera_pose = cam_file["pose"] + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.85, 0.1, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="mp3d", + label=scene + "_" + rgb_path, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.99, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/mvimgnet.py b/dust3r/datasets/mvimgnet.py new file mode 100644 index 0000000000000000000000000000000000000000..9563f7f5dcd6120b460486b46415ad0e57c214c8 --- /dev/null +++ b/dust3r/datasets/mvimgnet.py @@ -0,0 +1,145 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class MVImgNet_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = False + self.max_interval = 32 + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = os.listdir(self.ROOT) + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".jpg")] + ) + + num_imgs = len(basenames) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + start_img_ids.extend([(scene, id) for id in start_img_ids_]) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + self.invalid_scenes = {scene: False for scene in self.scenes} + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + invalid_seq = True + scene, start_id = self.start_img_ids[idx] + + while invalid_seq: + while self.invalid_scenes[scene]: + idx = rng.integers(low=0, high=len(self.start_img_ids)) + scene, start_id = self.start_img_ids[idx] + + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, start_id, all_image_ids, rng, max_interval=self.max_interval + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for view_idx in image_idxs: + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + try: + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".jpg")) + # Load depthmap, no depth, set to all ones + depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32) + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = np.eye(3) + intrinsics[0, 0] = cam["intrinsics"][0, 0] + intrinsics[1, 1] = cam["intrinsics"][0, 0] + intrinsics[0, 2] = cam["intrinsics"][1, 1] + intrinsics[1, 2] = cam["intrinsics"][0, 2] + except: + print(f"Error loading {scene} {basename}, skipping") + self.invalid_scenes[scene] = True + break + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="MVImgnet", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.98, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=True, + depth_only=False, + single_view=False, + reset=False, + ) + ) + if len(views) == num_views: + invalid_seq = False + return views diff --git a/dust3r/datasets/mvs_synth.py b/dust3r/datasets/mvs_synth.py new file mode 100644 index 0000000000000000000000000000000000000000..5492801a1bfadd28bae329c52c9cfd1da4e9c779 --- /dev/null +++ b/dust3r/datasets/mvs_synth.py @@ -0,0 +1,143 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class MVS_Synth_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = False + self.max_interval = 4 + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = os.listdir(self.ROOT) + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".jpg")] + ) + num_imgs = len(basenames) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=1.0, + fix_interval_prob=1.0, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".jpg")) + # Load depthmap + depthmap = np.load(osp.join(depth_dir, basename + ".npy")) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + threshold = ( + np.percentile(depthmap[depthmap > 0], 98) + if depthmap[depthmap > 0].size > 0 + else 0 + ) + depthmap[depthmap > threshold] = 0.0 + depthmap[depthmap > 1000] = 0.0 + + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.8, 0.15, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="MVS_Synth", + label=self.scenes[scene_id] + "_" + basename, + instance=osp.join(rgb_dir, basename + ".jpg"), + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/omniobject3d.py b/dust3r/datasets/omniobject3d.py new file mode 100644 index 0000000000000000000000000000000000000000..1d8e1019c94e30c70dd1d9dd2d50ff9dee46b924 --- /dev/null +++ b/dust3r/datasets/omniobject3d.py @@ -0,0 +1,146 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys +import json + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 +import re + + +def extract_number(filename): + match = re.search(r"\d+", filename) + if match: + return int(match.group()) + return 0 + + +class OmniObject3D_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = False + self.is_metric = False # True + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = [ + d + for d in os.listdir(self.ROOT) + if os.path.isdir(os.path.join(self.ROOT, d)) and not d.startswith('.') + ] + with open(os.path.join(self.ROOT, "scale.json"), "r") as f: + self.scales = json.load(f) + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")], + key=extract_number, + ) + + num_imgs = len(basenames) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + start_img_ids.extend([(scene, id) for id in start_img_ids_]) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + scene, start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, start_id, all_image_ids, rng, max_interval=100, video_prob=0.0 + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".png")) + depthmap = np.load(osp.join(depth_dir, basename + ".npy")) + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + scale = self.scales[self.scenes[scene_id]] + depthmap = depthmap / scale / 1000.0 + camera_pose[:3, 3] = camera_pose[:3, 3] / scale / 1000.0 + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.8, 0.15, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="OmniObject3D", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/pointodyssey.py b/dust3r/datasets/pointodyssey.py new file mode 100644 index 0000000000000000000000000000000000000000..9ced302f1bdaed09fc2294fd6c3a7dd8e248f964 --- /dev/null +++ b/dust3r/datasets/pointodyssey.py @@ -0,0 +1,178 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class PointOdyssey_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 4 + super().__init__(*args, **kwargs) + assert self.split in ["train", "test", "val"] + self.scenes_to_use = [ + # 'cab_h_bench_3rd', 'cab_h_bench_ego1', 'cab_h_bench_ego2', + "cnb_dlab_0215_3rd", + "cnb_dlab_0215_ego1", + "cnb_dlab_0225_3rd", + "cnb_dlab_0225_ego1", + "dancing", + "dancingroom0_3rd", + "footlab_3rd", + "footlab_ego1", + "footlab_ego2", + "girl", + "girl_egocentric", + "human_egocentric", + "human_in_scene", + "human_in_scene1", + "kg", + "kg_ego1", + "kg_ego2", + "kitchen_gfloor", + "kitchen_gfloor_ego1", + "kitchen_gfloor_ego2", + "scene_carb_h_tables", + "scene_carb_h_tables_ego1", + "scene_carb_h_tables_ego2", + "scene_j716_3rd", + "scene_j716_ego1", + "scene_j716_ego2", + "scene_recording_20210910_S05_S06_0_3rd", + "scene_recording_20210910_S05_S06_0_ego2", + "scene1_0129", + "scene1_0129_ego", + "seminar_h52_3rd", + "seminar_h52_ego1", + "seminar_h52_ego2", + ] + self.loaded_data = self._load_data(self.split) + + def _load_data(self, split): + root = os.path.join(self.ROOT, split) + self.scenes = [] + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(os.listdir(root)): + if scene not in self.scenes_to_use: + continue + scene_dir = osp.join(root, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".jpg")] + ) + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + # start_img_ids_ = img_ids[:-self.num_views+1] + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=1.0, + fix_interval_prob=1.0, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".jpg")) + # Load depthmap + depthmap = np.load(osp.join(depth_dir, basename + ".npy")) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + depthmap[depthmap > 1000] = 0.0 + + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.9, 0.05, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="PointOdyssey", + label=self.scenes[scene_id] + "_" + basename, + instance=osp.join(rgb_dir, basename + ".jpg"), + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/realestate10k.py b/dust3r/datasets/realestate10k.py new file mode 100644 index 0000000000000000000000000000000000000000..34526946529905640be4ee49d0530b950bafdb04 --- /dev/null +++ b/dust3r/datasets/realestate10k.py @@ -0,0 +1,139 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class RE10K_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = False + self.max_interval = 128 + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = os.listdir(self.ROOT) + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")], + key=lambda x: int(x), + ) + + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + start_img_ids.extend([(scene, id) for id in start_img_ids_]) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + self.invalid_scenes = {scene: False for scene in self.scenes} + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + invalid_seq = True + scene, start_id = self.start_img_ids[idx] + + while invalid_seq: + while self.invalid_scenes[scene]: + idx = rng.integers(low=0, high=len(self.start_img_ids)) + scene, start_id = self.start_img_ids[idx] + + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, start_id, all_image_ids, rng, max_interval=self.max_interval + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for view_idx in image_idxs: + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + try: + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".png")) + # Load depthmap, no depth, set to all ones + depthmap = np.ones_like(rgb_image[..., 0], dtype=np.float32) + cam = np.load(osp.join(cam_dir, basename + ".npz")) + intrinsics = cam["intrinsics"] + camera_pose = cam["pose"] + except: + print(f"Error loading {scene} {basename}, skipping") + self.invalid_scenes[scene] = True + break + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="realestate10k", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.98, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=True, + depth_only=False, + single_view=False, + reset=False, + ) + ) + if len(views) == num_views: + invalid_seq = False + return views diff --git a/dust3r/datasets/scannet.py b/dust3r/datasets/scannet.py new file mode 100644 index 0000000000000000000000000000000000000000..a4eb2fd3799a0bda6f1d3de6f0d73dee79b12d82 --- /dev/null +++ b/dust3r/datasets/scannet.py @@ -0,0 +1,148 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class ScanNet_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 30 + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data(self.split) + + def _load_data(self, split): + self.scene_root = osp.join( + self.ROOT, "scans_train" if split == "train" else "scans_test" + ) + self.scenes = [ + scene for scene in os.listdir(self.scene_root) if scene.startswith("scene") + ] + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.scene_root, scene) + with np.load( + osp.join(scene_dir, "new_scene_metadata.npz"), allow_pickle=True + ) as data: + basenames = data["images"] + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + cut_off = ( + self.num_views + if not self.allow_repeat + else max(self.num_views // 3, 3) + ) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=0.6, + fix_interval_prob=0.6, + block_shuffle=16, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.scene_root, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "color") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".jpg")) + # Load depthmap + depthmap = imread_cv2( + osp.join(depth_dir, basename + ".png"), cv2.IMREAD_UNCHANGED + ) + depthmap = depthmap.astype(np.float32) / 1000 + depthmap[~np.isfinite(depthmap)] = 0 # invalid + + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="ScanNet", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.98, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/scannetpp.py b/dust3r/datasets/scannetpp.py new file mode 100644 index 0000000000000000000000000000000000000000..5ef363ef49638f3b4599da865d545d32462f34e4 --- /dev/null +++ b/dust3r/datasets/scannetpp.py @@ -0,0 +1,191 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class ScanNetpp_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 3 + super().__init__(*args, **kwargs) + assert self.split == "train" + self.loaded_data = self._load_data() + + def _load_data(self): + with np.load(osp.join(self.ROOT, "all_metadata.npz")) as data: + self.scenes = data["scenes"] + offset = 0 + scenes = [] + sceneids = [] + images = [] + intrinsics = [] + trajectories = [] + groups = [] + id_ranges = [] + j = 0 + self.image_num = 0 + for scene in self.scenes: + scene_dir = osp.join(self.ROOT, scene) + with np.load( + osp.join(scene_dir, "new_scene_metadata.npz"), allow_pickle=True + ) as data: + imgs = data["images"] + self.image_num += len(imgs) + img_ids = np.arange(len(imgs)).tolist() + intrins = data["intrinsics"] + traj = data["trajectories"] + imgs_on_disk = sorted(os.listdir(osp.join(scene_dir, "images"))) + imgs_on_disk = list(map(lambda x: x[:-4], imgs_on_disk)) + + dslr_ids = [ + i + offset + for i in img_ids + if imgs[i].startswith("DSC") and imgs[i] in imgs_on_disk + ] + iphone_ids = [ + i + offset + for i in img_ids + if imgs[i].startswith("frame") and imgs[i] in imgs_on_disk + ] + + num_imgs = len(imgs) + assert max(dslr_ids) < min(iphone_ids) + assert "image_collection" in data + + img_groups = [] + img_id_ranges = [] + + for ref_id, group in data["image_collection"].item().items(): + if len(group) + 1 < self.num_views: + continue + group.insert(0, (ref_id, 1.0)) + sorted_group = sorted(group, key=lambda x: x[1], reverse=True) + group = [int(x[0] + offset) for x in sorted_group] + img_groups.append(sorted(group)) + + if imgs[ref_id].startswith("frame"): + img_id_ranges.append(dslr_ids) + else: + img_id_ranges.append(iphone_ids) + + if len(img_groups) == 0: + print(f"Skipping {scene}") + continue + scenes.append(scene) + sceneids.extend([j] * num_imgs) + images.extend(imgs) + intrinsics.append(intrins) + trajectories.append(traj) + + # offset groups + groups.extend(img_groups) + id_ranges.extend(img_id_ranges) + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.intrinsics = np.concatenate(intrinsics, axis=0) + self.trajectories = np.concatenate(trajectories, axis=0) + self.id_ranges = id_ranges + self.groups = groups + + def __len__(self): + return len(self.groups) * 10 + + def get_image_num(self): + return self.image_num + + def _get_views(self, idx, resolution, rng, num_views): + idx = idx // 10 + image_idxs = self.groups[idx] + rand_val = rng.random() + + image_idxs_video = self.id_ranges[idx] + cut_off = num_views if not self.allow_repeat else max(num_views // 3, 3) + start_image_idxs = image_idxs_video[: len(image_idxs_video) - cut_off + 1] + + if rand_val < 0.7 and len(start_image_idxs) > 0: + start_id = rng.choice(start_image_idxs) + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + image_idxs_video, + rng, + max_interval=self.max_interval, + video_prob=0.8, + fix_interval_prob=0.5, + block_shuffle=16, + ) + image_idxs = np.array(image_idxs_video)[pos] + + else: + ordered_video = True + # ordered video with varying intervals + num_candidates = len(image_idxs) + max_id = min(num_candidates, int(num_views * (2 + 2 * rng.random()))) + image_idxs = sorted(rng.permutation(image_idxs[:max_id])[:num_views]) + if rand_val > 0.75: + ordered_video = False + image_idxs = rng.permutation(image_idxs) + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + + intrinsics = self.intrinsics[view_idx] + camera_pose = self.trajectories[view_idx] + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(scene_dir, "images", basename + ".jpg")) + # Load depthmap + depthmap = imread_cv2( + osp.join(scene_dir, "depth", basename + ".png"), cv2.IMREAD_UNCHANGED + ) + depthmap = depthmap.astype(np.float32) / 1000 + depthmap[~np.isfinite(depthmap)] = 0 # invalid + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="ScanNet++", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.99, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/smartportraits.py b/dust3r/datasets/smartportraits.py new file mode 100644 index 0000000000000000000000000000000000000000..a5955aecd651f2bf1f6a666b0869b5d97816cf5f --- /dev/null +++ b/dust3r/datasets/smartportraits.py @@ -0,0 +1,85 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class SmartPortraits_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + scenes = os.listdir(self.ROOT) + img_names = [] + for scene in scenes: + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")] + ) + img_names.extend([(scene, basename) for basename in basenames]) + + self.img_names = img_names + + def __len__(self): + return len(self.img_names) + + def get_image_num(self): + return len(self.img_names) + + def _get_views(self, idx, resolution, rng, num_views): + new_seed = rng.integers(0, 2**32) + idx + new_rng = np.random.default_rng(new_seed) + img_names = new_rng.choice(self.img_names, num_views, replace=False) + + views = [] + for v, img_name in enumerate(img_names): + # Load RGB image + scene, img_name = img_name + rgb_image = imread_cv2(osp.join(self.ROOT, scene, "rgb", f"{img_name}.png")) + depthmap = np.load(osp.join(self.ROOT, scene, "depth", f"{img_name}.npy")) + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + + intrinsics = np.load(osp.join(self.ROOT, scene, "cam", f"{img_name}.npz"))[ + "intrinsics" + ] + # camera pose is not provided, placeholder + camera_pose = np.eye(4) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="SmartPortraits", + label=img_name, + instance=osp.join(self.ROOT, scene, "rgb", f"{img_name}.png"), + is_metric=self.is_metric, + is_video=False, + quantile=np.array(0.98, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=True, + reset=True, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/spring.py b/dust3r/datasets/spring.py new file mode 100644 index 0000000000000000000000000000000000000000..39bc760a36f56be0e5020e5adacd6eb913aaca6d --- /dev/null +++ b/dust3r/datasets/spring.py @@ -0,0 +1,137 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class Spring(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 16 + super().__init__(*args, **kwargs) + + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = os.listdir(self.ROOT) + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")] + ) + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + # start_img_ids_ = img_ids[:-self.num_views+1] + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=1.0, + fix_interval_prob=1.0, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".png")) + # Load depthmap + depthmap = np.load(osp.join(depth_dir, basename + ".npy")) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.85, 0.10, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="spring", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/synscapes.py b/dust3r/datasets/synscapes.py new file mode 100644 index 0000000000000000000000000000000000000000..92f4fc8506558ec16f50b71d2feacc07ea2f3a18 --- /dev/null +++ b/dust3r/datasets/synscapes.py @@ -0,0 +1,85 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class SynScapes(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = False + self.is_metric = True + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + rgb_dir = osp.join(self.ROOT, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")], + key=lambda x: int(x), + ) + self.img_names = basenames + + def __len__(self): + return len(self.img_names) + + def get_image_num(self): + return len(self.img_names) + + def _get_views(self, idx, resolution, rng, num_views): + new_seed = rng.integers(0, 2**32) + idx + new_rng = np.random.default_rng(new_seed) + img_names = new_rng.choice(self.img_names, num_views, replace=False) + + views = [] + for v, img_name in enumerate(img_names): + # Load RGB image + rgb_image = imread_cv2(osp.join(self.ROOT, "rgb", f"{img_name}.png")) + depthmap = np.load(osp.join(self.ROOT, "depth", f"{img_name}.npy")) + sky_mask = ( + imread_cv2(osp.join(self.ROOT, "sky_mask", f"{img_name}.png"))[..., 0] + >= 127 + ) + depthmap[sky_mask] = -1.0 + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + depthmap[depthmap > 200] = 0.0 + + intrinsics = np.load(osp.join(self.ROOT, "cam", f"{img_name}.npz"))[ + "intrinsics" + ] + # camera pose is not provided, placeholder + camera_pose = np.eye(4) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="synscapes", + label=img_name, + instance=f"{str(idx)}_{img_name}", + is_metric=self.is_metric, + is_video=False, + quantile=np.array(1.0, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=True, + reset=True, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/tartanair.py b/dust3r/datasets/tartanair.py new file mode 100644 index 0000000000000000000000000000000000000000..760d0e9d6921bb31354fbe505821b550d301f83a --- /dev/null +++ b/dust3r/datasets/tartanair.py @@ -0,0 +1,164 @@ +import os.path as osp +import numpy as np +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class TartanAir_Multi(BaseMultiViewDataset): + + def __init__(self, ROOT, *args, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 20 + super().__init__(*args, **kwargs) + # loading all + assert self.split is None + self._load_data() + + def _load_data(self): + scene_dirs = sorted( + [ + d + for d in os.listdir(self.ROOT) + if os.path.isdir(os.path.join(self.ROOT, d)) + ] + ) + + offset = 0 + scenes = [] + sceneids = [] + images = [] + scene_img_list = [] + start_img_ids = [] + j = 0 + + for scene in scene_dirs: + for mode in ["Easy", "Hard"]: + seq_dirs = sorted( + [ + os.path.join(self.ROOT, scene, mode, d) + for d in os.listdir(os.path.join(self.ROOT, scene, mode)) + if os.path.isdir(os.path.join(self.ROOT, scene, mode, d)) + ] + ) + for seq_dir in seq_dirs: + basenames = sorted( + [f[:-8] for f in os.listdir(seq_dir) if f.endswith(".png")] + ) + num_imgs = len(basenames) + cut_off = ( + self.num_views + if not self.allow_repeat + else max(self.num_views // 3, 3) + ) + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + scenes.append(seq_dir) + scene_img_list.append(img_ids) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + start_img_ids.extend(start_img_ids_) + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def get_stats(self): + return f"{len(self)} groups of views" + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + scene_id = self.sceneids[start_id] + all_image_ids = self.scene_img_list[scene_id] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=0.8, + fix_interval_prob=0.8, + block_shuffle=16, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = self.scenes[scene_id] + basename = self.images[view_idx] + + img = basename + "_rgb.png" + image = imread_cv2(osp.join(scene_dir, img)) + depthmap = np.load(osp.join(scene_dir, basename + "_depth.npy")) + camera_params = np.load(osp.join(scene_dir, basename + "_cam.npz")) + + intrinsics = camera_params["camera_intrinsics"] + camera_pose = camera_params["camera_pose"] + + sky_mask = depthmap >= 1000 + depthmap[sky_mask] = -1.0 # sky + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + threshold = ( + np.percentile(depthmap[depthmap > 0], 98) + if depthmap[depthmap > 0].size > 0 + else 0 + ) + depthmap[depthmap > threshold] = 0.0 + + image, depthmap, intrinsics = self._crop_resize_if_necessary( + image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img) + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=image, + depthmap=depthmap, + camera_pose=camera_pose, # cam2world + camera_intrinsics=intrinsics, + dataset="TartanAir", + label=scene_dir, + is_metric=self.is_metric, + instance=scene_dir + "_" + img, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/threedkb.py b/dust3r/datasets/threedkb.py new file mode 100644 index 0000000000000000000000000000000000000000..face09abd00f76cd62e7654b1b673e9d1d3394b7 --- /dev/null +++ b/dust3r/datasets/threedkb.py @@ -0,0 +1,111 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class ThreeDKenBurns(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = False + self.is_metric = False + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = os.listdir(self.ROOT) + + offset = 0 + scenes = [] + sceneids = [] + images = [] + img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")] + ) + + num_imgs = len(basenames) + img_ids_ = list(np.arange(num_imgs) + offset) + + img_ids.extend(img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.img_ids = img_ids + + def __len__(self): + return len(self.img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + new_seed = rng.integers(0, 2**32) + idx + new_rng = np.random.default_rng(new_seed) + image_idxs = new_rng.choice(self.img_ids, num_views, replace=False) + + views = [] + for view_idx in image_idxs: + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".png")) + depthmap = imread_cv2(osp.join(depth_dir, basename + ".exr")) + depthmap[depthmap > 20000] = 0.0 + depthmap = depthmap / 1000.0 + cam = np.load(osp.join(cam_dir, basename + ".npz")) + intrinsics = cam["intrinsics"] + camera_pose = np.eye(4) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="3DKenBurns", + label=self.scenes[scene_id] + "_" + basename, + instance=f"{str(idx)}_{str(view_idx)}", + is_metric=self.is_metric, + is_video=False, + quantile=np.array(1.0, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=True, + reset=True, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/uasol.py b/dust3r/datasets/uasol.py new file mode 100644 index 0000000000000000000000000000000000000000..b91b43bdd6a27691ac5016b22c183ac300d219a9 --- /dev/null +++ b/dust3r/datasets/uasol.py @@ -0,0 +1,148 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + +import re + + +def extract_number(filename): + match = re.search(r"\d+", filename) + if match: + return int(match.group()) + return 0 + + +class UASOL_Multi(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 40 + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + self.scenes = os.listdir(self.ROOT) + + offset = 0 + scenes = [] + sceneids = [] + scene_img_list = [] + images = [] + start_img_ids = [] + + j = 0 + for scene in tqdm(self.scenes): + scene_dir = osp.join(self.ROOT, scene) + rgb_dir = osp.join(scene_dir, "rgb") + basenames = sorted( + [f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")], + key=extract_number, + ) + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + # start_img_ids_ = img_ids[:-self.num_views+1] + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(scene) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=0.75, + fix_interval_prob=0.75, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + rgb_dir = osp.join(scene_dir, "rgb") + depth_dir = osp.join(scene_dir, "depth") + cam_dir = osp.join(scene_dir, "cam") + + basename = self.images[view_idx] + + # Load RGB image + rgb_image = imread_cv2(osp.join(rgb_dir, basename + ".png")) + # Load depthmap + depthmap = np.load(osp.join(depth_dir, basename + ".npy")) + depthmap[~np.isfinite(depthmap)] = 0 # invalid + depthmap[depthmap >= 20] = 0 # invalid + + cam = np.load(osp.join(cam_dir, basename + ".npz")) + camera_pose = cam["pose"] + intrinsics = cam["intrinsics"] + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="UASOL", + label=self.scenes[scene_id] + "_" + basename, + instance=osp.join(rgb_dir, basename + ".png"), + is_metric=self.is_metric, + is_video=ordered_video, + quantile=np.array(0.9, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/unreal4k.py b/dust3r/datasets/unreal4k.py new file mode 100644 index 0000000000000000000000000000000000000000..4d9092928daacf527c99e1958bbee85ef9110035 --- /dev/null +++ b/dust3r/datasets/unreal4k.py @@ -0,0 +1,159 @@ +import os.path as osp +import numpy as np +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + +R_conv = np.array([[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype( + np.float32 +) + + +class UnReal4K_Multi(BaseMultiViewDataset): + + def __init__(self, ROOT, *args, **kwargs): + self.ROOT = ROOT + self.max_interval = 2 + self.is_metric = True + super().__init__(*args, **kwargs) + # loading all + assert self.split is None + self._load_data() + + def _load_data(self): + scene_dirs = sorted( + [ + d + for d in os.listdir(self.ROOT) + if os.path.isdir(os.path.join(self.ROOT, d)) + ] + ) + + offset = 0 + scenes = [] + sceneids = [] + images = [] + start_img_ids = [] + scene_img_list = [] + j = 0 + + seq_dirs = sorted( + [ + os.path.join(self.ROOT, scene, mode) + for scene in scene_dirs + for mode in ["0", "1"] + ] + ) + for seq_dir in seq_dirs: + basenames = sorted( + [f[:-8] for f in os.listdir(seq_dir) if f.endswith(".png")] + ) + num_imgs = len(basenames) + img_ids = list(np.arange(num_imgs) + offset) + # start_img_ids_ = img_ids[:-self.num_views+1] + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + if num_imgs < cut_off: + print(f"Skipping {seq_dir}") + continue + + start_img_ids.extend(start_img_ids_) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + scenes.append(seq_dir) + scene_img_list.append(img_ids) + + # offset groups + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) * 10 + + def get_image_num(self): + return len(self.images) + + def get_stats(self): + return f"{len(self)//10} groups of views" + + def _get_views(self, idx, resolution, rng, num_views): + idx = idx // 10 + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + pos, ordered_video = self.get_seq_from_start_id( + num_views, start_id, all_image_ids, rng, max_interval=self.max_interval + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = self.scenes[scene_id] + basename = self.images[view_idx] + + img = basename + "_rgb.png" + image = imread_cv2(osp.join(scene_dir, img)) + depthmap = np.load(osp.join(scene_dir, basename + "_depth.npy")) + camera_params = np.load(osp.join(scene_dir, basename + ".npz")) + + intrinsics = camera_params["intrinsics"].astype(np.float32) + camera_pose = camera_params["cam2world"].astype(np.float32) + + camera_pose = R_conv @ camera_pose + + sky_mask = depthmap >= 1000 + depthmap[sky_mask] = -1.0 # sky + threshold = ( + np.percentile(depthmap[depthmap > 0], 98) + if depthmap[depthmap > 0].size > 0 + else 0 + ) + depthmap[depthmap > threshold] = 0.0 + image, depthmap, intrinsics = self._crop_resize_if_necessary( + image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img) + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.75, 0.2, 0.05] + ) + + views.append( + dict( + img=image, + depthmap=depthmap, + camera_pose=camera_pose, # cam2world + camera_intrinsics=intrinsics, + dataset="UnReal4K", + label=scene_dir, + is_metric=self.is_metric, + instance=scene_dir + "_" + img, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/urbansyn.py b/dust3r/datasets/urbansyn.py new file mode 100644 index 0000000000000000000000000000000000000000..f3654a1200fffc1ae1c23483c752e06452f91310 --- /dev/null +++ b/dust3r/datasets/urbansyn.py @@ -0,0 +1,82 @@ +import os.path as osp +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +from tqdm import tqdm +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class UrbanSyn(BaseMultiViewDataset): + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.video = False + self.is_metric = True + super().__init__(*args, **kwargs) + self.loaded_data = self._load_data() + + def _load_data(self): + rgb_dir = osp.join(self.ROOT, "rgb") + basenames = sorted([f[:-4] for f in os.listdir(rgb_dir) if f.endswith(".png")]) + self.img_names = basenames + + def __len__(self): + return len(self.img_names) + + def get_image_num(self): + return len(self.img_names) + + def _get_views(self, idx, resolution, rng, num_views): + new_seed = rng.integers(0, 2**32) + idx + new_rng = np.random.default_rng(new_seed) + img_names = new_rng.choice(self.img_names, num_views, replace=False) + + views = [] + for img_name in img_names: + # Load RGB image + rgb_image = imread_cv2(osp.join(self.ROOT, "rgb", f"{img_name}.png")) + depthmap = np.load(osp.join(self.ROOT, "depth", f"{img_name}.npy")) + sky_mask = ( + imread_cv2(osp.join(self.ROOT, "sky_mask", f"{img_name}.png"))[..., 0] + >= 127 + ) + depthmap[sky_mask] = -1.0 + depthmap = np.nan_to_num(depthmap, nan=0, posinf=0, neginf=0) + depthmap[depthmap > 200] = 0.0 + + intrinsics = np.load(osp.join(self.ROOT, "cam", f"{img_name}.npz"))[ + "intrinsics" + ] + # camera pose is not provided, placeholder + camera_pose = np.eye(4) + + rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary( + rgb_image, depthmap, intrinsics, resolution, rng=rng, info=img_name + ) + + views.append( + dict( + img=rgb_image, + depthmap=depthmap.astype(np.float32), + camera_pose=camera_pose.astype(np.float32), + camera_intrinsics=intrinsics.astype(np.float32), + dataset="urbansyn", + label=img_name, + instance=f"{str(idx)}_{img_name}", + is_metric=self.is_metric, + is_video=False, + quantile=np.array(1.0, dtype=np.float32), + img_mask=True, + ray_mask=False, + camera_only=False, + depth_only=False, + single_view=True, + reset=True, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/utils/__init__.py b/dust3r/datasets/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a32692113d830ddc4af4e6ed608f222fbe062e6e --- /dev/null +++ b/dust3r/datasets/utils/__init__.py @@ -0,0 +1,2 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). diff --git a/dust3r/datasets/utils/corr.py b/dust3r/datasets/utils/corr.py new file mode 100644 index 0000000000000000000000000000000000000000..a0413d4cc035f21acd9b02fb2bccebe36ab57736 --- /dev/null +++ b/dust3r/datasets/utils/corr.py @@ -0,0 +1,129 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import numpy as np +from dust3r.utils.device import to_numpy +from dust3r.utils.geometry import inv, geotrf + + +def reproject_view(pts3d, view2): + shape = view2["pts3d"].shape[:2] + return reproject( + pts3d, view2["camera_intrinsics"], inv(view2["camera_pose"]), shape + ) + + +def reproject(pts3d, K, world2cam, shape): + H, W, THREE = pts3d.shape + assert THREE == 3 + + # reproject in camera2 space + with np.errstate(divide="ignore", invalid="ignore"): + pos = geotrf(K @ world2cam[:3], pts3d, norm=1, ncol=2) + + # quantize to pixel positions + return (H, W), ravel_xy(pos, shape) + + +def ravel_xy(pos, shape): + H, W = shape + with np.errstate(invalid="ignore"): + qx, qy = pos.reshape(-1, 2).round().astype(np.int32).T + quantized_pos = qx.clip(min=0, max=W - 1, out=qx) + W * qy.clip( + min=0, max=H - 1, out=qy + ) + return quantized_pos + + +def unravel_xy(pos, shape): + # convert (x+W*y) back to 2d (x,y) coordinates + return np.unravel_index(pos, shape)[0].base[:, ::-1].copy() + + +def reciprocal_1d(corres_1_to_2, corres_2_to_1, ret_recip=False): + is_reciprocal1 = corres_2_to_1[corres_1_to_2] == np.arange(len(corres_1_to_2)) + pos1 = is_reciprocal1.nonzero()[0] + pos2 = corres_1_to_2[pos1] + if ret_recip: + return is_reciprocal1, pos1, pos2 + return pos1, pos2 + + +def extract_correspondences_from_pts3d( + view1, view2, target_n_corres, rng=np.random, ret_xy=True, nneg=0 +): + view1, view2 = to_numpy((view1, view2)) + # project pixels from image1 --> 3d points --> image2 pixels + shape1, corres1_to_2 = reproject_view(view1["pts3d"], view2) + shape2, corres2_to_1 = reproject_view(view2["pts3d"], view1) + + # compute reciprocal correspondences: + # pos1 == valid pixels (correspondences) in image1 + is_reciprocal1, pos1, pos2 = reciprocal_1d( + corres1_to_2, corres2_to_1, ret_recip=True + ) + is_reciprocal2 = corres1_to_2[corres2_to_1] == np.arange(len(corres2_to_1)) + + if target_n_corres is None: + if ret_xy: + pos1 = unravel_xy(pos1, shape1) + pos2 = unravel_xy(pos2, shape2) + return pos1, pos2 + + available_negatives = min((~is_reciprocal1).sum(), (~is_reciprocal2).sum()) + target_n_positives = int(target_n_corres * (1 - nneg)) + n_positives = min(len(pos1), target_n_positives) + n_negatives = min(target_n_corres - n_positives, available_negatives) + + if n_negatives + n_positives != target_n_corres: + # should be really rare => when there are not enough negatives + # in that case, break nneg and add a few more positives ? + n_positives = target_n_corres - n_negatives + assert n_positives <= len(pos1) + + assert n_positives <= len(pos1) + assert n_positives <= len(pos2) + assert n_negatives <= (~is_reciprocal1).sum() + assert n_negatives <= (~is_reciprocal2).sum() + assert n_positives + n_negatives == target_n_corres + + valid = np.ones(n_positives, dtype=bool) + if n_positives < len(pos1): + # random sub-sampling of valid correspondences + perm = rng.permutation(len(pos1))[:n_positives] + pos1 = pos1[perm] + pos2 = pos2[perm] + + if n_negatives > 0: + # add false correspondences if not enough + def norm(p): + return p / p.sum() + + pos1 = np.r_[ + pos1, + rng.choice( + shape1[0] * shape1[1], + size=n_negatives, + replace=False, + p=norm(~is_reciprocal1), + ), + ] + pos2 = np.r_[ + pos2, + rng.choice( + shape2[0] * shape2[1], + size=n_negatives, + replace=False, + p=norm(~is_reciprocal2), + ), + ] + valid = np.r_[valid, np.zeros(n_negatives, dtype=bool)] + + # convert (x+W*y) back to 2d (x,y) coordinates + if ret_xy: + pos1 = unravel_xy(pos1, shape1) + pos2 = unravel_xy(pos2, shape2) + return pos1, pos2, valid diff --git a/dust3r/datasets/utils/cropping.py b/dust3r/datasets/utils/cropping.py new file mode 100644 index 0000000000000000000000000000000000000000..6074f0d93b54ef5af36189276e0f179825a525fe --- /dev/null +++ b/dust3r/datasets/utils/cropping.py @@ -0,0 +1,147 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# croppping utilities +# -------------------------------------------------------- +import PIL.Image +import os + +os.environ["OPENCV_IO_ENABLE_OPENEXR"] = "1" +import cv2 # noqa +import numpy as np # noqa +from dust3r.utils.geometry import ( + colmap_to_opencv_intrinsics, + opencv_to_colmap_intrinsics, +) # noqa + +try: + lanczos = PIL.Image.Resampling.LANCZOS + bicubic = PIL.Image.Resampling.BICUBIC +except AttributeError: + lanczos = PIL.Image.LANCZOS + bicubic = PIL.Image.BICUBIC + + +class ImageList: + """Convenience class to aply the same operation to a whole set of images.""" + + def __init__(self, images): + if not isinstance(images, (tuple, list, set)): + images = [images] + self.images = [] + for image in images: + if not isinstance(image, PIL.Image.Image): + image = PIL.Image.fromarray(image) + self.images.append(image) + + def __len__(self): + return len(self.images) + + def to_pil(self): + return tuple(self.images) if len(self.images) > 1 else self.images[0] + + @property + def size(self): + sizes = [im.size for im in self.images] + assert all(sizes[0] == s for s in sizes) + return sizes[0] + + def resize(self, *args, **kwargs): + return ImageList(self._dispatch("resize", *args, **kwargs)) + + def crop(self, *args, **kwargs): + return ImageList(self._dispatch("crop", *args, **kwargs)) + + def _dispatch(self, func, *args, **kwargs): + return [getattr(im, func)(*args, **kwargs) for im in self.images] + + +def rescale_image_depthmap( + image, depthmap, camera_intrinsics, output_resolution, force=True +): + """Jointly rescale a (image, depthmap) + so that (out_width, out_height) >= output_res + """ + image = ImageList(image) + input_resolution = np.array(image.size) # (W,H) + output_resolution = np.array(output_resolution) + if depthmap is not None: + # can also use this with masks instead of depthmaps + assert tuple(depthmap.shape[:2]) == image.size[::-1] + + # define output resolution + assert output_resolution.shape == (2,) + scale_final = max(output_resolution / image.size) + 1e-8 + if scale_final >= 1 and not force: # image is already smaller than what is asked + return (image.to_pil(), depthmap, camera_intrinsics) + output_resolution = np.floor(input_resolution * scale_final).astype(int) + + # first rescale the image so that it contains the crop + image = image.resize( + output_resolution, resample=lanczos if scale_final < 1 else bicubic + ) + if depthmap is not None: + depthmap = cv2.resize( + depthmap, + output_resolution, + fx=scale_final, + fy=scale_final, + interpolation=cv2.INTER_NEAREST, + ) + + # no offset here; simple rescaling + camera_intrinsics = camera_matrix_of_crop( + camera_intrinsics, input_resolution, output_resolution, scaling=scale_final + ) + + return image.to_pil(), depthmap, camera_intrinsics + + +def camera_matrix_of_crop( + input_camera_matrix, + input_resolution, + output_resolution, + scaling=1, + offset_factor=0.5, + offset=None, +): + # Margins to offset the origin + margins = np.asarray(input_resolution) * scaling - output_resolution + assert np.all(margins >= 0.0) + if offset is None: + offset = offset_factor * margins + + # Generate new camera parameters + output_camera_matrix_colmap = opencv_to_colmap_intrinsics(input_camera_matrix) + output_camera_matrix_colmap[:2, :] *= scaling + output_camera_matrix_colmap[:2, 2] -= offset + output_camera_matrix = colmap_to_opencv_intrinsics(output_camera_matrix_colmap) + + return output_camera_matrix + + +def crop_image_depthmap(image, depthmap, camera_intrinsics, crop_bbox): + """ + Return a crop of the input view. + """ + image = ImageList(image) + l, t, r, b = crop_bbox + + image = image.crop((l, t, r, b)) + depthmap = depthmap[t:b, l:r] + + camera_intrinsics = camera_intrinsics.copy() + camera_intrinsics[0, 2] -= l + camera_intrinsics[1, 2] -= t + + return image.to_pil(), depthmap, camera_intrinsics + + +def bbox_from_intrinsics_in_out( + input_camera_matrix, output_camera_matrix, output_resolution +): + out_width, out_height = output_resolution + l, t = np.int32(np.round(input_camera_matrix[:2, 2] - output_camera_matrix[:2, 2])) + crop_bbox = (l, t, l + out_width, t + out_height) + return crop_bbox diff --git a/dust3r/datasets/utils/transforms.py b/dust3r/datasets/utils/transforms.py new file mode 100644 index 0000000000000000000000000000000000000000..39a4450e57e3482315e307e72c0f3b19e77dea3b --- /dev/null +++ b/dust3r/datasets/utils/transforms.py @@ -0,0 +1,80 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# DUST3R default transforms +# -------------------------------------------------------- +import torchvision.transforms as tvf +from dust3r.utils.image import ImgNorm + +# define the standard image transforms +ColorJitter = tvf.Compose([tvf.ColorJitter(0.5, 0.5, 0.5, 0.1), ImgNorm]) + + +def _check_input(value, center=1, bound=(0, float("inf")), clip_first_on_zero=True): + if isinstance(value, (int, float)): + if value < 0: + raise ValueError(f"If is a single number, it must be non negative.") + value = [center - float(value), center + float(value)] + if clip_first_on_zero: + value[0] = max(value[0], 0.0) + elif isinstance(value, (tuple, list)) and len(value) == 2: + value = [float(value[0]), float(value[1])] + else: + raise TypeError(f"should be a single number or a list/tuple with length 2.") + + if not bound[0] <= value[0] <= value[1] <= bound[1]: + raise ValueError(f"values should be between {bound}, but got {value}.") + + # if value is 0 or (1., 1.) for brightness/contrast/saturation + # or (0., 0.) for hue, do nothing + if value[0] == value[1] == center: + return None + else: + return tuple(value) + + +import torch +import torchvision.transforms.functional as F + + +def SeqColorJitter(): + """ + Return a color jitter transform with same random parameters + """ + brightness = _check_input(0.5) + contrast = _check_input(0.5) + saturation = _check_input(0.5) + hue = _check_input(0.1, center=0, bound=(-0.5, 0.5), clip_first_on_zero=False) + + fn_idx = torch.randperm(4) + brightness_factor = ( + None + if brightness is None + else float(torch.empty(1).uniform_(brightness[0], brightness[1])) + ) + contrast_factor = ( + None + if contrast is None + else float(torch.empty(1).uniform_(contrast[0], contrast[1])) + ) + saturation_factor = ( + None + if saturation is None + else float(torch.empty(1).uniform_(saturation[0], saturation[1])) + ) + hue_factor = None if hue is None else float(torch.empty(1).uniform_(hue[0], hue[1])) + + def _color_jitter(img): + for fn_id in fn_idx: + if fn_id == 0 and brightness_factor is not None: + img = F.adjust_brightness(img, brightness_factor) + elif fn_id == 1 and contrast_factor is not None: + img = F.adjust_contrast(img, contrast_factor) + elif fn_id == 2 and saturation_factor is not None: + img = F.adjust_saturation(img, saturation_factor) + elif fn_id == 3 and hue_factor is not None: + img = F.adjust_hue(img, hue_factor) + return ImgNorm(img) + + return _color_jitter diff --git a/dust3r/datasets/vkitti2.py b/dust3r/datasets/vkitti2.py new file mode 100644 index 0000000000000000000000000000000000000000..438e24f425fdb610b870c4d7b7f02b66ce8e3246 --- /dev/null +++ b/dust3r/datasets/vkitti2.py @@ -0,0 +1,169 @@ +import os.path as osp +import numpy as np +import cv2 +import numpy as np +import itertools +import os +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) + +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class VirtualKITTI2_Multi(BaseMultiViewDataset): + + def __init__(self, ROOT, *args, **kwargs): + self.ROOT = ROOT + self.video = True + self.is_metric = True + self.max_interval = 5 + super().__init__(*args, **kwargs) + # loading all + self._load_data(self.split) + + def _load_data(self, split=None): + scene_dirs = sorted( + [ + d + for d in os.listdir(self.ROOT) + if os.path.isdir(os.path.join(self.ROOT, d)) + ] + ) + if split == "train": + scene_dirs = scene_dirs[:-1] + elif split == "test": + scene_dirs = scene_dirs[-1:] + seq_dirs = [] + for scene in scene_dirs: + seq_dirs += sorted( + [ + os.path.join(scene, d) + for d in os.listdir(os.path.join(self.ROOT, scene)) + if os.path.isdir(os.path.join(self.ROOT, scene, d)) + ] + ) + offset = 0 + scenes = [] + sceneids = [] + images = [] + scene_img_list = [] + start_img_ids = [] + j = 0 + + for seq_idx, seq in enumerate(seq_dirs): + seq_path = osp.join(self.ROOT, seq) + for cam in ["Camera_0", "Camera_1"]: + basenames = sorted( + [ + f[:5] + for f in os.listdir(seq_path + "/" + cam) + if f.endswith(".jpg") + ] + ) + num_imgs = len(basenames) + cut_off = ( + self.num_views + if not self.allow_repeat + else max(self.num_views // 3, 3) + ) + if num_imgs < cut_off: + print(f"Skipping {scene}") + continue + img_ids = list(np.arange(num_imgs) + offset) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + scenes.append(seq + "/" + cam) + scene_img_list.append(img_ids) + sceneids.extend([j] * num_imgs) + images.extend(basenames) + start_img_ids.extend(start_img_ids_) + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def get_stats(self): + return f"{len(self)} groups of views" + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + scene_id = self.sceneids[start_id] + all_image_ids = self.scene_img_list[scene_id] + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=self.max_interval, + video_prob=1.0, + fix_interval_prob=0.9, + ) + image_idxs = np.array(all_image_ids)[pos] + + views = [] + + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir = osp.join(self.ROOT, self.scenes[scene_id]) + basename = self.images[view_idx] + + img = basename + "_rgb.jpg" + image = imread_cv2(osp.join(scene_dir, img)) + depthmap = ( + cv2.imread( + osp.join(scene_dir, basename + "_depth.png"), + cv2.IMREAD_ANYCOLOR | cv2.IMREAD_ANYDEPTH, + ).astype(np.float32) + / 100.0 + ) + camera_params = np.load(osp.join(scene_dir, basename + "_cam.npz")) + + intrinsics = camera_params["camera_intrinsics"] + camera_pose = camera_params["camera_pose"] + + sky_mask = depthmap >= 655 + depthmap[sky_mask] = -1.0 # sky + + image, depthmap, intrinsics = self._crop_resize_if_necessary( + image, depthmap, intrinsics, resolution, rng, info=(scene_dir, img) + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.85, 0.1, 0.05] + ) + + views.append( + dict( + img=image, + depthmap=depthmap, + camera_pose=camera_pose, # cam2world + camera_intrinsics=intrinsics, + dataset="VirtualKITTI2", + label=scene_dir, + is_metric=self.is_metric, + instance=scene_dir + "_" + img, + is_video=ordered_video, + quantile=np.array(1.0, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + assert len(views) == num_views + return views diff --git a/dust3r/datasets/waymo.py b/dust3r/datasets/waymo.py new file mode 100644 index 0000000000000000000000000000000000000000..b7f811f144c638b931cb99fd246702a0fa2d18e7 --- /dev/null +++ b/dust3r/datasets/waymo.py @@ -0,0 +1,178 @@ +import os.path as osp +import os +import numpy as np +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import h5py +from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset +from dust3r.utils.image import imread_cv2 + + +class Waymo_Multi(BaseMultiViewDataset): + """Dataset of outdoor street scenes, 5 images each time""" + + def __init__(self, *args, ROOT, **kwargs): + self.ROOT = ROOT + self.max_interval = 8 + self.video = True + self.is_metric = True + super().__init__(*args, **kwargs) + assert self.split is None + self._load_data() + + def load_invalid_dict(self, h5_file_path): + invalid_dict = {} + with h5py.File(h5_file_path, "r") as h5f: + for scene in h5f: + data = h5f[scene]["invalid_pairs"][:] + invalid_pairs = set( + tuple(pair.decode("utf-8").split("_")) for pair in data + ) + invalid_dict[scene] = invalid_pairs + return invalid_dict + + def _load_data(self): + invalid_dict = self.load_invalid_dict( + os.path.join(self.ROOT, "invalid_files.h5") + ) + scene_dirs = sorted( + [ + d + for d in os.listdir(self.ROOT) + if os.path.isdir(os.path.join(self.ROOT, d)) + ] + ) + offset = 0 + scenes = [] + sceneids = [] + images = [] + start_img_ids = [] + scene_img_list = [] + is_video = [] + j = 0 + + for scene in scene_dirs: + scene_dir = osp.join(self.ROOT, scene) + invalid_pairs = invalid_dict.get(scene, set()) + seq2frames = {} + for f in os.listdir(scene_dir): + if not f.endswith(".jpg"): + continue + basename = f[:-4] + frame_id = basename.split("_")[0] + seq_id = basename.split("_")[1] + if seq_id == "5": + continue + if (seq_id, frame_id) in invalid_pairs: + continue # Skip invalid files + if seq_id not in seq2frames: + seq2frames[seq_id] = [] + seq2frames[seq_id].append(frame_id) + + for seq_id, frame_ids in seq2frames.items(): + frame_ids = sorted(frame_ids) + num_imgs = len(frame_ids) + img_ids = list(np.arange(num_imgs) + offset) + cut_off = ( + self.num_views + if not self.allow_repeat + else max(self.num_views // 3, 3) + ) + start_img_ids_ = img_ids[: num_imgs - cut_off + 1] + + if num_imgs < cut_off: + print(f"Skipping {scene}_{seq_id}") + continue + + scenes.append((scene, seq_id)) + sceneids.extend([j] * num_imgs) + images.extend(frame_ids) + start_img_ids.extend(start_img_ids_) + scene_img_list.append(img_ids) + + offset += num_imgs + j += 1 + + self.scenes = scenes + self.sceneids = sceneids + self.images = images + self.start_img_ids = start_img_ids + self.scene_img_list = scene_img_list + self.is_video = is_video + + def __len__(self): + return len(self.start_img_ids) + + def get_image_num(self): + return len(self.images) + + def get_stats(self): + return f"{len(self)} groups of views" + + def _get_views(self, idx, resolution, rng, num_views): + start_id = self.start_img_ids[idx] + all_image_ids = self.scene_img_list[self.sceneids[start_id]] + _, seq_id = self.scenes[self.sceneids[start_id]] + max_interval = self.max_interval // 2 if seq_id == "4" else self.max_interval + pos, ordered_video = self.get_seq_from_start_id( + num_views, + start_id, + all_image_ids, + rng, + max_interval=max_interval, + video_prob=0.9, + fix_interval_prob=0.9, + block_shuffle=16, + ) + image_idxs = np.array(all_image_ids)[pos] + views = [] + ordered_video = True + + views = [] + + for v, view_idx in enumerate(image_idxs): + scene_id = self.sceneids[view_idx] + scene_dir, seq_id = self.scenes[scene_id] + scene_dir = osp.join(self.ROOT, scene_dir) + frame_id = self.images[view_idx] + + impath = f"{frame_id}_{seq_id}" + image = imread_cv2(osp.join(scene_dir, impath + ".jpg")) + depthmap = imread_cv2(osp.join(scene_dir, impath + ".exr")) + camera_params = np.load(osp.join(scene_dir, impath + ".npz")) + + intrinsics = np.float32(camera_params["intrinsics"]) + camera_pose = np.float32(camera_params["cam2world"]) + + image, depthmap, intrinsics = self._crop_resize_if_necessary( + image, depthmap, intrinsics, resolution, rng, info=(scene_dir, impath) + ) + + # generate img mask and raymap mask + img_mask, ray_mask = self.get_img_and_ray_masks( + self.is_metric, v, rng, p=[0.85, 0.10, 0.05] + ) + + views.append( + dict( + img=image, + depthmap=depthmap, + camera_pose=camera_pose, # cam2world + camera_intrinsics=intrinsics, + dataset="Waymo", + label=osp.relpath(scene_dir, self.ROOT), + is_metric=self.is_metric, + instance=osp.join(scene_dir, impath + ".jpg"), + is_video=ordered_video, + quantile=np.array(0.98, dtype=np.float32), + img_mask=img_mask, + ray_mask=ray_mask, + camera_only=False, + depth_only=False, + single_view=False, + reset=False, + ) + ) + + return views diff --git a/dust3r/datasets/wildrgbd.py b/dust3r/datasets/wildrgbd.py new file mode 100644 index 0000000000000000000000000000000000000000..9ba152e19b9dae9e3ddd254d632f19d779ccffbe --- /dev/null +++ b/dust3r/datasets/wildrgbd.py @@ -0,0 +1,56 @@ +import os.path as osp +import sys + +sys.path.append(osp.join(osp.dirname(__file__), "..", "..")) +import cv2 +import numpy as np + +from dust3r.datasets.co3d import Co3d_Multi +from dust3r.utils.image import imread_cv2 + + +class WildRGBD_Multi(Co3d_Multi): + def __init__(self, mask_bg="rand", *args, ROOT, **kwargs): + super().__init__(mask_bg, *args, ROOT=ROOT, **kwargs) + self.dataset_label = "WildRGBD" + self.is_metric = True + # load all scenes + self.scenes.pop(("box", "scenes/scene_257"), None) + self.scene_list = list(self.scenes.keys()) + cut_off = ( + self.num_views if not self.allow_repeat else max(self.num_views // 3, 3) + ) + self.cut_off = cut_off + self.all_ref_imgs = [ + (key, value) + for key, values in self.scenes.items() + for value in values[: len(values) - cut_off + 1] + ] + self.invalidate = {scene: {} for scene in self.scene_list} + self.invalid_scenes = {scene: False for scene in self.scene_list} + + def _get_metadatapath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "metadata", f"{view_idx:0>5d}.npz") + + def _get_impath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "rgb", f"{view_idx:0>5d}.jpg") + + def _get_depthpath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "depth", f"{view_idx:0>5d}.png") + + def _get_maskpath(self, obj, instance, view_idx): + return osp.join(self.ROOT, obj, instance, "masks", f"{view_idx:0>5d}.png") + + def _read_depthmap(self, depthpath, input_metadata): + # We store depths in the depth scale of 1000. + # That is, when we load depth image and divide by 1000, we could get depth in meters. + depthmap = imread_cv2(depthpath, cv2.IMREAD_UNCHANGED) + depthmap = depthmap.astype(np.float32) / 1000.0 + return depthmap + + def _get_views(self, idx, resolution, rng, num_views): + views = super()._get_views(idx, resolution, rng, num_views) + for view in views: + assert view["is_metric"] + view["quantile"] = np.array(0.96, dtype=np.float32) + return views diff --git a/dust3r/heads/__init__.py b/dust3r/heads/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..75cfc19c494f4c9faa0c9235864541902c75f4f6 --- /dev/null +++ b/dust3r/heads/__init__.py @@ -0,0 +1,35 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +from .linear_head import LinearPts3d, LinearPts3d_Desc, LinearPts3dPose +from .dpt_head import DPTPts3dPose + + +def head_factory( + head_type, + output_mode, + net, + has_conf=False, + has_depth=False, + has_rgb=False, + has_pose_conf=False, + has_pose=False, +): + """ " build a prediction head for the decoder""" + if head_type == "linear" and output_mode == "pts3d": + return LinearPts3d(net, has_conf, has_depth, has_rgb, has_pose_conf) + elif head_type == "linear" and output_mode == "pts3d+pose": + return LinearPts3dPose(net, has_conf, has_rgb, has_pose) + elif head_type == "linear" and output_mode.startswith("pts3d+desc"): + local_feat_dim = int(output_mode[10:]) + return LinearPts3d_Desc(net, has_conf, has_depth, local_feat_dim) + elif head_type == "dpt" and output_mode == "pts3d": + raise NotImplementedError(f"unexpected {head_type=} and {output_mode=}") + return create_dpt_head(net, has_conf=has_conf) + elif head_type == "dpt" and output_mode == "pts3d+pose": + return DPTPts3dPose(net, has_conf, has_rgb, has_pose) + else: + raise NotImplementedError(f"unexpected {head_type=} and {output_mode=}") diff --git a/dust3r/heads/dpt_head.py b/dust3r/heads/dpt_head.py new file mode 100644 index 0000000000000000000000000000000000000000..11bd08ed69679c09770a728d98afb6a1b1bc1cf3 --- /dev/null +++ b/dust3r/heads/dpt_head.py @@ -0,0 +1,260 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +from einops import rearrange +from typing import List +import torch +import torch.nn as nn +from dust3r.heads.postprocess import ( + postprocess, + postprocess_desc, + postprocess_rgb, + postprocess_pose_conf, + postprocess_pose, + reg_dense_conf, +) +import dust3r.utils.path_to_croco # noqa: F401 +from models.dpt_block import DPTOutputAdapter # noqa +from dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder +from dust3r.blocks import ConditionModulationBlock +from torch.utils.checkpoint import checkpoint + + +class DPTOutputAdapter_fix(DPTOutputAdapter): + """ + Adapt croco's DPTOutputAdapter implementation for dust3r: + remove duplicated weigths, and fix forward for dust3r + """ + + def init(self, dim_tokens_enc=768): + super().init(dim_tokens_enc) + + del self.act_1_postprocess + del self.act_2_postprocess + del self.act_3_postprocess + del self.act_4_postprocess + + def forward(self, encoder_tokens: List[torch.Tensor], image_size=None): + assert ( + self.dim_tokens_enc is not None + ), "Need to call init(dim_tokens_enc) function first" + + image_size = self.image_size if image_size is None else image_size + H, W = image_size + + N_H = H // (self.stride_level * self.P_H) + N_W = W // (self.stride_level * self.P_W) + + layers = [encoder_tokens[hook] for hook in self.hooks] + + layers = [self.adapt_tokens(l) for l in layers] + + layers = [ + rearrange(l, "b (nh nw) c -> b c nh nw", nh=N_H, nw=N_W) for l in layers + ] + + layers = [self.act_postprocess[idx](l) for idx, l in enumerate(layers)] + + layers = [self.scratch.layer_rn[idx](l) for idx, l in enumerate(layers)] + + path_4 = self.scratch.refinenet4(layers[3])[ + :, :, : layers[2].shape[2], : layers[2].shape[3] + ] + path_3 = self.scratch.refinenet3(path_4, layers[2]) + path_2 = self.scratch.refinenet2(path_3, layers[1]) + path_1 = self.scratch.refinenet1(path_2, layers[0]) + + out = self.head(path_1) + + return out + + +class PixelwiseTaskWithDPT(nn.Module): + """DPT module for dust3r, can return 3D points + confidence for all pixels""" + + def __init__( + self, + *, + n_cls_token=0, + hooks_idx=None, + dim_tokens=None, + output_width_ratio=1, + num_channels=1, + postprocess=None, + depth_mode=None, + conf_mode=None, + **kwargs + ): + super(PixelwiseTaskWithDPT, self).__init__() + self.return_all_layers = True # backbone needs to return all layers + self.postprocess = postprocess + self.depth_mode = depth_mode + self.conf_mode = conf_mode + + assert n_cls_token == 0, "Not implemented" + dpt_args = dict( + output_width_ratio=output_width_ratio, num_channels=num_channels, **kwargs + ) + if hooks_idx is not None: + dpt_args.update(hooks=hooks_idx) + self.dpt = DPTOutputAdapter_fix(**dpt_args) + dpt_init_args = {} if dim_tokens is None else {"dim_tokens_enc": dim_tokens} + self.dpt.init(**dpt_init_args) + + def forward(self, x, img_info): + out = self.dpt(x, image_size=(img_info[0], img_info[1])) + if self.postprocess: + out = self.postprocess(out, self.depth_mode, self.conf_mode) + return out + + +def create_dpt_head(net, has_conf=False): + """ + return PixelwiseTaskWithDPT for given net params + """ + assert net.dec_depth > 9 + l2 = net.dec_depth + feature_dim = 256 + last_dim = feature_dim // 2 + out_nchan = 3 + ed = net.enc_embed_dim + dd = net.dec_embed_dim + return PixelwiseTaskWithDPT( + num_channels=out_nchan + has_conf, + feature_dim=feature_dim, + last_dim=last_dim, + hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2], + dim_tokens=[ed, dd, dd, dd], + postprocess=postprocess, + depth_mode=net.depth_mode, + conf_mode=net.conf_mode, + head_type="regression", + ) + + +class DPTPts3dPose(nn.Module): + def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False): + super(DPTPts3dPose, self).__init__() + self.return_all_layers = True # backbone needs to return all layers + self.depth_mode = net.depth_mode + self.conf_mode = net.conf_mode + self.pose_mode = net.pose_mode + + self.has_conf = has_conf + self.has_rgb = has_rgb + self.has_pose = has_pose + + pts_channels = 3 + has_conf + rgb_channels = has_rgb * 3 + feature_dim = 256 + last_dim = feature_dim // 2 + ed = net.enc_embed_dim + dd = net.dec_embed_dim + hooks_idx = [0, 1, 2, 3] + dim_tokens = [ed, dd, dd, dd] + head_type = "regression" + output_width_ratio = 1 + + pts_dpt_args = dict( + output_width_ratio=output_width_ratio, + num_channels=pts_channels, + feature_dim=feature_dim, + last_dim=last_dim, + dim_tokens=dim_tokens, + hooks_idx=hooks_idx, + head_type=head_type, + ) + rgb_dpt_args = dict( + output_width_ratio=output_width_ratio, + num_channels=rgb_channels, + feature_dim=feature_dim, + last_dim=last_dim, + dim_tokens=dim_tokens, + hooks_idx=hooks_idx, + head_type=head_type, + ) + if hooks_idx is not None: + pts_dpt_args.update(hooks=hooks_idx) + rgb_dpt_args.update(hooks=hooks_idx) + + self.dpt_self = DPTOutputAdapter_fix(**pts_dpt_args) + dpt_init_args = {} if dim_tokens is None else {"dim_tokens_enc": dim_tokens} + self.dpt_self.init(**dpt_init_args) + + self.final_transform = nn.ModuleList( + [ + ConditionModulationBlock( + net.dec_embed_dim, + net.dec_num_heads, + mlp_ratio=4.0, + qkv_bias=True, + rope=net.rope, + ) + for _ in range(2) + ] + ) + + self.dpt_cross = DPTOutputAdapter_fix(**pts_dpt_args) + dpt_init_args = {} if dim_tokens is None else {"dim_tokens_enc": dim_tokens} + self.dpt_cross.init(**dpt_init_args) + + if has_rgb: + self.dpt_rgb = DPTOutputAdapter_fix(**rgb_dpt_args) + dpt_init_args = {} if dim_tokens is None else {"dim_tokens_enc": dim_tokens} + self.dpt_rgb.init(**dpt_init_args) + + if has_pose: + in_dim = net.dec_embed_dim + self.pose_head = PoseDecoder(hidden_size=in_dim) + + def forward(self, x, img_info, **kwargs): + if self.has_pose: + pose_token = x[-1][:, 0].clone() + token = x[-1][:, 1:] + with torch.cuda.amp.autocast(enabled=False): + pose = self.pose_head(pose_token) + + token_cross = token.clone() + for blk in self.final_transform: + token_cross = blk(token_cross, pose_token, kwargs.get("pos")) + x = x[:-1] + [token] + x_cross = x[:-1] + [token_cross] + + with torch.cuda.amp.autocast(enabled=False): + self_out = checkpoint( + self.dpt_self, + x, + image_size=(img_info[0], img_info[1]), + use_reentrant=False, + ) + + final_output = postprocess(self_out, self.depth_mode, self.conf_mode) + final_output["pts3d_in_self_view"] = final_output.pop("pts3d") + final_output["conf_self"] = final_output.pop("conf") + + if self.has_rgb: + rgb_out = checkpoint( + self.dpt_rgb, + x, + image_size=(img_info[0], img_info[1]), + use_reentrant=False, + ) + rgb_output = postprocess_rgb(rgb_out) + final_output.update(rgb_output) + + if self.has_pose: + pose = postprocess_pose(pose, self.pose_mode) + final_output["camera_pose"] = pose # B,7 + cross_out = checkpoint( + self.dpt_cross, + x_cross, + image_size=(img_info[0], img_info[1]), + use_reentrant=False, + ) + tmp = postprocess(cross_out, self.depth_mode, self.conf_mode) + final_output["pts3d_in_other_view"] = tmp.pop("pts3d") + final_output["conf"] = tmp.pop("conf") + return final_output diff --git a/dust3r/heads/linear_head.py b/dust3r/heads/linear_head.py new file mode 100644 index 0000000000000000000000000000000000000000..081cf21de8252c9ed51882cedf2ecae0c8364985 --- /dev/null +++ b/dust3r/heads/linear_head.py @@ -0,0 +1,346 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import torch +import torch.nn as nn +import torch.nn.functional as F +from dust3r.heads.postprocess import ( + postprocess, + postprocess_desc, + postprocess_rgb, + postprocess_pose_conf, + postprocess_pose, + reg_dense_conf, +) +import dust3r.utils.path_to_croco # noqa +from models.blocks import Mlp # noqa +from dust3r.utils.geometry import geotrf +from dust3r.utils.camera import pose_encoding_to_camera, PoseDecoder +from dust3r.blocks import ConditionModulationBlock + + +class LinearPts3d(nn.Module): + """ + Linear head for dust3r + Each token outputs: - 16x16 3D points (+ confidence) + """ + + def __init__( + self, net, has_conf=False, has_depth=False, has_rgb=False, has_pose_conf=False + ): + super().__init__() + self.patch_size = net.patch_embed.patch_size[0] + self.depth_mode = net.depth_mode + self.conf_mode = net.conf_mode + self.has_conf = has_conf + self.has_rgb = has_rgb + self.has_pose_conf = has_pose_conf + self.has_depth = has_depth + self.proj = Mlp( + net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2 + ) + if has_depth: + self.self_proj = Mlp( + net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2 + ) + if has_rgb: + self.rgb_proj = Mlp(net.dec_embed_dim, out_features=3 * self.patch_size**2) + + def setup(self, croconet): + pass + + def forward(self, decout, img_shape): + H, W = img_shape + tokens = decout[-1] + B, S, D = tokens.shape + + feat = self.proj(tokens) # B,S,D + feat = feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + feat = F.pixel_shuffle(feat, self.patch_size) # B,3,H,W + + final_output = postprocess(feat, self.depth_mode, self.conf_mode) + final_output["pts3d_in_other_view"] = final_output.pop("pts3d") + + if self.has_depth: + self_feat = self.self_proj(tokens) # B,S,D + self_feat = self_feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + self_feat = F.pixel_shuffle(self_feat, self.patch_size) # B,3,H,W + self_3d_output = postprocess(self_feat, self.depth_mode, self.conf_mode) + self_3d_output["pts3d_in_self_view"] = self_3d_output.pop("pts3d") + self_3d_output["conf_self"] = self_3d_output.pop("conf") + final_output.update(self_3d_output) + + if self.has_rgb: + rgb_feat = self.rgb_proj(tokens) + rgb_feat = rgb_feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size) # B,3,H,W + rgb_output = postprocess_rgb(rgb_feat) + final_output.update(rgb_output) + + if self.has_pose_conf: + pose_conf = self.pose_conf_proj(tokens) + pose_conf = pose_conf.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + pose_conf = F.pixel_shuffle(pose_conf, self.patch_size) + pose_conf_output = postprocess_pose_conf(pose_conf) + final_output.update(pose_conf_output) + + return final_output + + +class LinearPts3d_Desc(nn.Module): + """ + Linear head for dust3r + Each token outputs: - 16x16 3D points (+ confidence) + """ + + def __init__( + self, + net, + has_conf=False, + has_depth=False, + local_feat_dim=24, + hidden_dim_factor=4.0, + ): + super().__init__() + self.patch_size = net.patch_embed.patch_size[0] + self.depth_mode = net.depth_mode + self.conf_mode = net.conf_mode + self.has_conf = has_conf + self.double_channel = has_depth + self.local_feat_dim = local_feat_dim + + if not has_depth: + self.proj = nn.Linear( + net.dec_embed_dim, (3 + has_conf) * self.patch_size**2 + ) + else: + self.proj = nn.Linear( + net.dec_embed_dim, (3 + has_conf) * 2 * self.patch_size**2 + ) + idim = net.enc_embed_dim + net.dec_embed_dim + self.head_local_features = Mlp( + in_features=idim, + hidden_features=int(hidden_dim_factor * idim), + out_features=(self.local_feat_dim + 1) * self.patch_size**2, + ) + + def setup(self, croconet): + pass + + def forward(self, decout, img_shape): + H, W = img_shape + tokens = decout[-1] + B, S, D = tokens.shape + + feat = self.proj(tokens) # B,S,D + feat = feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + feat = F.pixel_shuffle(feat, self.patch_size) # B,3,H,W + + enc_output, dec_output = decout[0], decout[-1] + cat_output = torch.cat([enc_output, dec_output], dim=-1) + local_features = self.head_local_features(cat_output) # B,S,D + local_features = local_features.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + local_features = F.pixel_shuffle(local_features, self.patch_size) # B,d,H,W + feat = torch.cat([feat, local_features], dim=1) + + return postprocess_desc( + feat, + self.depth_mode, + self.conf_mode, + self.local_feat_dim, + self.double_channel, + ) + + +class LinearPts3dPoseDirect(nn.Module): + """ + Linear head for dust3r + Each token outputs: - 16x16 3D points (+ confidence) + """ + + def __init__(self, net, has_conf=False, has_rgb=False, has_pose=False): + super().__init__() + self.patch_size = net.patch_embed.patch_size[0] + self.depth_mode = net.depth_mode + self.conf_mode = net.conf_mode + self.pose_mode = net.pose_mode + self.has_conf = has_conf + self.has_rgb = has_rgb + self.has_pose = has_pose + + self.proj = Mlp( + net.dec_embed_dim, out_features=(3 + has_conf) * self.patch_size**2 + ) + if has_rgb: + self.rgb_proj = Mlp(net.dec_embed_dim, out_features=3 * self.patch_size**2) + if has_pose: + self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim) + if has_conf: + self.cross_conf_proj = Mlp( + net.dec_embed_dim, out_features=self.patch_size**2 + ) + + def setup(self, croconet): + pass + + def forward(self, decout, img_shape): + H, W = img_shape + tokens = decout[-1] + if self.has_pose: + pose_token = tokens[:, 0] + tokens = tokens[:, 1:] + B, S, D = tokens.shape + + feat = self.proj(tokens) # B,S,D + feat = feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + feat = F.pixel_shuffle(feat, self.patch_size) # B,3,H,W + final_output = postprocess(feat, self.depth_mode, self.conf_mode) + final_output["pts3d_in_self_view"] = final_output.pop("pts3d") + final_output["conf_self"] = final_output.pop("conf") + + if self.has_rgb: + rgb_feat = self.rgb_proj(tokens) + rgb_feat = rgb_feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size) # B,3,H,W + rgb_output = postprocess_rgb(rgb_feat) + final_output.update(rgb_output) + + if self.has_pose: + pose = self.pose_head(pose_token) + pose = postprocess_pose(pose, self.pose_mode) + final_output["camera_pose"] = pose # B,7 + final_output["pts3d_in_other_view"] = geotrf( + pose_encoding_to_camera(final_output["camera_pose"]), + final_output["pts3d_in_self_view"], + ) + + if self.has_conf: + cross_conf = self.cross_conf_proj(tokens) + cross_conf = cross_conf.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + cross_conf = F.pixel_shuffle(cross_conf, self.patch_size)[:, 0] + final_output["conf"] = reg_dense_conf(cross_conf, mode=self.conf_mode) + return final_output + + +class LinearPts3dPose(nn.Module): + """ + Linear head for dust3r + Each token outputs: - 16x16 3D points (+ confidence) + """ + + def __init__( + self, net, has_conf=False, has_rgb=False, has_pose=False, mlp_ratio=4.0 + ): + super().__init__() + self.patch_size = net.patch_embed.patch_size[0] + self.depth_mode = net.depth_mode + self.conf_mode = net.conf_mode + self.pose_mode = net.pose_mode + self.has_conf = has_conf + self.has_rgb = has_rgb + self.has_pose = has_pose + + self.proj = Mlp( + net.dec_embed_dim, + hidden_features=int(mlp_ratio * net.dec_embed_dim), + out_features=(3 + has_conf) * self.patch_size**2, + ) + if has_rgb: + self.rgb_proj = Mlp( + net.dec_embed_dim, + hidden_features=int(mlp_ratio * net.dec_embed_dim), + out_features=3 * self.patch_size**2, + ) + if has_pose: + self.pose_head = PoseDecoder(hidden_size=net.dec_embed_dim) + self.final_transform = nn.ModuleList( + [ + ConditionModulationBlock( + net.dec_embed_dim, + net.dec_num_heads, + mlp_ratio=4.0, + qkv_bias=True, + rope=net.rope, + ) + for _ in range(2) + ] + ) + self.cross_proj = Mlp( + net.dec_embed_dim, + hidden_features=int(mlp_ratio * net.dec_embed_dim), + out_features=(3 + has_conf) * self.patch_size**2, + ) + + def setup(self, croconet): + pass + + def forward(self, decout, img_shape, **kwargs): + H, W = img_shape + tokens = decout[-1] + if self.has_pose: + pose_token = tokens[:, 0] + tokens = tokens[:, 1:] + with torch.cuda.amp.autocast(enabled=False): + pose = self.pose_head(pose_token) + cross_tokens = tokens + for blk in self.final_transform: + cross_tokens = blk(cross_tokens, pose_token, kwargs.get("pos")) + + with torch.cuda.amp.autocast(enabled=False): + B, S, D = tokens.shape + + feat = self.proj(tokens) # B,S,D + feat = feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + feat = F.pixel_shuffle(feat, self.patch_size) # B,3,H,W + final_output = postprocess( + feat, self.depth_mode, self.conf_mode, pos_z=True + ) + final_output["pts3d_in_self_view"] = final_output.pop("pts3d") + final_output["conf_self"] = final_output.pop("conf") + + if self.has_rgb: + rgb_feat = self.rgb_proj(tokens) + rgb_feat = rgb_feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + rgb_feat = F.pixel_shuffle(rgb_feat, self.patch_size) # B,3,H,W + rgb_output = postprocess_rgb(rgb_feat) + final_output.update(rgb_output) + + if self.has_pose: + pose = postprocess_pose(pose, self.pose_mode) + final_output["camera_pose"] = pose # B,7 + + cross_feat = self.cross_proj(cross_tokens) # B,S,D + cross_feat = cross_feat.transpose(-1, -2).view( + B, -1, H // self.patch_size, W // self.patch_size + ) + cross_feat = F.pixel_shuffle(cross_feat, self.patch_size) # B,3,H,W + tmp = postprocess(cross_feat, self.depth_mode, self.conf_mode) + final_output["pts3d_in_other_view"] = tmp.pop("pts3d") + final_output["conf"] = tmp.pop("conf") + + return final_output diff --git a/dust3r/heads/postprocess.py b/dust3r/heads/postprocess.py new file mode 100644 index 0000000000000000000000000000000000000000..63cf3211b4b2dc5a9782c1d1d53eff17886d54cd --- /dev/null +++ b/dust3r/heads/postprocess.py @@ -0,0 +1,167 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import torch +import torch.nn.functional as F + + +def postprocess(out, depth_mode, conf_mode, pos_z=False): + """ + extract 3D points/confidence from prediction head output + """ + fmap = out.permute(0, 2, 3, 1) # B,H,W,3 + res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode, pos_z=pos_z)) + + if conf_mode is not None: + res["conf"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode) + return res + + +def postprocess_rgb(out, eps=1e-6): + fmap = out.permute(0, 2, 3, 1) # B,H,W,3 + res = torch.sigmoid(fmap) * (1 - 2 * eps) + eps + res = (res - 0.5) * 2 + return dict(rgb=res) + + +def postprocess_pose(out, mode, inverse=False): + """ + extract pose from prediction head output + """ + mode, vmin, vmax = mode + + no_bounds = (vmin == -float("inf")) and (vmax == float("inf")) + assert no_bounds + trans = out[..., 0:3] + quats = out[..., 3:7] + + if mode == "linear": + if no_bounds: + return trans # [-inf, +inf] + return trans.clip(min=vmin, max=vmax) + + d = trans.norm(dim=-1, keepdim=True) + + if mode == "square": + if inverse: + scale = d / d.square().clip(min=1e-8) + else: + scale = d.square() / d.clip(min=1e-8) + + if mode == "exp": + if inverse: + scale = d / torch.expm1(d).clip(min=1e-8) + else: + scale = torch.expm1(d) / d.clip(min=1e-8) + + trans = trans * scale + quats = standardize_quaternion(quats) + + return torch.cat([trans, quats], dim=-1) + + +def postprocess_pose_conf(out): + fmap = out.permute(0, 2, 3, 1) # B,H,W,1 + return dict(pose_conf=torch.sigmoid(fmap)) + + +def postprocess_desc(out, depth_mode, conf_mode, desc_dim, double_channel=False): + """ + extract 3D points/confidence from prediction head output + """ + fmap = out.permute(0, 2, 3, 1) # B,H,W,3 + res = dict(pts3d=reg_dense_depth(fmap[:, :, :, 0:3], mode=depth_mode)) + + if conf_mode is not None: + res["conf"] = reg_dense_conf(fmap[:, :, :, 3], mode=conf_mode) + + if double_channel: + res["pts3d_self"] = reg_dense_depth( + fmap[ + :, :, :, 3 + int(conf_mode is not None) : 6 + int(conf_mode is not None) + ], + mode=depth_mode, + ) + if conf_mode is not None: + res["conf_self"] = reg_dense_conf( + fmap[:, :, :, 6 + int(conf_mode is not None)], mode=conf_mode + ) + + start = ( + 3 + + int(conf_mode is not None) + + int(double_channel) * (3 + int(conf_mode is not None)) + ) + res["desc"] = reg_desc(fmap[:, :, :, start : start + desc_dim], mode="norm") + res["desc_conf"] = reg_dense_conf(fmap[:, :, :, start + desc_dim], mode=conf_mode) + assert start + desc_dim + 1 == fmap.shape[-1] + + return res + + +def reg_desc(desc, mode="norm"): + if "norm" in mode: + desc = desc / desc.norm(dim=-1, keepdim=True) + else: + raise ValueError(f"Unknown desc mode {mode}") + return desc + + +def reg_dense_depth(xyz, mode, pos_z=False): + """ + extract 3D points from prediction head output + """ + mode, vmin, vmax = mode + + no_bounds = (vmin == -float("inf")) and (vmax == float("inf")) + assert no_bounds + + if mode == "linear": + if no_bounds: + return xyz # [-inf, +inf] + return xyz.clip(min=vmin, max=vmax) + + if pos_z: + sign = torch.sign(xyz[..., -1:]) + xyz *= sign + d = xyz.norm(dim=-1, keepdim=True) + xyz = xyz / d.clip(min=1e-8) + + if mode == "square": + return xyz * d.square() + + if mode == "exp": + return xyz * torch.expm1(d) + + raise ValueError(f"bad {mode=}") + + +def reg_dense_conf(x, mode): + """ + extract confidence from prediction head output + """ + mode, vmin, vmax = mode + if mode == "exp": + return vmin + x.exp().clip(max=vmax - vmin) + if mode == "sigmoid": + return (vmax - vmin) * torch.sigmoid(x) + vmin + raise ValueError(f"bad {mode=}") + + +def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor: + """ + Convert a unit quaternion to a standard form: one in which the real + part is non negative. + + Args: + quaternions: Quaternions with real part first, + as tensor of shape (..., 4). + + Returns: + Standardized quaternions as tensor of shape (..., 4). + """ + quaternions = F.normalize(quaternions, p=2, dim=-1) + return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions) diff --git a/dust3r/inference.py b/dust3r/inference.py new file mode 100644 index 0000000000000000000000000000000000000000..bc2b0961d854f6947b5763e63072d897052a017a --- /dev/null +++ b/dust3r/inference.py @@ -0,0 +1,214 @@ +import tqdm +import torch +from dust3r.utils.device import to_cpu, collate_with_cat +from dust3r.utils.misc import invalid_to_nans +from dust3r.utils.geometry import depthmap_to_pts3d, geotrf +from dust3r.model import ARCroco3DStereo +from accelerate import Accelerator +import re +import time + +def sample_query_points(mask, M): + B, H, W = mask.shape + yx = [] + for b in range(B): + ys, xs = torch.where(mask[b]) + if len(xs) == 0 or len(xs) < M: + pts = torch.zeros(M, 2, device=mask.device) + else: + idx = torch.randint(0, len(xs), (M,)) + pts = torch.stack([xs[idx], ys[idx]], dim=-1) + yx.append(pts) + return torch.stack(yx, dim=0) + +def custom_sort_key(key): + text = key.split("/") + if len(text) > 1: + text, num = text[0], text[-1] + return (text, int(num)) + else: + return (key, -1) + + +def merge_chunk_dict(old_dict, curr_dict, add_number): + new_dict = {} + for key, value in curr_dict.items(): + + match = re.search(r"(\d+)$", key) + if match: + + num_part = int(match.group()) + add_number + + new_key = re.sub(r"(\d+)$", str(num_part), key, 1) + new_dict[new_key] = value + else: + new_dict[key] = value + new_dict = old_dict | new_dict + return {k: new_dict[k] for k in sorted(new_dict.keys(), key=custom_sort_key)} + + +def _interleave_imgs(img1, img2): + res = {} + for key, value1 in img1.items(): + value2 = img2[key] + if isinstance(value1, torch.Tensor): + value = torch.stack((value1, value2), dim=1).flatten(0, 1) + else: + value = [x for pair in zip(value1, value2) for x in pair] + res[key] = value + return res + + +def make_batch_symmetric(batch): + view1, view2 = batch + view1, view2 = (_interleave_imgs(view1, view2), _interleave_imgs(view2, view1)) + return view1, view2 + + +def loss_of_one_batch( + batch, + model, + criterion, + accelerator: Accelerator, + teacher=None, + symmetrize_batch=False, + use_amp=False, + ret=None, + img_mask=None, + inference=False, +): + if len(batch) > 2: + assert ( + symmetrize_batch is False + ), "cannot symmetrize batch with more than 2 views" + if symmetrize_batch: + batch = make_batch_symmetric(batch) + if "valid_mask" in batch[0]: + query_pts = sample_query_points(batch[0]['valid_mask'], M=64).to(device=batch[0]["img"].device) + else: + query_pts = None + dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] >= 8 else torch.float16 + with torch.cuda.amp.autocast(dtype=dtype): + if inference: + with torch.no_grad(): + output = model.inference(batch, query_pts) + preds, batch = output.ress, output.views + result = dict(views=batch, pred=preds) + return result[ret] if ret else result + else: + output = model(batch, query_pts) + preds, batch = output.ress, output.views + + if teacher is not None: + with torch.no_grad(): + knowledge = teacher(batch, query_pts) + gts, batch = knowledge.ress, knowledge.views + + with torch.cuda.amp.autocast(enabled=False): + loss = criterion(gts, preds) if criterion is not None else None + else: + with torch.cuda.amp.autocast(enabled=False): + loss = criterion(batch, preds) if criterion is not None else None + + result = dict(views=batch, pred=preds, loss=loss) + return result[ret] if ret else result + + + +def check_if_same_size(pairs): + shapes1 = [img1["img"].shape[-2:] for img1, img2 in pairs] + shapes2 = [img2["img"].shape[-2:] for img1, img2 in pairs] + return all(shapes1[0] == s for s in shapes1) and all( + shapes2[0] == s for s in shapes2 + ) + + +def get_pred_pts3d(gt, pred, use_pose=False, inplace=False): + if "depth" in pred and "pseudo_focal" in pred: + try: + pp = gt["camera_intrinsics"][..., :2, 2] + except KeyError: + pp = None + pts3d = depthmap_to_pts3d(**pred, pp=pp) + + elif "pts3d" in pred: + + pts3d = pred["pts3d"] + + elif "pts3d_in_other_view" in pred: + + assert use_pose is True + return ( + pred["pts3d_in_other_view"] + if inplace + else pred["pts3d_in_other_view"].clone() + ) + + if use_pose: + camera_pose = pred.get("camera_pose") + assert camera_pose is not None + pts3d = geotrf(camera_pose, pts3d) + + return pts3d + + +def find_opt_scaling( + gt_pts1, + gt_pts2, + pr_pts1, + pr_pts2=None, + fit_mode="weiszfeld_stop_grad", + valid1=None, + valid2=None, +): + assert gt_pts1.ndim == pr_pts1.ndim == 4 + assert gt_pts1.shape == pr_pts1.shape + if gt_pts2 is not None: + assert gt_pts2.ndim == pr_pts2.ndim == 4 + assert gt_pts2.shape == pr_pts2.shape + + nan_gt_pts1 = invalid_to_nans(gt_pts1, valid1).flatten(1, 2) + nan_gt_pts2 = ( + invalid_to_nans(gt_pts2, valid2).flatten(1, 2) if gt_pts2 is not None else None + ) + + pr_pts1 = invalid_to_nans(pr_pts1, valid1).flatten(1, 2) + pr_pts2 = ( + invalid_to_nans(pr_pts2, valid2).flatten(1, 2) if pr_pts2 is not None else None + ) + + all_gt = ( + torch.cat((nan_gt_pts1, nan_gt_pts2), dim=1) + if gt_pts2 is not None + else nan_gt_pts1 + ) + all_pr = torch.cat((pr_pts1, pr_pts2), dim=1) if pr_pts2 is not None else pr_pts1 + + dot_gt_pr = (all_pr * all_gt).sum(dim=-1) + dot_gt_gt = all_gt.square().sum(dim=-1) + + if fit_mode.startswith("avg"): + + scaling = dot_gt_pr.nanmean(dim=1) / dot_gt_gt.nanmean(dim=1) + elif fit_mode.startswith("median"): + scaling = (dot_gt_pr / dot_gt_gt).nanmedian(dim=1).values + elif fit_mode.startswith("weiszfeld"): + + scaling = dot_gt_pr.nanmean(dim=1) / dot_gt_gt.nanmean(dim=1) + + for iter in range(10): + + dis = (all_pr - scaling.view(-1, 1, 1) * all_gt).norm(dim=-1) + + w = dis.clip_(min=1e-8).reciprocal() + + scaling = (w * dot_gt_pr).nanmean(dim=1) / (w * dot_gt_gt).nanmean(dim=1) + else: + raise ValueError(f"bad {fit_mode=}") + + if fit_mode.endswith("stop_grad"): + scaling = scaling.detach() + + scaling = scaling.clip(min=1e-3) + + return scaling diff --git a/dust3r/losses.py b/dust3r/losses.py new file mode 100644 index 0000000000000000000000000000000000000000..fdaca4f15858defee4621f766bcde5effd5083dd --- /dev/null +++ b/dust3r/losses.py @@ -0,0 +1,1383 @@ +from copy import copy, deepcopy +import torch +import torch.nn as nn +import torch.nn.functional as F + +from dust3r.inference import get_pred_pts3d, find_opt_scaling +from dust3r.utils.geometry import ( + inv, + geotrf, + normalize_pointcloud, + normalize_pointcloud_group, +) +from dust3r.utils.geometry import ( + get_group_pointcloud_depth, + get_group_pointcloud_center_scale, + weighted_procrustes, +) +from gsplat import rasterization +import numpy as np +import lpips +from dust3r.utils.camera import ( + pose_encoding_to_camera, + camera_to_pose_encoding, + relative_pose_absT_quatR, +) + + + +def Sum(*losses_and_masks): + loss, mask = losses_and_masks[0] + if loss.ndim > 0: + # we are actually returning the loss for every pixels + return losses_and_masks + else: + # we are returning the global loss + for loss2, mask2 in losses_and_masks[1:]: + loss = loss + loss2 + return loss + + +class BaseCriterion(nn.Module): + def __init__(self, reduction="mean"): + super().__init__() + self.reduction = reduction + + +class LLoss(BaseCriterion): + """L-norm loss""" + + def forward(self, a, b): + assert ( + a.shape == b.shape and a.ndim >= 2 and 1 <= a.shape[-1] <= 3 + ), f"Bad shape = {a.shape}" + dist = self.distance(a, b) + if self.reduction == "none": + return dist + if self.reduction == "sum": + return dist.sum() + if self.reduction == "mean": + return dist.mean() if dist.numel() > 0 else dist.new_zeros(()) + raise ValueError(f"bad {self.reduction=} mode") + + def distance(self, a, b): + raise NotImplementedError() + + +class L21Loss(LLoss): + """Euclidean distance between 3d points""" + + def distance(self, a, b): + return torch.norm(a - b, dim=-1) # normalized L2 distance + + +L21 = L21Loss() + + +class MSELoss(LLoss): + def distance(self, a, b): + return (a - b) ** 2 + + +MSE = MSELoss() + + +class Criterion(nn.Module): + def __init__(self, criterion=None): + super().__init__() + assert isinstance( + criterion, BaseCriterion + ), f"{criterion} is not a proper criterion!" + self.criterion = copy(criterion) + + def get_name(self): + return f"{type(self).__name__}({self.criterion})" + + def with_reduction(self, mode="none"): + res = loss = deepcopy(self) + while loss is not None: + assert isinstance(loss, Criterion) + loss.criterion.reduction = mode # make it return the loss for each sample + loss = loss._loss2 # we assume loss is a Multiloss + return res + + +class MultiLoss(nn.Module): + """Easily combinable losses (also keep track of individual loss values): + loss = MyLoss1() + 0.1*MyLoss2() + Usage: + Inherit from this class and override get_name() and compute_loss() + """ + + def __init__(self): + super().__init__() + self._alpha = 1 + self._loss2 = None + + def compute_loss(self, *args, **kwargs): + raise NotImplementedError() + + def get_name(self): + raise NotImplementedError() + + def __mul__(self, alpha): + assert isinstance(alpha, (int, float)) + res = copy(self) + res._alpha = alpha + return res + + __rmul__ = __mul__ # same + + def __add__(self, loss2): + assert isinstance(loss2, MultiLoss) + res = cur = copy(self) + # find the end of the chain + while cur._loss2 is not None: + cur = cur._loss2 + cur._loss2 = loss2 + return res + + def __repr__(self): + name = self.get_name() + if self._alpha != 1: + name = f"{self._alpha:g}*{name}" + if self._loss2: + name = f"{name} + {self._loss2}" + return name + + def forward(self, *args, **kwargs): + loss = self.compute_loss(*args, **kwargs) + if isinstance(loss, tuple): + loss, details = loss + elif loss.ndim == 0: + details = {self.get_name(): float(loss)} + else: + details = {} + loss = loss * self._alpha + + if self._loss2: + loss2, details2 = self._loss2(*args, **kwargs) + loss = loss + loss2 + details |= details2 + + return loss, details + + +class SSIM(nn.Module): + """Layer to compute the SSIM loss between a pair of images""" + + def __init__(self): + super(SSIM, self).__init__() + self.mu_x_pool = nn.AvgPool2d(3, 1) + self.mu_y_pool = nn.AvgPool2d(3, 1) + self.sig_x_pool = nn.AvgPool2d(3, 1) + self.sig_y_pool = nn.AvgPool2d(3, 1) + self.sig_xy_pool = nn.AvgPool2d(3, 1) + + self.refl = nn.ReflectionPad2d(1) + + self.C1 = 0.01**2 + self.C2 = 0.03**2 + + def forward(self, x, y): + x = self.refl(x) + y = self.refl(y) + + mu_x = self.mu_x_pool(x) + mu_y = self.mu_y_pool(y) + + sigma_x = self.sig_x_pool(x**2) - mu_x**2 + sigma_y = self.sig_y_pool(y**2) - mu_y**2 + sigma_xy = self.sig_xy_pool(x * y) - mu_x * mu_y + + SSIM_n = (2 * mu_x * mu_y + self.C1) * (2 * sigma_xy + self.C2) + SSIM_d = (mu_x**2 + mu_y**2 + self.C1) * (sigma_x + sigma_y + self.C2) + + return torch.clamp((1 - SSIM_n / SSIM_d) / 2, 0, 1) + + +class RGBLoss(Criterion, MultiLoss): + def __init__(self, criterion): + super().__init__(criterion) + self.ssim = SSIM() + + def img_loss(self, a, b): + return self.criterion(a, b) + + def compute_loss(self, gts, preds, **kw): + gt_rgbs = [gt["img"].permute(0, 2, 3, 1) for gt in gts] + pred_rgbs = [pred["rgb"] for pred in preds] + ls = [ + self.img_loss(pred_rgb, gt_rgb) + for pred_rgb, gt_rgb in zip(pred_rgbs, gt_rgbs) + ] + details = {} + self_name = type(self).__name__ + for i, l in enumerate(ls): + details[self_name + f"_rgb/{i+1}"] = float(l) + details[f"pred_rgb_{i+1}"] = pred_rgbs[i] + rgb_loss = sum(ls) / len(ls) + return rgb_loss, details + + +class DepthScaleShiftInvLoss(BaseCriterion): + """scale and shift invariant loss""" + + def __init__(self, reduction="none"): + super().__init__(reduction) + + def forward(self, pred, gt, mask): + assert pred.shape == gt.shape and pred.ndim == 3, f"Bad shape = {pred.shape}" + dist = self.distance(pred, gt, mask) + # assert dist.ndim == a.ndim - 1 # one dimension less + if self.reduction == "none": + return dist + if self.reduction == "sum": + return dist.sum() + if self.reduction == "mean": + return dist.mean() if dist.numel() > 0 else dist.new_zeros(()) + raise ValueError(f"bad {self.reduction=} mode") + + def normalize(self, x, mask): + x_valid = x[mask] + splits = mask.sum(dim=(1, 2)).tolist() + x_valid_list = torch.split(x_valid, splits) + shift = [x.mean() for x in x_valid_list] + x_valid_centered = [x - m for x, m in zip(x_valid_list, shift)] + scale = [x.abs().mean() for x in x_valid_centered] + scale = torch.stack(scale) + shift = torch.stack(shift) + x = (x - shift.view(-1, 1, 1)) / scale.view(-1, 1, 1).clamp(min=1e-6) + return x + + def distance(self, pred, gt, mask): + pred = self.normalize(pred, mask) + gt = self.normalize(gt, mask) + return torch.abs((pred - gt)[mask]) + + +class ScaleInvLoss(BaseCriterion): + """scale invariant loss""" + + def __init__(self, reduction="none"): + super().__init__(reduction) + + def forward(self, pred, gt, mask): + assert pred.shape == gt.shape and pred.ndim == 4, f"Bad shape = {pred.shape}" + dist = self.distance(pred, gt, mask) + # assert dist.ndim == a.ndim - 1 # one dimension less + if self.reduction == "none": + return dist + if self.reduction == "sum": + return dist.sum() + if self.reduction == "mean": + return dist.mean() if dist.numel() > 0 else dist.new_zeros(()) + raise ValueError(f"bad {self.reduction=} mode") + + def distance(self, pred, gt, mask): + pred_norm_factor = (torch.norm(pred, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum( + dim=(1, 2) + ).clamp(min=1e-6) + gt_norm_factor = (torch.norm(gt, dim=-1) * mask).sum(dim=(1, 2)) / mask.sum( + dim=(1, 2) + ).clamp(min=1e-6) + pred = pred / pred_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6) + gt = gt / gt_norm_factor.view(-1, 1, 1, 1).clamp(min=1e-6) + return torch.norm(pred - gt, dim=-1)[mask] + + +class Regr3DPose(Criterion, MultiLoss): + """Ensure that all 3D points are correct. + Asymmetric loss: view1 is supposed to be the anchor. + + P1 = RT1 @ D1 + P2 = RT2 @ D2 + loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1) + loss2 = (RT21 @ pred_D2) - (RT1^-1 @ P2) + = (RT21 @ pred_D2) - (RT1^-1 @ RT2 @ D2) + """ + + def __init__( + self, + criterion, + norm_mode="?avg_dis", + gt_scale=False, + sky_loss_value=2, + max_metric_scale=False, + ): + super().__init__(criterion) + if norm_mode.startswith("?"): + # do no norm pts from metric scale datasets + self.norm_all = False + self.norm_mode = norm_mode[1:] + else: + self.norm_all = True + self.norm_mode = norm_mode + self.gt_scale = gt_scale + + self.sky_loss_value = sky_loss_value + self.max_metric_scale = max_metric_scale + + def get_norm_factor_point_cloud( + self, pts_cross, valids, conf_cross, norm_self_only=False + ): + pts = [x for x in pts_cross] + valids = [x for x in valids] + confs = [x for x in conf_cross] + norm_factor = normalize_pointcloud_group( + pts, self.norm_mode, valids, confs, ret_factor_only=True + ) + return norm_factor + + def get_norm_factor_poses(self, gt_trans, pr_trans, not_metric_mask): + + if self.norm_mode and not self.gt_scale: + gt_trans = [x[:, None, None, :].clone() for x in gt_trans] + valids = [torch.ones_like(x[..., 0], dtype=torch.bool) for x in gt_trans] + norm_factor_gt = ( + normalize_pointcloud_group( + gt_trans, + self.norm_mode, + valids, + ret_factor_only=True, + ) + .squeeze(-1) + .squeeze(-1) + ) + else: + norm_factor_gt = torch.ones( + len(gt_trans), dtype=gt_trans[0].dtype, device=gt_trans[0].device + ) + + norm_factor_pr = norm_factor_gt.clone() + if self.norm_mode and not_metric_mask.sum() > 0 and not self.gt_scale: + pr_trans_not_metric = [ + x[not_metric_mask][:, None, None, :].clone() for x in pr_trans + ] + valids = [ + torch.ones_like(x[..., 0], dtype=torch.bool) + for x in pr_trans_not_metric + ] + norm_factor_pr_not_metric = ( + normalize_pointcloud_group( + pr_trans_not_metric, + self.norm_mode, + valids, + ret_factor_only=True, + ) + .squeeze(-1) + .squeeze(-1) + ) + norm_factor_pr[not_metric_mask] = norm_factor_pr_not_metric + return norm_factor_gt, norm_factor_pr + + def get_all_pts3d( + self, + gts, + preds, + dist_clip=None, + norm_self_only=False, + norm_pose_separately=False, + eps=1e-3, + camera1=None, + ): + # everything is normalized w.r.t. camera of view1 + in_camera1 = inv(gts[0]["camera_pose"]) if camera1 is None else inv(camera1) + gt_pts_cross = [geotrf(in_camera1, gt["pts3d"]) for gt in gts] + valids = [gt["valid_mask"].clone() for gt in gts] + camera_only = gts[0]["camera_only"] + + if dist_clip is not None: + # points that are too far-away == invalid + dis = [gt_pt.norm(dim=-1) for gt_pt in gt_pts_cross] + valids = [valid & (dis <= dist_clip) for valid, dis in zip(valids, dis)] + + pr_pts_cross = [pred["pts3d_in_other_view"] for pred in preds] + conf_cross = [torch.log(pred["conf"]).detach().clip(eps) for pred in preds] + + # valids = torch.stack(valids, dim=0) # S B H W + # valids = valids.permute(1, 0, 2, 3) # B S H W + # valids_masks = preprocess_mask(valids, mode="pad") # (B, S, H, W) + # + # valids = torch.unbind(valids_masks, dim=1) # [S] (B, H, W) + + if not self.norm_all: + if self.max_metric_scale: + B = valids[0].shape[0] + dist = [ + torch.where(valid, torch.linalg.norm(gt_pt_cross, dim=-1), 0).view( + B, -1 + ) + for valid, gt_pt_cross in zip(valids, gt_pts_cross) + ] + for d in dist: + gts[0]["is_metric"] = gts[0]["is_metric_scale"] & ( + d.max(dim=-1).values < self.max_metric_scale + ) + not_metric_mask = ~gts[0]["is_metric"] + else: + not_metric_mask = torch.ones_like(gts[0]["is_metric"]) + + # normalize 3d points + # compute the scale using only the self view point maps + if self.norm_mode and not self.gt_scale: + norm_factor_gt = self.get_norm_factor_point_cloud( + gt_pts_cross, + valids, + conf_cross, + norm_self_only=norm_self_only, + ) + else: + norm_factor_gt = torch.ones_like( + preds[0]["pts3d_in_other_view"][:, :1, :1, :1] + ) + + norm_factor_pr = norm_factor_gt.clone() + if self.norm_mode and not_metric_mask.sum() > 0 and not self.gt_scale: + norm_factor_pr_not_metric = self.get_norm_factor_point_cloud( + [pr_pt_cross[not_metric_mask] for pr_pt_cross in pr_pts_cross], + [valid[not_metric_mask] for valid in valids], + [conf[not_metric_mask] for conf in conf_cross], + norm_self_only=norm_self_only, + ) + norm_factor_pr[not_metric_mask] = norm_factor_pr_not_metric + + norm_factor_gt = norm_factor_gt.clip(eps) + norm_factor_pr = norm_factor_pr.clip(eps) + + gt_pts_cross = [pts / norm_factor_gt for pts in gt_pts_cross] + pr_pts_cross = [pts / norm_factor_pr for pts in pr_pts_cross] + + # [(Bx3, BX4), (BX3, BX4), ...], 3 for translation, 4 for quaternion + gt_poses = [ + camera_to_pose_encoding(in_camera1 @ gt["camera_pose"]).clone() + for gt in gts + ] + pr_poses = [pred["camera_pose"].clone() for pred in preds] + pose_norm_factor_gt = norm_factor_gt.clone().squeeze(2, 3) + pose_norm_factor_pr = norm_factor_pr.clone().squeeze(2, 3) + + if norm_pose_separately: + gt_trans = [gt[:, :3] for gt in gt_poses] + pr_trans = [pr[:, :3] for pr in pr_poses] + pose_norm_factor_gt, pose_norm_factor_pr = self.get_norm_factor_poses( + gt_trans, pr_trans, not_metric_mask + ) + elif any(camera_only): + gt_trans = [gt[:, :3] for gt in gt_poses] + pr_trans = [pr[:, :3] for pr in pr_poses] + pose_only_norm_factor_gt, pose_only_norm_factor_pr = ( + self.get_norm_factor_poses(gt_trans, pr_trans, not_metric_mask) + ) + pose_norm_factor_gt = torch.where( + camera_only[:, None], pose_only_norm_factor_gt, pose_norm_factor_gt + ) + pose_norm_factor_pr = torch.where( + camera_only[:, None], pose_only_norm_factor_pr, pose_norm_factor_pr + ) + + gt_poses = [ + (gt[:, :3] / pose_norm_factor_gt.clip(eps), gt[:, 3:]) for gt in gt_poses + ] + pr_poses = [ + (pr[:, :3] / pose_norm_factor_pr.clip(eps), pr[:, 3:]) for pr in pr_poses + ] + pose_masks = (pose_norm_factor_gt.squeeze(-1) > eps) & ( + pose_norm_factor_pr.squeeze(-1) > eps + ) + + + skys = [gt["sky_mask"] & ~valid for gt, valid in zip(gts, valids)] + return ( + gt_pts_cross, + pr_pts_cross, + gt_poses, + pr_poses, + valids, + skys, + pose_masks, + {}, + ) + + def get_all_pts3d_with_scale_loss( + self, + gts, + preds, + dist_clip=None, + norm_self_only=False, + norm_pose_separately=False, + eps=1e-3, + ): + # everything is normalized w.r.t. camera of view1 + in_camera1 = inv(gts[0]["camera_pose"]) + gt_pts_self = [geotrf(inv(gt["camera_pose"]), gt["pts3d"]) for gt in gts] + gt_pts_cross = [geotrf(in_camera1, gt["pts3d"]) for gt in gts] + valids = [gt["valid_mask"].clone() for gt in gts] + camera_only = gts[0]["camera_only"] + + if dist_clip is not None: + # points that are too far-away == invalid + dis = [gt_pt.norm(dim=-1) for gt_pt in gt_pts_cross] + valids = [valid & (dis <= dist_clip) for valid, dis in zip(valids, dis)] + + pr_pts_self = [pred["pts3d_in_self_view"] for pred in preds] + pr_pts_cross = [pred["pts3d_in_other_view"] for pred in preds] + conf_self = [torch.log(pred["conf_self"]).detach().clip(eps) for pred in preds] + conf_cross = [torch.log(pred["conf"]).detach().clip(eps) for pred in preds] + + if not self.norm_all: + if self.max_metric_scale: + B = valids[0].shape[0] + dist = [ + torch.where(valid, torch.linalg.norm(gt_pt_cross, dim=-1), 0).view( + B, -1 + ) + for valid, gt_pt_cross in zip(valids, gt_pts_cross) + ] + for d in dist: + gts[0]["is_metric"] = gts[0]["is_metric_scale"] & ( + d.max(dim=-1).values < self.max_metric_scale + ) + not_metric_mask = ~gts[0]["is_metric"] + else: + not_metric_mask = torch.ones_like(gts[0]["is_metric"]) + + # normalize 3d points + # compute the scale using only the self view point maps + if self.norm_mode and not self.gt_scale: + norm_factor_gt = self.get_norm_factor_point_cloud( + gt_pts_self[:1], + gt_pts_cross[:1], + valids[:1], + conf_self[:1], + conf_cross[:1], + norm_self_only=norm_self_only, + ) + else: + norm_factor_gt = torch.ones_like( + preds[0]["pts3d_in_other_view"][:, :1, :1, :1] + ) + + if self.norm_mode: + norm_factor_pr = self.get_norm_factor_point_cloud( + pr_pts_self[:1], + pr_pts_cross[:1], + valids[:1], + conf_self[:1], + conf_cross[:1], + norm_self_only=norm_self_only, + ) + else: + raise NotImplementedError + # only add loss to metric scale norm factor + if (~not_metric_mask).sum() > 0: + pts_scale_loss = torch.abs( + norm_factor_pr[~not_metric_mask] - norm_factor_gt[~not_metric_mask] + ).mean() + else: + pts_scale_loss = 0.0 + + norm_factor_gt = norm_factor_gt.clip(eps) + norm_factor_pr = norm_factor_pr.clip(eps) + + gt_pts_self = [pts / norm_factor_gt for pts in gt_pts_self] + gt_pts_cross = [pts / norm_factor_gt for pts in gt_pts_cross] + pr_pts_self = [pts / norm_factor_pr for pts in pr_pts_self] + pr_pts_cross = [pts / norm_factor_pr for pts in pr_pts_cross] + + # [(Bx3, BX4), (BX3, BX4), ...], 3 for translation, 4 for quaternion + gt_poses = [ + camera_to_pose_encoding(in_camera1 @ gt["camera_pose"]).clone() + for gt in gts + ] + pr_poses = [pred["camera_pose"].clone() for pred in preds] + pose_norm_factor_gt = norm_factor_gt.clone().squeeze(2, 3) + pose_norm_factor_pr = norm_factor_pr.clone().squeeze(2, 3) + + if norm_pose_separately: + gt_trans = [gt[:, :3] for gt in gt_poses][:1] + pr_trans = [pr[:, :3] for pr in pr_poses][:1] + pose_norm_factor_gt, pose_norm_factor_pr = self.get_norm_factor_poses( + gt_trans, pr_trans, torch.ones_like(not_metric_mask) + ) + elif any(camera_only): + gt_trans = [gt[:, :3] for gt in gt_poses][:1] + pr_trans = [pr[:, :3] for pr in pr_poses][:1] + pose_only_norm_factor_gt, pose_only_norm_factor_pr = ( + self.get_norm_factor_poses( + gt_trans, pr_trans, torch.ones_like(not_metric_mask) + ) + ) + pose_norm_factor_gt = torch.where( + camera_only[:, None], pose_only_norm_factor_gt, pose_norm_factor_gt + ) + pose_norm_factor_pr = torch.where( + camera_only[:, None], pose_only_norm_factor_pr, pose_norm_factor_pr + ) + # only add loss to metric scale norm factor + if (~not_metric_mask).sum() > 0: + pose_scale_loss = torch.abs( + pose_norm_factor_pr[~not_metric_mask] + - pose_norm_factor_gt[~not_metric_mask] + ).mean() + else: + pose_scale_loss = 0.0 + gt_poses = [ + (gt[:, :3] / pose_norm_factor_gt.clip(eps), gt[:, 3:]) for gt in gt_poses + ] + pr_poses = [ + (pr[:, :3] / pose_norm_factor_pr.clip(eps), pr[:, 3:]) for pr in pr_poses + ] + + pose_masks = (pose_norm_factor_gt.squeeze() > eps) & ( + pose_norm_factor_pr.squeeze() > eps + ) + + if any(camera_only): + # this is equal to a loss for camera intrinsics + gt_pts_self = [ + torch.where( + camera_only[:, None, None, None], + (gt / gt[..., -1:].clip(1e-6)).clip(-2, 2), + gt, + ) + for gt in gt_pts_self + ] + pr_pts_self = [ + torch.where( + camera_only[:, None, None, None], + (pr / pr[..., -1:].clip(1e-6)).clip(-2, 2), + pr, + ) + for pr in pr_pts_self + ] + # # do not add cross view loss when there is only camera supervision + + skys = [gt["sky_mask"] & ~valid for gt, valid in zip(gts, valids)] + return ( + gt_pts_self, + gt_pts_cross, + pr_pts_self, + pr_pts_cross, + gt_poses, + pr_poses, + valids, + skys, + pose_masks, + {"scale_loss": pose_scale_loss + pts_scale_loss}, + ) + + def compute_relative_pose_loss( + self, gt_trans, gt_quats, pr_trans, pr_quats, masks=None + ): + if masks is None: + masks = torch.ones(len(gt_trans), dtype=torch.bool, device=gt_trans.device) + gt_trans_matrix1 = gt_trans[:, :, None, :].repeat(1, 1, gt_trans.shape[1], 1)[ + masks + ] + gt_trans_matrix2 = gt_trans[:, None, :, :].repeat(1, gt_trans.shape[1], 1, 1)[ + masks + ] + gt_quats_matrix1 = gt_quats[:, :, None, :].repeat(1, 1, gt_quats.shape[1], 1)[ + masks + ] + gt_quats_matrix2 = gt_quats[:, None, :, :].repeat(1, gt_quats.shape[1], 1, 1)[ + masks + ] + pr_trans_matrix1 = pr_trans[:, :, None, :].repeat(1, 1, pr_trans.shape[1], 1)[ + masks + ] + pr_trans_matrix2 = pr_trans[:, None, :, :].repeat(1, pr_trans.shape[1], 1, 1)[ + masks + ] + pr_quats_matrix1 = pr_quats[:, :, None, :].repeat(1, 1, pr_quats.shape[1], 1)[ + masks + ] + pr_quats_matrix2 = pr_quats[:, None, :, :].repeat(1, pr_quats.shape[1], 1, 1)[ + masks + ] + + gt_rel_trans, gt_rel_quats = relative_pose_absT_quatR( + gt_trans_matrix1, gt_quats_matrix1, gt_trans_matrix2, gt_quats_matrix2 + ) + pr_rel_trans, pr_rel_quats = relative_pose_absT_quatR( + pr_trans_matrix1, pr_quats_matrix1, pr_trans_matrix2, pr_quats_matrix2 + ) + rel_trans_err = torch.norm(gt_rel_trans - pr_rel_trans, dim=-1) + rel_quats_err = torch.norm(gt_rel_quats - pr_rel_quats, dim=-1) + return rel_trans_err.mean() + rel_quats_err.mean() + + def compute_pose_loss(self, gt_poses, pred_poses, masks=None): + """ + gt_pose: list of (Bx3, Bx4) + pred_pose: list of (Bx3, Bx4) + masks: None, or B + """ + gt_trans = torch.stack([gt[0] for gt in gt_poses], dim=1) # BxNx3 + gt_quats = torch.stack([gt[1] for gt in gt_poses], dim=1) # BXNX3 + pred_trans = torch.stack([pr[0] for pr in pred_poses], dim=1) # BxNx4 + pred_quats = torch.stack([pr[1] for pr in pred_poses], dim=1) # BxNx4 + if masks == None: + pose_loss = ( + torch.norm(pred_trans - gt_trans, dim=-1).mean() + + torch.norm(pred_quats - gt_quats, dim=-1).mean() + ) + else: + if not any(masks): + return torch.tensor(0.0) + pose_loss = ( + torch.norm(pred_trans - gt_trans, dim=-1)[masks].mean() + + torch.norm(pred_quats - gt_quats, dim=-1)[masks].mean() + ) + + return pose_loss + + def compute_loss(self, gts, preds, **kw): + ( + gt_pts_cross, + pred_pts_cross, + gt_poses, + pr_poses, + masks, + skys, + pose_masks, + monitoring, + ) = self.get_all_pts3d(gts, preds, **kw) + + if self.sky_loss_value > 0: + assert ( + self.criterion.reduction == "none" + ), "sky_loss_value should be 0 if no conf loss" + masks = [mask | sky for mask, sky in zip(masks, skys)] + + + # if self.sky_loss_value > 0: + # assert ( + # self.criterion.reduction == "none" + # ), "sky_loss_value should be 0 if no conf loss" + # for i, l in enumerate(ls_self): + # ls_self[i] = torch.where(skys[i][masks[i]], self.sky_loss_value, l) + + self_name = type(self).__name__ + + details = {} + + # cross view loss and details + camera_only = gts[0]["camera_only"] + pred_pts_cross = [pred_pts[~camera_only] for pred_pts in pred_pts_cross] + gt_pts_cross = [gt_pts[~camera_only] for gt_pts in gt_pts_cross] + masks_cross = [mask[~camera_only] for mask in masks] + skys_cross = [sky[~camera_only] for sky in skys] + + if "Quantile" in self.criterion.__class__.__name__: + # quantile masks have already been determined by self view losses, here pass in None as quantile + ls_cross, _ = self.criterion( + pred_pts_cross, gt_pts_cross, masks_cross, None + ) + else: + ls_cross = [ + self.criterion(pred_pt[mask], gt_pt[mask]) + for pred_pt, gt_pt, mask in zip( + pred_pts_cross, gt_pts_cross, masks_cross + ) + ] + + for i in range(len(ls_cross)): + details[f"gt_img{i + 1}"] = gts[i]["img"].permute(0, 2, 3, 1).detach() + details[f"valid_mask_{i + 1}"] = masks[i].detach() + + if "img_mask" in gts[i] and "ray_mask" in gts[i]: + details[f"img_mask_{i + 1}"] = gts[i]["img_mask"].detach() + details[f"ray_mask_{i + 1}"] = gts[i]["ray_mask"].detach() + + if "desc" in preds[i]: + details[f"desc_{i + 1}"] = preds[i]["desc"].detach() + + if self.sky_loss_value > 0: + assert ( + self.criterion.reduction == "none" + ), "sky_loss_value should be 0 if no conf loss" + for i, l in enumerate(ls_cross): + ls_cross[i] = torch.where( + skys_cross[i][masks_cross[i]], self.sky_loss_value, l + ) + + for i in range(len(ls_cross)): + details[self_name + f"_pts3d/{i+1}"] = float( + ls_cross[i].mean() if ls_cross[i].numel() > 0 else 0 + ) + details[f"conf_{i+1}"] = preds[i]["conf"].detach() + + ls = ls_cross + masks = masks_cross + details["img_ids"] = ( + np.arange(len(ls_cross)).tolist() + ) + details["pose_loss"] = self.compute_pose_loss(gt_poses, pr_poses, pose_masks) + + return Sum(*list(zip(ls, masks))), (details | monitoring) + + +class Regr3DPoseBatchList(Regr3DPose): + """Ensure that all 3D points are correct. + Asymmetric loss: view1 is supposed to be the anchor. + + P1 = RT1 @ D1 + P2 = RT2 @ D2 + loss1 = (I @ pred_D1) - (RT1^-1 @ RT1 @ D1) + loss2 = (RT21 @ pred_D2) - (RT1^-1 @ P2) + = (RT21 @ pred_D2) - (RT1^-1 @ RT2 @ D2) + """ + + def __init__( + self, + criterion, + norm_mode="?avg_dis", + gt_scale=False, + sky_loss_value=2, + max_metric_scale=False, + ): + super().__init__( + criterion, norm_mode, gt_scale, sky_loss_value, max_metric_scale + ) + self.depth_only_criterion = DepthScaleShiftInvLoss() + self.single_view_criterion = ScaleInvLoss() + + def reorg(self, ls_b, masks_b): + ids_split = [mask.sum(dim=(1, 2)) for mask in masks_b] + ls = [[] for _ in range(len(masks_b[0]))] + for i in range(len(ls_b)): + ls_splitted_i = torch.split(ls_b[i], ids_split[i].tolist()) + for j in range(len(masks_b[0])): + ls[j].append(ls_splitted_i[j]) + ls = [torch.cat(l) for l in ls] + return ls + + def compute_loss(self, gts, preds, **kw): + ( + gt_pts_cross, + pred_pts_cross, + gt_poses, + pr_poses, + masks, + skys, + pose_masks, + monitoring, + ) = self.get_all_pts3d(gts, preds, **kw) + + if self.sky_loss_value > 0: + assert ( + self.criterion.reduction == "none" + ), "sky_loss_value should be 0 if no conf loss" + masks = [mask | sky for mask, sky in zip(masks, skys)] + + camera_only = gts[0]["camera_only"] + depth_only = gts[0]["depth_only"] + single_view = gts[0]["single_view"] + is_metric = gts[0]["is_metric"] + + # self view loss and details + if "Quantile" in self.criterion.__class__.__name__: + raise NotImplementedError + else: + # list [(B, h, w, 3)] x num_views -> list [num_views, h, w, 3] x B + masks_b = torch.unbind(torch.stack(masks, dim=1), dim=0) + + + self_name = type(self).__name__ + + gt_pts_cross_b = torch.unbind( + torch.stack(gt_pts_cross, dim=1)[~camera_only], dim=0 + ) + pred_pts_cross_b = torch.unbind( + torch.stack(pred_pts_cross, dim=1)[~camera_only], dim=0 + ) + masks_cross_b = torch.unbind(torch.stack(masks, dim=1)[~camera_only], dim=0) + ls_cross_b = [] + for i in range(len(gt_pts_cross_b)): + if depth_only[~camera_only][i]: + ls_cross_b.append( + self.depth_only_criterion( + pred_pts_cross_b[i][..., -1], + gt_pts_cross_b[i][..., -1], + masks_cross_b[i], + ) + ) + elif single_view[~camera_only][i] and not is_metric[~camera_only][i]: + ls_cross_b.append( + self.single_view_criterion( + pred_pts_cross_b[i], gt_pts_cross_b[i], masks_cross_b[i] + ) + ) + else: + ls_cross_b.append( + self.criterion( + pred_pts_cross_b[i][masks_cross_b[i]], + gt_pts_cross_b[i][masks_cross_b[i]], + ) + ) + ls_cross = self.reorg(ls_cross_b, masks_cross_b) + + if self.sky_loss_value > 0: + assert ( + self.criterion.reduction == "none" + ), "sky_loss_value should be 0 if no conf loss" + masks_cross = [mask[~camera_only] for mask in masks] + skys_cross = [sky[~camera_only] for sky in skys] + for i, l in enumerate(ls_cross): + ls_cross[i] = torch.where( + skys_cross[i][masks_cross[i]], self.sky_loss_value, l + ) + + details = {} + for i in range(len(ls_cross)): + details[f"gt_img{i + 1}"] = gts[i]["img"].permute(0, 2, 3, 1).detach() + details[f"valid_mask_{i + 1}"] = masks[i].detach() + + if "img_mask" in gts[i] and "ray_mask" in gts[i]: + details[f"img_mask_{i + 1}"] = gts[i]["img_mask"].detach() + details[f"ray_mask_{i + 1}"] = gts[i]["ray_mask"].detach() + + if "desc" in preds[i]: + details[f"desc_{i + 1}"] = preds[i]["desc"].detach() + + for i in range(len(ls_cross)): + details[self_name + f"_pts3d/{i+1}"] = float( + ls_cross[i].mean() if ls_cross[i].numel() > 0 else 0 + ) + details[f"conf_{i+1}"] = preds[i]["conf"].detach() + + ls = ls_cross + masks = masks_cross + details["img_ids"] = ( + np.arange(len(ls_cross)).tolist() + ) + pose_masks = pose_masks * gts[i]["img_mask"] + details["pose_loss"] = self.compute_pose_loss(gt_poses, pr_poses, pose_masks) + + return Sum(*list(zip(ls, masks))), (details | monitoring) + + +class ConfLoss(MultiLoss): + """Weighted regression by learned confidence. + Assuming the input pixel_loss is a pixel-level regression loss. + + Principle: + high-confidence means high conf = 0.1 ==> conf_loss = x / 10 + alpha*log(10) + low confidence means low conf = 10 ==> conf_loss = x * 10 - alpha*log(10) + + alpha: hyperparameter + """ + + def __init__(self, pixel_loss, alpha=1): + super().__init__() + assert alpha > 0 + self.alpha = alpha + self.pixel_loss = pixel_loss.with_reduction("none") + + def get_name(self): + return f"ConfLoss({self.pixel_loss})" + + def get_conf_log(self, x): + return x, torch.log(x) + + def compute_loss(self, gts, preds, **kw): + # compute per-pixel loss + losses_and_masks, details = self.pixel_loss(gts, preds, **kw) + if "is_self" in details and "img_ids" in details: + img_ids = details["img_ids"] + else: + img_ids = list(range(len(losses_and_masks))) + + # weight by confidence + conf_losses = [] + + for i in range(len(losses_and_masks)): + pred = preds[img_ids[i]] + conf_key = "conf" + + camera_only = gts[0]["camera_only"] + conf, log_conf = self.get_conf_log( + pred[conf_key][~camera_only][losses_and_masks[i][1]] + ) + + conf_loss = losses_and_masks[i][0] * conf - self.alpha * log_conf + conf_loss = conf_loss.mean() if conf_loss.numel() > 0 else 0 + conf_losses.append(conf_loss) + + + + details[self.get_name() + f"_conf_loss/{img_ids[i]+1}"] = float( + conf_loss + ) + + details.pop("img_ids", None) + + final_loss = sum(conf_losses) / len(conf_losses) * 2.0 + if "pose_loss" in details: + final_loss = ( + final_loss + details["pose_loss"].clip(max=0.3) * 5.0 + ) # , details + if "scale_loss" in details: + final_loss = final_loss + details["scale_loss"] + return final_loss, details + + +class Regr3DPose_ScaleInv(Regr3DPose): + """Same than Regr3D but invariant to depth shift. + if gt_scale == True: enforce the prediction to take the same scale than GT + """ + + def get_all_pts3d(self, gts, preds): + # compute depth-normalized points + ( + gt_pts_cross, + pr_pts_cross, + gt_poses, + pr_poses, + masks, + skys, + pose_masks, + monitoring, + ) = super().get_all_pts3d(gts, preds) + + # measure scene scale + + _, gt_scale_cross = get_group_pointcloud_center_scale(gt_pts_cross, masks) + _, pred_scale_cross = get_group_pointcloud_center_scale(pr_pts_cross, masks) + + # prevent predictions to be in a ridiculous range + pred_scale_cross = pred_scale_cross.clip(min=1e-3, max=1e3) + + # subtract the median depth + if self.gt_scale: + + pr_pts_cross = [ + pr_pt_cross * gt_scale_cross / pred_scale_cross + for pr_pt_cross in pr_pts_cross + ] + else: + gt_pts_cross = [ + gt_pt_cross / gt_scale_cross for gt_pt_cross in gt_pts_cross + ] + pr_pts_cross = [ + pr_pt_cross / pred_scale_cross for pr_pt_cross in pr_pts_cross + ] + + return ( + gt_pts_cross, + pr_pts_cross, + gt_poses, + pr_poses, + masks, + skys, + pose_masks, + monitoring, + ) + +def closed_form_scale_and_shift(pred, gt): + """ + Args: + pred: (B, H, W, C) + gt: (B, H, W, C) + valid_mask: (B, H, W) + Returns: + scale: (B,) + shift: (B,) + """ + assert pred.dim() == 4 and gt.dim() == 4, "Inputs must be 4D tensors" + B, H, W, C = pred.shape + device = pred.device + + pred_flat = pred.view(-1, C) # (N, C) + gt_flat = gt.view(-1, C) # (N, C) + + if C == 1: + pred_mean = pred_flat.mean(dim=0) + gt_mean = gt_flat.mean(dim=0) + + numerator = ((pred_flat - pred_mean) * (gt_flat - gt_mean)).sum(dim=0) + denominator = ((pred_flat - pred_mean) ** 2).sum(dim=0).clamp(min=1e-6) + scale = numerator / denominator + + shift = gt_mean - scale * pred_mean + return scale, shift + + elif C == 3: + pred_mean = pred_flat.mean(0) + gt_mean = gt_flat.mean(0) + pred_centered = pred_flat - pred_mean + gt_centered = gt_flat - gt_mean + + scale = (pred_centered * gt_centered).sum() / (pred_centered ** 2).sum().clamp(min=1e-6) + shift = gt_mean - scale * pred_mean + return scale, shift + + else: + raise ValueError(f"Unsupported channel dimension C={C}. Only 1 or 3 channels are supported.") + +def normalize_pointcloud(pts3d, valid_mask, eps=1e-3): + """ + pts3d: B, H, W, 3 + valid_mask: B, H, W + """ + dist = pts3d.norm(dim=-1) + dist_sum = (dist * valid_mask).sum(dim=[1,2]) + valid_count = valid_mask.sum(dim=[1,2]) + + avg_scale = (dist_sum / (valid_count + eps)).clamp(min=eps, max=1e3) + + # avg_scale = avg_scale.view(-1, 1, 1, 1, 1) + + pts3d = pts3d / avg_scale.view(-1, 1, 1, 1) + return pts3d, avg_scale + +def point_map_to_normal(point_map, mask, eps=1e-6): + """ + point_map: (B, H, W, 3) - 3D points laid out in a 2D grid + mask: (B, H, W) - valid pixels (bool) + + Returns: + normals: (4, B, H, W, 3) - normal vectors for each of the 4 cross-product directions + valids: (4, B, H, W) - corresponding valid masks + """ + + with torch.cuda.amp.autocast(enabled=False): + padded_mask = F.pad(mask, (1, 1, 1, 1), mode='constant', value=0) + pts = F.pad(point_map.permute(0, 3, 1, 2), (1,1,1,1), mode='constant', value=0).permute(0, 2, 3, 1) + + center = pts[:, 1:-1, 1:-1, :] # B,H,W,3 + up = pts[:, :-2, 1:-1, :] + left = pts[:, 1:-1, :-2 , :] + down = pts[:, 2:, 1:-1, :] + right = pts[:, 1:-1, 2:, :] + + up_dir = up - center + left_dir = left - center + down_dir = down - center + right_dir = right - center + + n1 = torch.cross(up_dir, left_dir, dim=-1) # up x left + n2 = torch.cross(left_dir, down_dir, dim=-1) # left x down + n3 = torch.cross(down_dir, right_dir, dim=-1) # down x right + n4 = torch.cross(right_dir,up_dir, dim=-1) # right x up + + v1 = padded_mask[:, :-2, 1:-1] & padded_mask[:, 1:-1, 1:-1] & padded_mask[:, 1:-1, :-2] + v2 = padded_mask[:, 1:-1, :-2 ] & padded_mask[:, 1:-1, 1:-1] & padded_mask[:, 2:, 1:-1] + v3 = padded_mask[:, 2:, 1:-1] & padded_mask[:, 1:-1, 1:-1] & padded_mask[:, 1:-1, 2:] + v4 = padded_mask[:, 1:-1, 2: ] & padded_mask[:, 1:-1, 1:-1] & padded_mask[:, :-2, 1:-1] + + normals = torch.stack([n1, n2, n3, n4], dim=0) # shape [4, B, H, W, 3] + valids = torch.stack([v1, v2, v3, v4], dim=0) # shape [4, B, H, W] + + normals = F.normalize(normals, p=2, dim=-1, eps=eps) + + + # Zero out invalid entries so they don't pollute subsequent computations + # normals = normals * valids.unsqueeze(-1) + + return normals, valids + +class HuberLoss(nn.Module): + def __init__(self, delta=1e-1, reduction="mean"): + super().__init__() + self.delta = delta + self.reduction = reduction + def forward(self, pred, target): + err = pred - target + abs_err = err.abs() + sq = 0.5 * err.pow(2) / self.delta + lin = abs_err - 0.5 * self.delta + loss = torch.where(abs_err <= self.delta, sq, lin) + if self.reduction == "mean": + return loss.mean() + if self.reduction == "sum": + return loss.sum() + return loss # 'none' + +class CameraLoss(nn.Module): + def __init__(self, delta=1e-1, weights=(1.0, 1.0, 0.5)): + super().__init__() + self.huber = HuberLoss(delta=delta) + self.weights = weights + def forward(self, pred_pose, gt_pose): + loss_T = self.huber(pred_pose[..., :3], gt_pose[..., :3]) + loss_R = self.huber(pred_pose[..., 3:7], gt_pose[..., 3:7]) + loss_fl = self.huber(pred_pose[..., 7:], gt_pose[..., 7:]) + return (self.weights[0] * loss_T + self.weights[1] * loss_R + self.weights[2] * loss_fl) + +class DepthOrPmapLoss(nn.Module): + def __init__(self, alpha=0.01): + super().__init__() + self.alpha = alpha + self.grad_scales = 3 + self.gamma = 1.0 + + def gradient_loss_multi_scale(self, pred, gt, mask): + total = 0 + for s in range(self.grad_scales): + step = 2 ** s + pred_s = pred[:, ::step, ::step] + gt_s = gt[:, ::step, ::step] + mask_s = mask[:, ::step, ::step] + total += self.normal_loss(pred_s, gt_s, mask_s) + return total / self.grad_scales + + def normal_loss(self, pred, gt, mask): + pred_norm, _ = point_map_to_normal(pred, mask) + gt_norm, _ = point_map_to_normal(gt, mask) + cos_sim = F.cosine_similarity(pred_norm, gt_norm, dim=-1) + return 1 - cos_sim.mean() + + def image_gradient_loss(self, pred, gt, mask): + assert pred.dim() == 4 and pred.shape[-1] == 1 + assert gt.shape == pred.shape + + B, H, W, _ = pred.shape + device = pred.device + + dx_pred = pred[:, :, 1:] - pred[:, :, :-1] # [B,H,W-1,1] + dx_gt = gt[:, :, 1:] - gt[:, :, :-1] + dx_mask = mask[:, :, 1:] & mask[:, :, :-1] # [B,H,W-1] + + dy_pred = pred[:, 1:, :] - pred[:, :-1, :] # [B,H-1,W,1] + dy_gt = gt[:, 1:, :] - gt[:, :-1, :] + dy_mask = mask[:, 1:, :] & mask[:, :-1, :] # [B,H-1,W] + + min_h = min(dy_pred.shape[1], dx_pred.shape[1]) + min_w = min(dx_pred.shape[2], dy_pred.shape[2]) + + dx_pred = dx_pred[:, :min_h, :min_w, :] # [B,H-1,W-1,1] + dx_gt = dx_gt[:, :min_h, :min_w, :] + dx_mask = dx_mask[:, :min_h, :min_w] # [B,H-1,W-1] + + dy_pred = dy_pred[:, :min_h, :min_w, :] # [B,H-1,W-1,1] + dy_gt = dy_gt[:, :min_h, :min_w, :] + dy_mask = dy_mask[:, :min_h, :min_w] # [B,H-1,W-1] + + loss_dx = F.l1_loss(dx_pred * dx_mask.unsqueeze(-1), + dx_gt * dx_mask.unsqueeze(-1)) + loss_dy = F.l1_loss(dy_pred * dy_mask.unsqueeze(-1), + dy_gt * dy_mask.unsqueeze(-1)) + + return (loss_dx + loss_dy) / 2 + + def forward(self, pred, gt, sigma_p, sigma_g, valid_mask): + if self.training: + pred_normalized, _ = normalize_pointcloud(pred, valid_mask) + gt_normalized, _ = normalize_pointcloud(gt, valid_mask) + else: + pred_normalized, gt_normalized = pred, gt + scale, shift = closed_form_scale_and_shift( + pred_normalized, gt_normalized + ) + pred_aligned = pred_normalized * scale + shift + sigma_p = sigma_p.clamp(min=1e-6) + sigma_g = sigma_g.clamp(min=1e-6) + #sigma = 0.5 * (sigma_p + sigma_g) + sigma = sigma_p + diff = (pred_aligned - gt_normalized).abs() + + C = diff.shape[-1] + + main_loss = (sigma[..., None].expand(-1, -1, -1, C) * diff)[valid_mask[..., None].expand(-1, -1, -1, C)].mean() + + if pred.shape[-1] == 1: + grad_loss = self.image_gradient_loss(pred_aligned, gt_normalized, valid_mask) + else: + grad_loss = self.gradient_loss_multi_scale(pred_aligned, gt_normalized, valid_mask) + reg_loss = -self.alpha * torch.log(sigma.clamp(min=1e-6))[valid_mask].mean() + # return main + reg + return self.gamma * main_loss + grad_loss + reg_loss + +class TrackLoss(nn.Module): + def __init__(self): + super().__init__() + self.bce = nn.BCEWithLogitsLoss(reduction="none") + self.alpha = 0.2 + self.gamma = 1.0 + def forward(self, y_pr, y_gt, vis_pr, vis_gt, w_p, w_g): + #w = 0.5 * (w_p + w_g) + w = w_p + l_pos = (y_pr - y_gt).norm(dim=-1) + l_pos = (w * l_pos).mean() + + l_vis = self.bce(vis_pr, vis_gt.float()) + l_vis = (w * l_vis).mean() + return l_pos + l_vis + +class DistillLoss(MultiLoss): + def __init__(self, lambda_track=0.05): + super().__init__() + self.cam_loss = CameraLoss( + delta=0.1, + weights=(1.0, 1.0, 0.5) + ) + self.depth_loss = DepthOrPmapLoss(alpha=0.1)#init 0.01 now 0.1 + self.pmap_loss = DepthOrPmapLoss(alpha=0.1) + self.track_loss = TrackLoss() + self.lambda_track = lambda_track + + def get_name(self): return "DistillLoss" + + def compute_loss(self, gts, preds, + track_queries=None, track_preds=None): + # ---------- Lcamera ---------- + cam_gt = torch.stack([g['camera_pose'] for g in gts], dim=1) + cam_pr = torch.stack([p['camera_pose'] for p in preds], dim=1) + Lcamera = self.cam_loss(cam_pr, cam_gt) + + # ---------- Ldepth ---------- + depth_terms = [] + for g,p in zip(gts, preds): + if ('depth' in g) and ('depth' in p): + sigma_p = p['depth_conf'] + sigma_g = g['depth_conf'] + valid_mask = g['valid_mask'] + if not valid_mask.any(): + valid_mask = torch.ones_like(g['valid_mask']) + depth_terms.append(self.depth_loss(p['depth'], g['depth'], sigma_p, sigma_g, valid_mask)) + Ldepth = torch.stack(depth_terms).mean() if depth_terms else torch.zeros_like(Lcamera) + + # ---------- Lpmap ---------- + pmap_terms = [] + for g,p in zip(gts,preds): + sigma_p = p['conf'] + sigma_g = g['conf'] + valid_mask = g['valid_mask'] + if not valid_mask.any(): + valid_mask = torch.ones_like(g['valid_mask']) + pmap_terms.append( + self.pmap_loss(p['pts3d_in_other_view'], + g['pts3d_in_other_view'], + sigma_p, + sigma_g, + valid_mask)) + Lpmap = torch.stack(pmap_terms).mean() + + # ---------- Ltrack ---------- + if ('track' in gts[0]) and ('track' in preds[0]): + y_gt = torch.stack([g['track'] for g in gts], dim=1) + vis_gt = torch.stack([g['vis'] for g in gts], dim=1) + + y_pr = torch.stack([p['track'] for p in preds], dim=1) + vis_pr = torch.stack([p['vis'] for p in preds], dim=1) + + w_p = torch.stack([p['track_conf'] for p in preds], dim=1) + w_g = torch.stack([g['track_conf'] for g in gts], dim=1) + + + Ltrack = self.track_loss(y_pr, y_gt, vis_pr, vis_gt, w_p, w_g) + else: + Ltrack = torch.zeros_like(Lcamera) + + total = Lcamera * 20 + Ldepth * 20 + Lpmap * 10 + self.lambda_track * 10 * Ltrack + details = {} + + details['Lcamera'] = float(Lcamera) * 20 + details['Ldepth'] = float(Ldepth) * 20 + details['Lpmap'] = float(Lpmap) * 10 + details['Ltrack'] = float(Ltrack) * self.lambda_track * 10 + details['total'] = float(total) + + return total, details \ No newline at end of file diff --git a/dust3r/model.py b/dust3r/model.py new file mode 100644 index 0000000000000000000000000000000000000000..7ed9f6106fb063686990c874ede99876ebc939ab --- /dev/null +++ b/dust3r/model.py @@ -0,0 +1,1123 @@ +import sys +import os + +sys.path.append(os.path.dirname(os.path.dirname(__file__))) +from collections import OrderedDict +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch.utils.checkpoint import checkpoint +from copy import deepcopy +from functools import partial +from typing import Optional, Tuple, List, Any +from dataclasses import dataclass +from transformers import PretrainedConfig +from transformers import PreTrainedModel +from transformers.modeling_outputs import BaseModelOutput +from transformers.file_utils import ModelOutput +import time +from dust3r.utils.misc import ( + fill_default_args, + freeze_all_params, + is_symmetrized, + interleave, + transpose_to_landscape, +) +from dust3r.heads import head_factory +from dust3r.utils.camera import PoseEncoder +from dust3r.patch_embed import get_patch_embed +import dust3r.utils.path_to_croco # noqa: F401 +from models.croco import CroCoNet, CrocoConfig # noqa +from dust3r.blocks import ( + Block, + DecoderBlock, + Mlp, + Attention, + CrossAttention, + DropPath, + CustomDecoderBlock, +) # noqa + +inf = float("inf") +from accelerate.logging import get_logger + +printer = get_logger(__name__, log_level="DEBUG") + + +@dataclass +class ARCroco3DStereoOutput(ModelOutput): + """ + Custom output class for ARCroco3DStereo. + """ + + ress: Optional[List[Any]] = None + views: Optional[List[Any]] = None + + +def strip_module(state_dict): + """ + Removes the 'module.' prefix from the keys of a state_dict. + Args: + state_dict (dict): The original state_dict with possible 'module.' prefixes. + Returns: + OrderedDict: A new state_dict with 'module.' prefixes removed. + """ + new_state_dict = OrderedDict() + for k, v in state_dict.items(): + name = k[7:] if k.startswith("module.") else k + new_state_dict[name] = v + return new_state_dict + + +def load_model(model_path, device, verbose=True): + if verbose: + print("... loading model from", model_path) + ckpt = torch.load(model_path, map_location="cpu") + args = ckpt["args"].model.replace( + "ManyAR_PatchEmbed", "PatchEmbedDust3R" + ) # ManyAR only for aspect ratio not consistent + if "landscape_only" not in args: + args = args[:-2] + ", landscape_only=False))" + else: + args = args.replace(" ", "").replace( + "landscape_only=True", "landscape_only=False" + ) + assert "landscape_only=False" in args + if verbose: + print(f"instantiating : {args}") + net = eval(args) + s = net.load_state_dict(ckpt["model"], strict=False) + if verbose: + print(s) + return net.to(device) + + +class ARCroco3DStereoConfig(PretrainedConfig): + model_type = "arcroco_3d_stereo" + + def __init__( + self, + output_mode="pts3d", + head_type="linear", # or dpt + depth_mode=("exp", -float("inf"), float("inf")), + conf_mode=("exp", 1, float("inf")), + pose_mode=("exp", -float("inf"), float("inf")), + freeze="none", + landscape_only=True, + patch_embed_cls="PatchEmbedDust3R", + ray_enc_depth=2, + state_size=324, + local_mem_size=256, + state_pe="2d", + state_dec_num_heads=16, + depth_head=False, + rgb_head=False, + pose_conf_head=False, + pose_head=False, + **croco_kwargs, + ): + super().__init__() + self.output_mode = output_mode + self.head_type = head_type + self.depth_mode = depth_mode + self.conf_mode = conf_mode + self.pose_mode = pose_mode + self.freeze = freeze + self.landscape_only = landscape_only + self.patch_embed_cls = patch_embed_cls + self.ray_enc_depth = ray_enc_depth + self.state_size = state_size + self.state_pe = state_pe + self.state_dec_num_heads = state_dec_num_heads + self.local_mem_size = local_mem_size + self.depth_head = depth_head + self.rgb_head = rgb_head + self.pose_conf_head = pose_conf_head + self.pose_head = pose_head + self.croco_kwargs = croco_kwargs + + +class LocalMemory(nn.Module): + def __init__( + self, + size, + k_dim, + v_dim, + num_heads, + depth=2, + mlp_ratio=4.0, + qkv_bias=False, + drop=0.0, + attn_drop=0.0, + drop_path=0.0, + act_layer=nn.GELU, + norm_layer=nn.LayerNorm, + norm_mem=True, + rope=None, + ) -> None: + super().__init__() + self.v_dim = v_dim + self.proj_q = nn.Linear(k_dim, v_dim) + self.masked_token = nn.Parameter( + torch.randn(1, 1, v_dim) * 0.2, requires_grad=True + ) + self.mem = nn.Parameter( + torch.randn(1, size, 2 * v_dim) * 0.2, requires_grad=True + ) + self.write_blocks = nn.ModuleList( + [ + DecoderBlock( + 2 * v_dim, + num_heads, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + norm_layer=norm_layer, + attn_drop=attn_drop, + drop=drop, + drop_path=drop_path, + act_layer=act_layer, + norm_mem=norm_mem, + rope=rope, + ) + for _ in range(depth) + ] + ) + self.read_blocks = nn.ModuleList( + [ + DecoderBlock( + 2 * v_dim, + num_heads, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + norm_layer=norm_layer, + attn_drop=attn_drop, + drop=drop, + drop_path=drop_path, + act_layer=act_layer, + norm_mem=norm_mem, + rope=rope, + ) + for _ in range(depth) + ] + ) + + def update_mem(self, mem, feat_k, feat_v): + """ + mem_k: [B, size, C] + mem_v: [B, size, C] + feat_k: [B, 1, C] + feat_v: [B, 1, C] + """ + feat_k = self.proj_q(feat_k) # [B, 1, C] + feat = torch.cat([feat_k, feat_v], dim=-1) + for blk in self.write_blocks: + mem, _ = blk(mem, feat, None, None) + return mem + + def inquire(self, query, mem): + x = self.proj_q(query) # [B, 1, C] + x = torch.cat([x, self.masked_token.expand(x.shape[0], -1, -1)], dim=-1) + for blk in self.read_blocks: + x, _ = blk(x, mem, None, None) + return x[..., -self.v_dim :] + + +class ARCroco3DStereo(CroCoNet): + config_class = ARCroco3DStereoConfig + base_model_prefix = "arcroco3dstereo" + supports_gradient_checkpointing = True + + def __init__(self, config: ARCroco3DStereoConfig): + self.gradient_checkpointing = False + self.fixed_input_length = True + config.croco_kwargs = fill_default_args( + config.croco_kwargs, CrocoConfig.__init__ + ) + self.config = config + self.patch_embed_cls = config.patch_embed_cls + self.croco_args = config.croco_kwargs + croco_cfg = CrocoConfig(**self.croco_args) + super().__init__(croco_cfg) + self.enc_blocks_ray_map = nn.ModuleList( + [ + Block( + self.enc_embed_dim, + 16, + 4, + qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), + rope=self.rope, + ) + for _ in range(config.ray_enc_depth) + ] + ) + self.enc_norm_ray_map = nn.LayerNorm(self.enc_embed_dim, eps=1e-6) + self.dec_num_heads = self.croco_args["dec_num_heads"] + self.pose_head_flag = config.pose_head + if self.pose_head_flag: + self.pose_token = nn.Parameter( + torch.randn(1, 1, self.dec_embed_dim) * 0.02, requires_grad=True + ) + self.pose_retriever = LocalMemory( + size=config.local_mem_size, + k_dim=self.enc_embed_dim, + v_dim=self.dec_embed_dim, + num_heads=self.dec_num_heads, + mlp_ratio=4, + qkv_bias=True, + attn_drop=0.0, + norm_layer=partial(nn.LayerNorm, eps=1e-6), + rope=None, + ) + self.register_tokens = nn.Embedding(config.state_size, self.enc_embed_dim) + self.state_size = config.state_size + self.state_pe = config.state_pe + self.masked_img_token = nn.Parameter( + torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True + ) + self.masked_ray_map_token = nn.Parameter( + torch.randn(1, self.enc_embed_dim) * 0.02, requires_grad=True + ) + self._set_state_decoder( + self.enc_embed_dim, + self.dec_embed_dim, + config.state_dec_num_heads, + self.dec_depth, + self.croco_args.get("mlp_ratio", None), + self.croco_args.get("norm_layer", None), + self.croco_args.get("norm_im2_in_dec", None), + ) + self.set_downstream_head( + config.output_mode, + config.head_type, + config.landscape_only, + config.depth_mode, + config.conf_mode, + config.pose_mode, + config.depth_head, + config.rgb_head, + config.pose_conf_head, + config.pose_head, + **self.croco_args, + ) + self.set_freeze(config.freeze) + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path, **kw): + if os.path.isfile(pretrained_model_name_or_path): + return load_model(pretrained_model_name_or_path, device="cpu") + else: + try: + model = super(ARCroco3DStereo, cls).from_pretrained( + pretrained_model_name_or_path, **kw + ) + except TypeError as e: + raise Exception( + f"tried to load {pretrained_model_name_or_path} from huggingface, but failed" + ) + return model + + def _set_patch_embed(self, img_size=224, patch_size=16, enc_embed_dim=768): + self.patch_embed = get_patch_embed( + self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3 + ) + self.patch_embed_ray_map = get_patch_embed( + self.patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=6 + ) + + def _set_decoder( + self, + enc_embed_dim, + dec_embed_dim, + dec_num_heads, + dec_depth, + mlp_ratio, + norm_layer, + norm_im2_in_dec, + ): + self.dec_depth = dec_depth + self.dec_embed_dim = dec_embed_dim + self.decoder_embed = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True) + self.dec_blocks = nn.ModuleList( + [ + DecoderBlock( + dec_embed_dim, + dec_num_heads, + mlp_ratio=mlp_ratio, + qkv_bias=True, + norm_layer=norm_layer, + norm_mem=norm_im2_in_dec, + rope=self.rope, + ) + for i in range(dec_depth) + ] + ) + self.dec_norm = norm_layer(dec_embed_dim) + + def _set_state_decoder( + self, + enc_embed_dim, + dec_embed_dim, + dec_num_heads, + dec_depth, + mlp_ratio, + norm_layer, + norm_im2_in_dec, + ): + self.dec_depth_state = dec_depth + self.dec_embed_dim_state = dec_embed_dim + self.decoder_embed_state = nn.Linear(enc_embed_dim, dec_embed_dim, bias=True) + self.dec_blocks_state = nn.ModuleList( + [ + DecoderBlock( + dec_embed_dim, + dec_num_heads, + mlp_ratio=mlp_ratio, + qkv_bias=True, + norm_layer=norm_layer, + norm_mem=norm_im2_in_dec, + rope=self.rope, + ) + for i in range(dec_depth) + ] + ) + self.dec_norm_state = norm_layer(dec_embed_dim) + + def load_state_dict(self, ckpt, **kw): + if all(k.startswith("module") for k in ckpt): + ckpt = strip_module(ckpt) + new_ckpt = dict(ckpt) + if not any(k.startswith("dec_blocks_state") for k in ckpt): + for key, value in ckpt.items(): + if key.startswith("dec_blocks"): + new_ckpt[key.replace("dec_blocks", "dec_blocks_state")] = value + try: + return super().load_state_dict(new_ckpt, **kw) + except: + try: + new_new_ckpt = { + k: v + for k, v in new_ckpt.items() + if not k.startswith("dec_blocks") + and not k.startswith("dec_norm") + and not k.startswith("decoder_embed") + } + return super().load_state_dict(new_new_ckpt, **kw) + except: + new_new_ckpt = {} + for key in new_ckpt: + if key in self.state_dict(): + if new_ckpt[key].size() == self.state_dict()[key].size(): + new_new_ckpt[key] = new_ckpt[key] + else: + printer.info( + f"Skipping '{key}': size mismatch (ckpt: {new_ckpt[key].size()}, model: {self.state_dict()[key].size()})" + ) + else: + printer.info(f"Skipping '{key}': not found in model") + return super().load_state_dict(new_new_ckpt, **kw) + + def set_freeze(self, freeze): # this is for use by downstream models + self.freeze = freeze + to_be_frozen = { + "none": [], + "mask": [self.mask_token] if hasattr(self, "mask_token") else [], + "encoder": [ + self.patch_embed, + self.patch_embed_ray_map, + self.masked_img_token, + self.masked_ray_map_token, + self.enc_blocks, + self.enc_blocks_ray_map, + self.enc_norm, + self.enc_norm_ray_map, + ], + "encoder_and_head": [ + self.patch_embed, + self.patch_embed_ray_map, + self.masked_img_token, + self.masked_ray_map_token, + self.enc_blocks, + self.enc_blocks_ray_map, + self.enc_norm, + self.enc_norm_ray_map, + self.downstream_head, + ], + "encoder_and_decoder": [ + self.patch_embed, + self.patch_embed_ray_map, + self.masked_img_token, + self.masked_ray_map_token, + self.enc_blocks, + self.enc_blocks_ray_map, + self.enc_norm, + self.enc_norm_ray_map, + self.dec_blocks, + self.dec_blocks_state, + self.pose_retriever, + self.pose_token, + self.register_tokens, + self.decoder_embed_state, + self.decoder_embed, + self.dec_norm, + self.dec_norm_state, + ], + "decoder": [ + self.dec_blocks, + self.dec_blocks_state, + self.pose_retriever, + self.pose_token, + ], + } + freeze_all_params(to_be_frozen[freeze]) + + def _set_prediction_head(self, *args, **kwargs): + """No prediction head""" + return + + def set_downstream_head( + self, + output_mode, + head_type, + landscape_only, + depth_mode, + conf_mode, + pose_mode, + depth_head, + rgb_head, + pose_conf_head, + pose_head, + patch_size, + img_size, + **kw, + ): + assert ( + img_size[0] % patch_size == 0 and img_size[1] % patch_size == 0 + ), f"{img_size=} must be multiple of {patch_size=}" + self.output_mode = output_mode + self.head_type = head_type + self.depth_mode = depth_mode + self.conf_mode = conf_mode + self.pose_mode = pose_mode + self.downstream_head = head_factory( + head_type, + output_mode, + self, + has_conf=bool(conf_mode), + has_depth=bool(depth_head), + has_rgb=bool(rgb_head), + has_pose_conf=bool(pose_conf_head), + has_pose=bool(pose_head), + ) + self.head = transpose_to_landscape( + self.downstream_head, activate=landscape_only + ) + + def _encode_image(self, image, true_shape): + x, pos = self.patch_embed(image, true_shape=true_shape) + assert self.enc_pos_embed is None + for blk in self.enc_blocks: + if self.gradient_checkpointing and self.training: + x = checkpoint(blk, x, pos, use_reentrant=False) + else: + x = blk(x, pos) + x = self.enc_norm(x) + return [x], pos, None + + def _encode_ray_map(self, ray_map, true_shape): + x, pos = self.patch_embed_ray_map(ray_map, true_shape=true_shape) + assert self.enc_pos_embed is None + for blk in self.enc_blocks_ray_map: + if self.gradient_checkpointing and self.training: + x = checkpoint(blk, x, pos, use_reentrant=False) + else: + x = blk(x, pos) + x = self.enc_norm_ray_map(x) + return [x], pos, None + + def _encode_state(self, image_tokens, image_pos): + batch_size = image_tokens.shape[0] + state_feat = self.register_tokens( + torch.arange(self.state_size, device=image_pos.device) + ) + if self.state_pe == "1d": + state_pos = ( + torch.tensor( + [[i, i] for i in range(self.state_size)], + dtype=image_pos.dtype, + device=image_pos.device, + )[None] + .expand(batch_size, -1, -1) + .contiguous() + ) # .long() + elif self.state_pe == "2d": + width = int(self.state_size**0.5) + width = width + 1 if width % 2 == 1 else width + state_pos = ( + torch.tensor( + [[i // width, i % width] for i in range(self.state_size)], + dtype=image_pos.dtype, + device=image_pos.device, + )[None] + .expand(batch_size, -1, -1) + .contiguous() + ) + elif self.state_pe == "none": + state_pos = None + state_feat = state_feat[None].expand(batch_size, -1, -1) + return state_feat, state_pos, None + + def _encode_views(self, views, img_mask=None, ray_mask=None): + device = views[0]["img"].device + batch_size = views[0]["img"].shape[0] + given = True + if img_mask is None and ray_mask is None: + given = False + if not given: + img_mask = torch.stack( + [view["img_mask"] for view in views], dim=0 + ) # Shape: (num_views, batch_size) + ray_mask = torch.stack( + [view["ray_mask"] for view in views], dim=0 + ) # Shape: (num_views, batch_size) + imgs = torch.stack( + [view["img"] for view in views], dim=0 + ) # Shape: (num_views, batch_size, C, H, W) + ray_maps = torch.stack( + [view["ray_map"] for view in views], dim=0 + ) # Shape: (num_views, batch_size, H, W, C) + shapes = [] + for view in views: + if "true_shape" in view: + shapes.append(view["true_shape"]) + else: + shape = torch.tensor(view["img"].shape[-2:], device=device) + shapes.append(shape.unsqueeze(0).repeat(batch_size, 1)) + shapes = torch.stack(shapes, dim=0).to( + imgs.device + ) # Shape: (num_views, batch_size, 2) + imgs = imgs.view( + -1, *imgs.shape[2:] + ) # Shape: (num_views * batch_size, C, H, W) + ray_maps = ray_maps.view( + -1, *ray_maps.shape[2:] + ) # Shape: (num_views * batch_size, H, W, C) + shapes = shapes.view(-1, 2) # Shape: (num_views * batch_size, 2) + img_masks_flat = img_mask.view(-1) # Shape: (num_views * batch_size) + ray_masks_flat = ray_mask.view(-1) + selected_imgs = imgs[img_masks_flat] + selected_shapes = shapes[img_masks_flat] + if selected_imgs.size(0) > 0: + img_out, img_pos, _ = self._encode_image(selected_imgs, selected_shapes) + else: + raise NotImplementedError + full_out = [ + torch.zeros( + len(views) * batch_size, *img_out[0].shape[1:], device=img_out[0].device + ) + for _ in range(len(img_out)) + ] + full_pos = torch.zeros( + len(views) * batch_size, + *img_pos.shape[1:], + device=img_pos.device, + dtype=img_pos.dtype, + ) + for i in range(len(img_out)): + full_out[i][img_masks_flat] += img_out[i] + full_out[i][~img_masks_flat] += self.masked_img_token + full_pos[img_masks_flat] += img_pos + ray_maps = ray_maps.permute(0, 3, 1, 2) # Change shape to (N, C, H, W) + selected_ray_maps = ray_maps[ray_masks_flat] + selected_shapes_ray = shapes[ray_masks_flat] + if selected_ray_maps.size(0) > 0: + ray_out, ray_pos, _ = self._encode_ray_map( + selected_ray_maps, selected_shapes_ray + ) + assert len(ray_out) == len(full_out), f"{len(ray_out)}, {len(full_out)}" + for i in range(len(ray_out)): + full_out[i][ray_masks_flat] += ray_out[i] + full_out[i][~ray_masks_flat] += self.masked_ray_map_token + full_pos[ray_masks_flat] += ( + ray_pos * (~img_masks_flat[ray_masks_flat][:, None, None]).long() + ) + else: + raymaps = torch.zeros( + 1, 6, imgs[0].shape[-2], imgs[0].shape[-1], device=img_out[0].device + ) + ray_mask_flat = torch.zeros_like(img_masks_flat) + ray_mask_flat[:1] = True + ray_out, ray_pos, _ = self._encode_ray_map(raymaps, shapes[ray_mask_flat]) + for i in range(len(ray_out)): + full_out[i][ray_mask_flat] += ray_out[i] * 0.0 + full_out[i][~ray_mask_flat] += self.masked_ray_map_token * 0.0 + return ( + shapes.chunk(len(views), dim=0), + [out.chunk(len(views), dim=0) for out in full_out], + full_pos.chunk(len(views), dim=0), + ) + + def _decoder(self, f_state, pos_state, f_img, pos_img, f_pose, pos_pose): + final_output = [(f_state, f_img)] # before projection + assert f_state.shape[-1] == self.dec_embed_dim + f_img = self.decoder_embed(f_img) + if self.pose_head_flag: + assert f_pose is not None and pos_pose is not None + f_img = torch.cat([f_pose, f_img], dim=1) + pos_img = torch.cat([pos_pose, pos_img], dim=1) + final_output.append((f_state, f_img)) + for blk_state, blk_img in zip(self.dec_blocks_state, self.dec_blocks): + if ( + self.gradient_checkpointing + and self.training + and torch.is_grad_enabled() + ): + f_state, _ = checkpoint( + blk_state, + *final_output[-1][::+1], + pos_state, + pos_img, + use_reentrant=not self.fixed_input_length, + ) + f_img, _ = checkpoint( + blk_img, + *final_output[-1][::-1], + pos_img, + pos_state, + use_reentrant=not self.fixed_input_length, + ) + else: + f_state, _ = blk_state(*final_output[-1][::+1], pos_state, pos_img) + f_img, _ = blk_img(*final_output[-1][::-1], pos_img, pos_state) + final_output.append((f_state, f_img)) + del final_output[1] # duplicate with final_output[0] + final_output[-1] = ( + self.dec_norm_state(final_output[-1][0]), + self.dec_norm(final_output[-1][1]), + ) + return zip(*final_output) + + def _downstream_head(self, decout, img_shape, **kwargs): + B, S, D = decout[-1].shape + head = getattr(self, f"head") + return head(decout, img_shape, **kwargs) + + def _init_state(self, image_tokens, image_pos): + """ + Current Version: input the first frame img feature and pose to initialize the state feature and pose + """ + state_feat, state_pos, _ = self._encode_state(image_tokens, image_pos) + state_feat = self.decoder_embed_state(state_feat) + return state_feat, state_pos + + def _recurrent_rollout( + self, + state_feat, + state_pos, + current_feat, + current_pos, + pose_feat, + pose_pos, + init_state_feat, + img_mask=None, + reset_mask=None, + update=None, + ): + new_state_feat, dec = self._decoder( + state_feat, state_pos, current_feat, current_pos, pose_feat, pose_pos + ) + new_state_feat = new_state_feat[-1] + return new_state_feat, dec + + def _get_img_level_feat(self, feat): + return torch.mean(feat, dim=1, keepdim=True) + + def _forward_encoder(self, views): + shape, feat_ls, pos = self._encode_views(views) + feat = feat_ls[-1] + state_feat, state_pos = self._init_state(feat[0], pos[0]) + mem = self.pose_retriever.mem.expand(feat[0].shape[0], -1, -1) + init_state_feat = state_feat.clone() + init_mem = mem.clone() + return (feat, pos, shape), ( + init_state_feat, + init_mem, + state_feat, + state_pos, + mem, + ) + + def _forward_decoder_step( + self, + views, + i, + feat_i, + pos_i, + shape_i, + init_state_feat, + init_mem, + state_feat, + state_pos, + mem, + ): + if self.pose_head_flag: + global_img_feat_i = self._get_img_level_feat(feat_i) + if i == 0: + pose_feat_i = self.pose_token.expand(feat_i.shape[0], -1, -1) + else: + pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem) + pose_pos_i = -torch.ones( + feat_i.shape[0], 1, 2, device=feat_i.device, dtype=pos_i.dtype + ) + else: + pose_feat_i = None + pose_pos_i = None + new_state_feat, dec = self._recurrent_rollout( + state_feat, + state_pos, + feat_i, + pos_i, + pose_feat_i, + pose_pos_i, + init_state_feat, + img_mask=views[i]["img_mask"], + reset_mask=views[i]["reset"], + update=views[i].get("update", None), + ) + out_pose_feat_i = dec[-1][:, 0:1] + new_mem = self.pose_retriever.update_mem( + mem, global_img_feat_i, out_pose_feat_i + ) + head_input = [ + dec[0].float(), + dec[self.dec_depth * 2 // 4][:, 1:].float(), + dec[self.dec_depth * 3 // 4][:, 1:].float(), + dec[self.dec_depth].float(), + ] + res = self._downstream_head(head_input, shape_i, pos=pos_i) + img_mask = views[i]["img_mask"] + update = views[i].get("update", None) + if update is not None: + update_mask = img_mask & update # if don't update, then whatever img_mask + else: + update_mask = img_mask + update_mask = update_mask[:, None, None].float() + state_feat = new_state_feat * update_mask + state_feat * ( + 1 - update_mask + ) # update global state + mem = new_mem * update_mask + mem * (1 - update_mask) # then update local state + reset_mask = views[i]["reset"] + if reset_mask is not None: + reset_mask = reset_mask[:, None, None].float() + state_feat = init_state_feat * reset_mask + state_feat * (1 - reset_mask) + mem = init_mem * reset_mask + mem * (1 - reset_mask) + return res, (state_feat, mem) + + def _forward_impl(self, views, ret_state=False): + shape, feat_ls, pos = self._encode_views(views) + feat = feat_ls[-1] + state_feat, state_pos = self._init_state(feat[0], pos[0]) + mem = self.pose_retriever.mem.expand(feat[0].shape[0], -1, -1) + init_state_feat = state_feat.clone() + init_mem = mem.clone() + all_state_args = [(state_feat, state_pos, init_state_feat, mem, init_mem)] + ress = [] + for i in range(len(views)): + feat_i = feat[i] + pos_i = pos[i] + if self.pose_head_flag: + global_img_feat_i = self._get_img_level_feat(feat_i) + if i == 0: + pose_feat_i = self.pose_token.expand(feat_i.shape[0], -1, -1) + else: + pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem) + pose_pos_i = -torch.ones( + feat_i.shape[0], 1, 2, device=feat_i.device, dtype=pos_i.dtype + ) + else: + pose_feat_i = None + pose_pos_i = None + new_state_feat, dec = self._recurrent_rollout( + state_feat, + state_pos, + feat_i, + pos_i, + pose_feat_i, + pose_pos_i, + init_state_feat, + img_mask=views[i]["img_mask"], + reset_mask=views[i]["reset"], + update=views[i].get("update", None), + ) + out_pose_feat_i = dec[-1][:, 0:1] + new_mem = self.pose_retriever.update_mem( + mem, global_img_feat_i, out_pose_feat_i + ) + assert len(dec) == self.dec_depth + 1 + head_input = [ + dec[0].float(), + dec[self.dec_depth * 2 // 4][:, 1:].float(), + dec[self.dec_depth * 3 // 4][:, 1:].float(), + dec[self.dec_depth].float(), + ] + res = self._downstream_head(head_input, shape[i], pos=pos_i) + ress.append(res) + img_mask = views[i]["img_mask"] + update = views[i].get("update", None) + if update is not None: + update_mask = ( + img_mask & update + ) # if don't update, then whatever img_mask + else: + update_mask = img_mask + update_mask = update_mask[:, None, None].float() + state_feat = new_state_feat * update_mask + state_feat * ( + 1 - update_mask + ) # update global state + mem = new_mem * update_mask + mem * ( + 1 - update_mask + ) # then update local state + reset_mask = views[i]["reset"] + if reset_mask is not None: + reset_mask = reset_mask[:, None, None].float() + state_feat = init_state_feat * reset_mask + state_feat * ( + 1 - reset_mask + ) + mem = init_mem * reset_mask + mem * (1 - reset_mask) + all_state_args.append( + (state_feat, state_pos, init_state_feat, mem, init_mem) + ) + if ret_state: + return ress, views, all_state_args + return ress, views + + def forward(self, views, ret_state=False): + if ret_state: + ress, views, state_args = self._forward_impl(views, ret_state=ret_state) + return ARCroco3DStereoOutput(ress=ress, views=views), state_args + else: + ress, views = self._forward_impl(views, ret_state=ret_state) + return ARCroco3DStereoOutput(ress=ress, views=views) + + def inference_step( + self, view, state_feat, state_pos, init_state_feat, mem, init_mem + ): + batch_size = view["img"].shape[0] + raymaps = [] + shapes = [] + for j in range(batch_size): + assert view["ray_mask"][j] + raymap = view["ray_map"][[j]].permute(0, 3, 1, 2) + raymaps.append(raymap) + shapes.append( + view.get( + "true_shape", + torch.tensor(view["ray_map"].shape[-2:])[None].repeat( + view["ray_map"].shape[0], 1 + ), + )[[j]] + ) + + raymaps = torch.cat(raymaps, dim=0) + shape = torch.cat(shapes, dim=0).to(raymaps.device) + feat_ls, pos, _ = self._encode_ray_map(raymaps, shapes) + + feat_i = feat_ls[-1] + pos_i = pos + if self.pose_head_flag: + global_img_feat_i = self._get_img_level_feat(feat_i) + pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem) + pose_pos_i = -torch.ones( + feat_i.shape[0], 1, 2, device=feat_i.device, dtype=pos_i.dtype + ) + else: + pose_feat_i = None + pose_pos_i = None + new_state_feat, dec = self._recurrent_rollout( + state_feat, + state_pos, + feat_i, + pos_i, + pose_feat_i, + pose_pos_i, + init_state_feat, + img_mask=view["img_mask"], + reset_mask=view["reset"], + update=view.get("update", None), + ) + + out_pose_feat_i = dec[-1][:, 0:1] + new_mem = self.pose_retriever.update_mem( + mem, global_img_feat_i, out_pose_feat_i + ) + assert len(dec) == self.dec_depth + 1 + head_input = [ + dec[0].float(), + dec[self.dec_depth * 2 // 4][:, 1:].float(), + dec[self.dec_depth * 3 // 4][:, 1:].float(), + dec[self.dec_depth].float(), + ] + res = self._downstream_head(head_input, shape, pos=pos_i) + return res, view + + def forward_recurrent(self, views, device, ret_state=False): + ress = [] + all_state_args = [] + for i, view in enumerate(views): + device = view["img"].device + batch_size = view["img"].shape[0] + img_mask = view["img_mask"].reshape( + -1, batch_size + ) # Shape: (1, batch_size) + ray_mask = view["ray_mask"].reshape( + -1, batch_size + ) # Shape: (1, batch_size) + imgs = view["img"].unsqueeze(0) # Shape: (1, batch_size, C, H, W) + ray_maps = view["ray_map"].unsqueeze( + 0 + ) # Shape: (num_views, batch_size, H, W, C) + shapes = ( + view["true_shape"].unsqueeze(0) + if "true_shape" in view + else torch.tensor(view["img"].shape[-2:], device=device) + .unsqueeze(0) + .repeat(batch_size, 1) + .unsqueeze(0) + ) # Shape: (num_views, batch_size, 2) + imgs = imgs.view( + -1, *imgs.shape[2:] + ) # Shape: (num_views * batch_size, C, H, W) + ray_maps = ray_maps.view( + -1, *ray_maps.shape[2:] + ) # Shape: (num_views * batch_size, H, W, C) + shapes = shapes.view(-1, 2).to( + imgs.device + ) # Shape: (num_views * batch_size, 2) + img_masks_flat = img_mask.view(-1) # Shape: (num_views * batch_size) + ray_masks_flat = ray_mask.view(-1) + selected_imgs = imgs[img_masks_flat] + selected_shapes = shapes[img_masks_flat] + if selected_imgs.size(0) > 0: + img_out, img_pos, _ = self._encode_image(selected_imgs, selected_shapes) + else: + img_out, img_pos = None, None + ray_maps = ray_maps.permute(0, 3, 1, 2) # Change shape to (N, C, H, W) + selected_ray_maps = ray_maps[ray_masks_flat] + selected_shapes_ray = shapes[ray_masks_flat] + if selected_ray_maps.size(0) > 0: + ray_out, ray_pos, _ = self._encode_ray_map( + selected_ray_maps, selected_shapes_ray + ) + else: + ray_out, ray_pos = None, None + + shape = shapes + if img_out is not None and ray_out is None: + feat_i = img_out[-1] + pos_i = img_pos + elif img_out is None and ray_out is not None: + feat_i = ray_out[-1] + pos_i = ray_pos + elif img_out is not None and ray_out is not None: + feat_i = img_out[-1] + ray_out[-1] + pos_i = img_pos + else: + raise NotImplementedError + + if i == 0: + state_feat, state_pos = self._init_state(feat_i, pos_i) + mem = self.pose_retriever.mem.expand(feat_i.shape[0], -1, -1) + init_state_feat = state_feat.clone() + init_mem = mem.clone() + all_state_args.append( + (state_feat, state_pos, init_state_feat, mem, init_mem) + ) + + if self.pose_head_flag: + global_img_feat_i = self._get_img_level_feat(feat_i) + if i == 0: + pose_feat_i = self.pose_token.expand(feat_i.shape[0], -1, -1) + else: + pose_feat_i = self.pose_retriever.inquire(global_img_feat_i, mem) + pose_pos_i = -torch.ones( + feat_i.shape[0], 1, 2, device=feat_i.device, dtype=pos_i.dtype + ) + else: + pose_feat_i = None + pose_pos_i = None + new_state_feat, dec = self._recurrent_rollout( + state_feat, + state_pos, + feat_i, + pos_i, + pose_feat_i, + pose_pos_i, + init_state_feat, + img_mask=view["img_mask"], + reset_mask=view["reset"], + update=view.get("update", None), + ) + out_pose_feat_i = dec[-1][:, 0:1] + new_mem = self.pose_retriever.update_mem( + mem, global_img_feat_i, out_pose_feat_i + ) + assert len(dec) == self.dec_depth + 1 + head_input = [ + dec[0].float(), + dec[self.dec_depth * 2 // 4][:, 1:].float(), + dec[self.dec_depth * 3 // 4][:, 1:].float(), + dec[self.dec_depth].float(), + ] + res = self._downstream_head(head_input, shape, pos=pos_i) + ress.append(res) + img_mask = view["img_mask"] + update = view.get("update", None) + if update is not None: + update_mask = ( + img_mask & update + ) # if don't update, then whatever img_mask + else: + update_mask = img_mask + update_mask = update_mask[:, None, None].float() + state_feat = new_state_feat * update_mask + state_feat * ( + 1 - update_mask + ) # update global state + mem = new_mem * update_mask + mem * ( + 1 - update_mask + ) # then update local state + reset_mask = view["reset"] + if reset_mask is not None: + reset_mask = reset_mask[:, None, None].float() + state_feat = init_state_feat * reset_mask + state_feat * ( + 1 - reset_mask + ) + mem = init_mem * reset_mask + mem * (1 - reset_mask) + all_state_args.append( + (state_feat, state_pos, init_state_feat, mem, init_mem) + ) + if ret_state: + return ress, views, all_state_args + return ress, views + + +if __name__ == "__main__": + print(ARCroco3DStereo.mro()) + cfg = ARCroco3DStereoConfig( + state_size=256, + pos_embed="RoPE100", + rgb_head=True, + pose_head=True, + img_size=(224, 224), + head_type="linear", + output_mode="pts3d+pose", + depth_mode=("exp", -inf, inf), + conf_mode=("exp", 1, inf), + pose_mode=("exp", -inf, inf), + enc_embed_dim=1024, + enc_depth=24, + enc_num_heads=16, + dec_embed_dim=768, + dec_depth=12, + dec_num_heads=12, + ) + ARCroco3DStereo(cfg) diff --git a/dust3r/patch_embed.py b/dust3r/patch_embed.py new file mode 100644 index 0000000000000000000000000000000000000000..6cc177f0b05940b5e9ee01b9053fbf24be6d1905 --- /dev/null +++ b/dust3r/patch_embed.py @@ -0,0 +1,93 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import torch +import dust3r.utils.path_to_croco # noqa: F401 +from models.blocks import PatchEmbed # noqa + + +def get_patch_embed(patch_embed_cls, img_size, patch_size, enc_embed_dim, in_chans=3): + assert patch_embed_cls in ["PatchEmbedDust3R", "ManyAR_PatchEmbed"] + patch_embed = eval(patch_embed_cls)(img_size, patch_size, in_chans, enc_embed_dim) + return patch_embed + + +class PatchEmbedDust3R(PatchEmbed): + def forward(self, x, **kw): + B, C, H, W = x.shape + assert ( + H % self.patch_size[0] == 0 + ), f"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]})." + assert ( + W % self.patch_size[1] == 0 + ), f"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]})." + x = self.proj(x) + pos = self.position_getter(B, x.size(2), x.size(3), x.device) + if self.flatten: + x = x.flatten(2).transpose(1, 2) # BCHW -> BNC + x = self.norm(x) + return x, pos + + +class ManyAR_PatchEmbed(PatchEmbed): + """Handle images with non-square aspect ratio. + All images in the same batch have the same aspect ratio. + true_shape = [(height, width) ...] indicates the actual shape of each image. + """ + + def __init__( + self, + img_size=224, + patch_size=16, + in_chans=3, + embed_dim=768, + norm_layer=None, + flatten=True, + ): + self.embed_dim = embed_dim + super().__init__(img_size, patch_size, in_chans, embed_dim, norm_layer, flatten) + + def forward(self, img, true_shape): + B, C, H, W = img.shape + + assert ( + H % self.patch_size[0] == 0 + ), f"Input image height ({H}) is not a multiple of patch size ({self.patch_size[0]})." + assert ( + W % self.patch_size[1] == 0 + ), f"Input image width ({W}) is not a multiple of patch size ({self.patch_size[1]})." + assert true_shape.shape == ( + B, + 2, + ), f"true_shape has the wrong shape={true_shape.shape}" + + W //= self.patch_size[0] + H //= self.patch_size[1] + n_tokens = H * W + + height, width = true_shape.T + + is_landscape = torch.ones_like(width, dtype=torch.bool) + is_portrait = ~is_landscape + + x = img.new_zeros((B, n_tokens, self.embed_dim)) + pos = img.new_zeros((B, n_tokens, 2), dtype=torch.int64) + + x[is_landscape] = ( + self.proj(img[is_landscape]).permute(0, 2, 3, 1).flatten(1, 2).float() + ) + x[is_portrait] = ( + self.proj(img[is_portrait].swapaxes(-1, -2)) + .permute(0, 2, 3, 1) + .flatten(1, 2) + .float() + ) + + pos[is_landscape] = self.position_getter(1, H, W, pos.device) + pos[is_portrait] = self.position_getter(1, W, H, pos.device) + + x = self.norm(x) + return x, pos diff --git a/dust3r/post_process.py b/dust3r/post_process.py new file mode 100644 index 0000000000000000000000000000000000000000..04a6597b33f2074f32b05477437dde2b940b3532 --- /dev/null +++ b/dust3r/post_process.py @@ -0,0 +1,64 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import numpy as np +import torch +from dust3r.utils.geometry import xy_grid + + +def estimate_focal_knowing_depth( + pts3d, pp, focal_mode="median", min_focal=0.0, max_focal=np.inf +): + """Reprojection method, for when the absolute depth is known: + 1) estimate the camera focal using a robust estimator + 2) reproject points onto true rays, minimizing a certain error + """ + B, H, W, THREE = pts3d.shape + assert THREE == 3 + + pixels = xy_grid(W, H, device=pts3d.device).view(1, -1, 2) - pp.view( + -1, 1, 2 + ) # B,HW,2 + pts3d = pts3d.flatten(1, 2) # (B, HW, 3) + + if focal_mode == "median": + with torch.no_grad(): + + u, v = pixels.unbind(dim=-1) + x, y, z = pts3d.unbind(dim=-1) + fx_votes = (u * z) / x + fy_votes = (v * z) / y + + f_votes = torch.cat((fx_votes.view(B, -1), fy_votes.view(B, -1)), dim=-1) + focal = torch.nanmedian(f_votes, dim=-1).values + + elif focal_mode == "weiszfeld": + + xy_over_z = (pts3d[..., :2] / pts3d[..., 2:3]).nan_to_num( + posinf=0, neginf=0 + ) # homogeneous (x,y,1) + + dot_xy_px = (xy_over_z * pixels).sum(dim=-1) + dot_xy_xy = xy_over_z.square().sum(dim=-1) + + focal = dot_xy_px.mean(dim=1) / dot_xy_xy.mean(dim=1) + + for iter in range(10): + + dis = (pixels - focal.view(-1, 1, 1) * xy_over_z).norm(dim=-1) + + w = dis.clip(min=1e-8).reciprocal() + + focal = (w * dot_xy_px).mean(dim=1) / (w * dot_xy_xy).mean(dim=1) + else: + raise ValueError(f"bad {focal_mode=}") + + focal_base = max(H, W) / ( + 2 * np.tan(np.deg2rad(60) / 2) + ) # size / 1.1547005383792515 + focal = focal.clip(min=min_focal * focal_base, max=max_focal * focal_base) + + return focal diff --git a/dust3r/utils/__init__.py b/dust3r/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8b137891791fe96927ad78e64b0aad7bded08bdc --- /dev/null +++ b/dust3r/utils/__init__.py @@ -0,0 +1 @@ + diff --git a/dust3r/utils/camera.py b/dust3r/utils/camera.py new file mode 100644 index 0000000000000000000000000000000000000000..a76b52fcae78a004f74ae4fc1a4c187b743c5e57 --- /dev/null +++ b/dust3r/utils/camera.py @@ -0,0 +1,463 @@ +from typing import Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from croco.models.blocks import Mlp + + +inf = float("inf") + + +class PoseDecoder(nn.Module): + def __init__( + self, + hidden_size=768, + mlp_ratio=4, + pose_encoding_type="absT_quaR", + ): + super().__init__() + + self.pose_encoding_type = pose_encoding_type + if self.pose_encoding_type == "absT_quaR": + self.target_dim = 7 + + self.mlp = Mlp( + in_features=hidden_size, + hidden_features=int(hidden_size * mlp_ratio), + out_features=self.target_dim, + drop=0, + ) + + def forward( + self, + pose_feat, + ): + """ + pose_feat: BxC + preliminary_cameras: cameras in opencv coordinate. + """ + + pred_cameras = self.mlp(pose_feat) # Bx7, 3 for absT, 4 for quaR + return pred_cameras + + +class PoseEncoder(nn.Module): + def __init__( + self, + hidden_size=768, + mlp_ratio=4, + pose_mode=("exp", -inf, inf), + pose_encoding_type="absT_quaR", + ): + super().__init__() + self.pose_encoding_type = pose_encoding_type + self.pose_mode = pose_mode + + if self.pose_encoding_type == "absT_quaR": + self.target_dim = 7 + + self.embed_pose = PoseEmbedding( + target_dim=self.target_dim, + out_dim=hidden_size, + n_harmonic_functions=10, + append_input=True, + ) + self.pose_encoder = Mlp( + in_features=self.embed_pose.out_dim, + hidden_features=int(hidden_size * mlp_ratio), + out_features=hidden_size, + drop=0, + ) + + def forward(self, camera): + from dust3r.heads.postprocess import postprocess_pose + pose_enc = camera_to_pose_encoding( + camera, + pose_encoding_type=self.pose_encoding_type, + ).to(camera.dtype) + pose_enc = postprocess_pose(pose_enc, self.pose_mode, inverse=True) + pose_feat = self.embed_pose(pose_enc) + pose_feat = self.pose_encoder(pose_feat) + return pose_feat + + +class HarmonicEmbedding(torch.nn.Module): + def __init__( + self, + n_harmonic_functions: int = 6, + omega_0: float = 1.0, + logspace: bool = True, + append_input: bool = True, + ) -> None: + """ + The harmonic embedding layer supports the classical + Nerf positional encoding described in + `NeRF `_ + and the integrated position encoding in + `MIP-NeRF `_. + + During the inference you can provide the extra argument `diag_cov`. + + If `diag_cov is None`, it converts + rays parametrized with a `ray_bundle` to 3D points by + extending each ray according to the corresponding length. + Then it converts each feature + (i.e. vector along the last dimension) in `x` + into a series of harmonic features `embedding`, + where for each i in range(dim) the following are present + in embedding[...]:: + + [ + sin(f_1*x[..., i]), + sin(f_2*x[..., i]), + ... + sin(f_N * x[..., i]), + cos(f_1*x[..., i]), + cos(f_2*x[..., i]), + ... + cos(f_N * x[..., i]), + x[..., i], # only present if append_input is True. + ] + + where N corresponds to `n_harmonic_functions-1`, and f_i is a scalar + denoting the i-th frequency of the harmonic embedding. + + + If `diag_cov is not None`, it approximates + conical frustums following a ray bundle as gaussians, + defined by x, the means of the gaussians and diag_cov, + the diagonal covariances. + Then it converts each gaussian + into a series of harmonic features `embedding`, + where for each i in range(dim) the following are present + in embedding[...]:: + + [ + sin(f_1*x[..., i]) * exp(0.5 * f_1**2 * diag_cov[..., i,]), + sin(f_2*x[..., i]) * exp(0.5 * f_2**2 * diag_cov[..., i,]), + ... + sin(f_N * x[..., i]) * exp(0.5 * f_N**2 * diag_cov[..., i,]), + cos(f_1*x[..., i]) * exp(0.5 * f_1**2 * diag_cov[..., i,]), + cos(f_2*x[..., i]) * exp(0.5 * f_2**2 * diag_cov[..., i,]),, + ... + cos(f_N * x[..., i]) * exp(0.5 * f_N**2 * diag_cov[..., i,]), + x[..., i], # only present if append_input is True. + ] + + where N equals `n_harmonic_functions-1`, and f_i is a scalar + denoting the i-th frequency of the harmonic embedding. + + If `logspace==True`, the frequencies `[f_1, ..., f_N]` are + powers of 2: + `f_1, ..., f_N = 2**torch.arange(n_harmonic_functions)` + + If `logspace==False`, frequencies are linearly spaced between + `1.0` and `2**(n_harmonic_functions-1)`: + `f_1, ..., f_N = torch.linspace( + 1.0, 2**(n_harmonic_functions-1), n_harmonic_functions + )` + + Note that `x` is also premultiplied by the base frequency `omega_0` + before evaluating the harmonic functions. + + Args: + n_harmonic_functions: int, number of harmonic + features + omega_0: float, base frequency + logspace: bool, Whether to space the frequencies in + logspace or linear space + append_input: bool, whether to concat the original + input to the harmonic embedding. If true the + output is of the form (embed.sin(), embed.cos(), x) + """ + super().__init__() + + if logspace: + frequencies = 2.0 ** torch.arange(n_harmonic_functions, dtype=torch.float32) + else: + frequencies = torch.linspace( + 1.0, + 2.0 ** (n_harmonic_functions - 1), + n_harmonic_functions, + dtype=torch.float32, + ) + + self.register_buffer("_frequencies", frequencies * omega_0, persistent=False) + self.register_buffer( + "_zero_half_pi", + torch.tensor([0.0, 0.5 * torch.pi]), + persistent=False, + ) + self.append_input = append_input + + def forward( + self, x: torch.Tensor, diag_cov: Optional[torch.Tensor] = None, **kwargs + ) -> torch.Tensor: + """ + Args: + x: tensor of shape [..., dim] + diag_cov: An optional tensor of shape `(..., dim)` + representing the diagonal covariance matrices of our Gaussians, joined with x + as means of the Gaussians. + + Returns: + embedding: a harmonic embedding of `x` of shape + [..., (n_harmonic_functions * 2 + int(append_input)) * num_points_per_ray] + """ + + embed = x[..., None] * self._frequencies + + embed = embed[..., None, :, :] + self._zero_half_pi[..., None, None] + + embed = embed.sin() + if diag_cov is not None: + x_var = diag_cov[..., None] * torch.pow(self._frequencies, 2) + exp_var = torch.exp(-0.5 * x_var) + + embed = embed * exp_var[..., None, :, :] + + embed = embed.reshape(*x.shape[:-1], -1) + + if self.append_input: + return torch.cat([embed, x], dim=-1) + return embed + + @staticmethod + def get_output_dim_static( + input_dims: int, n_harmonic_functions: int, append_input: bool + ) -> int: + """ + Utility to help predict the shape of the output of `forward`. + + Args: + input_dims: length of the last dimension of the input tensor + n_harmonic_functions: number of embedding frequencies + append_input: whether or not to concat the original + input to the harmonic embedding + Returns: + int: the length of the last dimension of the output tensor + """ + return input_dims * (2 * n_harmonic_functions + int(append_input)) + + def get_output_dim(self, input_dims: int = 3) -> int: + """ + Same as above. The default for input_dims is 3 for 3D applications + which use harmonic embedding for positional encoding, + so the input might be xyz. + """ + return self.get_output_dim_static( + input_dims, len(self._frequencies), self.append_input + ) + + +class PoseEmbedding(nn.Module): + def __init__(self, target_dim, out_dim, n_harmonic_functions=10, append_input=True): + super().__init__() + + self._emb_pose = HarmonicEmbedding( + n_harmonic_functions=n_harmonic_functions, append_input=append_input + ) + + self.out_dim = self._emb_pose.get_output_dim(target_dim) + + def forward(self, pose_encoding): + e_pose_encoding = self._emb_pose(pose_encoding) + return e_pose_encoding + + +def _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor: + """ + Returns torch.sqrt(torch.max(0, x)) + but with a zero subgradient where x is 0. + """ + ret = torch.zeros_like(x) + positive_mask = x > 0 + ret[positive_mask] = torch.sqrt(x[positive_mask]) + return ret + + +def matrix_to_quaternion(matrix: torch.Tensor) -> torch.Tensor: + """ + Convert rotations given as rotation matrices to quaternions. + + Args: + matrix: Rotation matrices as tensor of shape (..., 3, 3). + + Returns: + quaternions with real part first, as tensor of shape (..., 4). + """ + if matrix.size(-1) != 3 or matrix.size(-2) != 3: + raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.") + + batch_dim = matrix.shape[:-2] + m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind( + matrix.reshape(batch_dim + (9,)), dim=-1 + ) + + q_abs = _sqrt_positive_part( + torch.stack( + [ + 1.0 + m00 + m11 + m22, + 1.0 + m00 - m11 - m22, + 1.0 - m00 + m11 - m22, + 1.0 - m00 - m11 + m22, + ], + dim=-1, + ) + ) + + quat_by_rijk = torch.stack( + [ + torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1), + torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1), + torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1), + torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1), + ], + dim=-2, + ) + + flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device) + quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr)) + + out = quat_candidates[ + F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, : + ].reshape(batch_dim + (4,)) + return standardize_quaternion(out) + + +def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor: + """ + Convert a unit quaternion to a standard form: one in which the real + part is non negative. + + Args: + quaternions: Quaternions with real part first, + as tensor of shape (..., 4). + + Returns: + Standardized quaternions as tensor of shape (..., 4). + """ + quaternions = F.normalize(quaternions, p=2, dim=-1) + return torch.where(quaternions[..., 0:1] < 0, -quaternions, quaternions) + + +def camera_to_pose_encoding( + camera, + pose_encoding_type="absT_quaR", +): + """ + Inverse to pose_encoding_to_camera + camera: opencv, cam2world + """ + if pose_encoding_type == "absT_quaR": + + quaternion_R = matrix_to_quaternion(camera[:, :3, :3]) + + pose_encoding = torch.cat([camera[:, :3, 3], quaternion_R], dim=-1) + else: + raise ValueError(f"Unknown pose encoding {pose_encoding_type}") + + return pose_encoding + + +def quaternion_to_matrix(quaternions: torch.Tensor) -> torch.Tensor: + """ + Convert rotations given as quaternions to rotation matrices. + + Args: + quaternions: quaternions with real part first, + as tensor of shape (..., 4). + + Returns: + Rotation matrices as tensor of shape (..., 3, 3). + """ + r, i, j, k = torch.unbind(quaternions, -1) + + two_s = 2.0 / (quaternions * quaternions).sum(-1) + + o = torch.stack( + ( + 1 - two_s * (j * j + k * k), + two_s * (i * j - k * r), + two_s * (i * k + j * r), + two_s * (i * j + k * r), + 1 - two_s * (i * i + k * k), + two_s * (j * k - i * r), + two_s * (i * k - j * r), + two_s * (j * k + i * r), + 1 - two_s * (i * i + j * j), + ), + -1, + ) + return o.reshape(quaternions.shape[:-1] + (3, 3)) + + +def pose_encoding_to_camera( + pose_encoding, + pose_encoding_type="absT_quaR", +): + """ + Args: + pose_encoding: A tensor of shape `BxC`, containing a batch of + `B` `C`-dimensional pose encodings. + pose_encoding_type: The type of pose encoding, + """ + + if pose_encoding_type == "absT_quaR": + + abs_T = pose_encoding[:, :3] + quaternion_R = pose_encoding[:, 3:7] + R = quaternion_to_matrix(quaternion_R) + else: + raise ValueError(f"Unknown pose encoding {pose_encoding_type}") + + c2w_mats = torch.eye(4, 4).to(R.dtype).to(R.device) + c2w_mats = c2w_mats[None].repeat(len(R), 1, 1) + c2w_mats[:, :3, :3] = R + c2w_mats[:, :3, 3] = abs_T + + return c2w_mats + + +def quaternion_conjugate(q): + """Compute the conjugate of quaternion q (w, x, y, z).""" + + q_conj = torch.cat([q[..., :1], -q[..., 1:]], dim=-1) + return q_conj + + +def quaternion_multiply(q1, q2): + """Multiply two quaternions q1 and q2.""" + w1, x1, y1, z1 = q1.unbind(dim=-1) + w2, x2, y2, z2 = q2.unbind(dim=-1) + + w = w1 * w2 - x1 * x2 - y1 * y2 - z1 * z2 + x = w1 * x2 + x1 * w2 + y1 * z2 - z1 * y2 + y = w1 * y2 - x1 * z2 + y1 * w2 + z1 * x2 + z = w1 * z2 + x1 * y2 - y1 * x2 + z1 * w2 + + return torch.stack((w, x, y, z), dim=-1) + + +def rotate_vector(q, v): + """Rotate vector v by quaternion q.""" + q_vec = q[..., 1:] + q_w = q[..., :1] + + t = 2.0 * torch.cross(q_vec, v, dim=-1) + v_rot = v + q_w * t + torch.cross(q_vec, t, dim=-1) + return v_rot + + +def relative_pose_absT_quatR(t1, q1, t2, q2): + """Compute the relative translation and quaternion between two poses.""" + + q1_inv = quaternion_conjugate(q1) + + q_rel = quaternion_multiply(q1_inv, q2) + + delta_t = t2 - t1 + t_rel = rotate_vector(q1_inv, delta_t) + return t_rel, q_rel diff --git a/dust3r/utils/device.py b/dust3r/utils/device.py new file mode 100644 index 0000000000000000000000000000000000000000..ad5e8a44a0e634b4590695063f028847818bf12f --- /dev/null +++ b/dust3r/utils/device.py @@ -0,0 +1,88 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import numpy as np +import torch + + +def todevice(batch, device, callback=None, non_blocking=False): + """Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy). + + batch: list, tuple, dict of tensors or other things + device: pytorch device or 'numpy' + callback: function that would be called on every sub-elements. + """ + if callback: + batch = callback(batch) + + if isinstance(batch, dict): + return {k: todevice(v, device) for k, v in batch.items()} + + if isinstance(batch, (tuple, list)): + return type(batch)(todevice(x, device) for x in batch) + + x = batch + if device == "numpy": + if isinstance(x, torch.Tensor): + x = x.detach().cpu().numpy() + elif x is not None: + if isinstance(x, np.ndarray): + x = torch.from_numpy(x) + if torch.is_tensor(x): + x = x.to(device, non_blocking=non_blocking) + return x + + +to_device = todevice # alias + + +def to_numpy(x): + return todevice(x, "numpy") + + +def to_cpu(x): + return todevice(x, "cpu") + + +def to_cuda(x): + return todevice(x, "cuda") + + +def collate_with_cat(whatever, lists=False): + if isinstance(whatever, dict): + return {k: collate_with_cat(vals, lists=lists) for k, vals in whatever.items()} + + elif isinstance(whatever, (tuple, list)): + if len(whatever) == 0: + return whatever + elem = whatever[0] + T = type(whatever) + + if elem is None: + return None + if isinstance(elem, (bool, float, int, str)): + return whatever + if isinstance(elem, tuple): + return T(collate_with_cat(x, lists=lists) for x in zip(*whatever)) + if isinstance(elem, dict): + return { + k: collate_with_cat([e[k] for e in whatever], lists=lists) for k in elem + } + + if isinstance(elem, torch.Tensor): + return listify(whatever) if lists else torch.cat(whatever) + if isinstance(elem, np.ndarray): + return ( + listify(whatever) + if lists + else torch.cat([torch.from_numpy(x) for x in whatever]) + ) + + return sum(whatever, T()) + + +def listify(elems): + return [x for e in elems for x in e] diff --git a/dust3r/utils/geometry.py b/dust3r/utils/geometry.py new file mode 100644 index 0000000000000000000000000000000000000000..1c103094978f777e4cf3fa79b2f6cdf7aa4075cd --- /dev/null +++ b/dust3r/utils/geometry.py @@ -0,0 +1,555 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import torch +import numpy as np +from scipy.spatial import cKDTree as KDTree + +from dust3r.utils.misc import invalid_to_zeros, invalid_to_nans +from dust3r.utils.device import to_numpy + + +def xy_grid( + W, + H, + device=None, + origin=(0, 0), + unsqueeze=None, + cat_dim=-1, + homogeneous=False, + **arange_kw, +): + """Output a (H,W,2) array of int32 + with output[j,i,0] = i + origin[0] + output[j,i,1] = j + origin[1] + """ + if device is None: + + arange, meshgrid, stack, ones = np.arange, np.meshgrid, np.stack, np.ones + else: + + arange = lambda *a, **kw: torch.arange(*a, device=device, **kw) + meshgrid, stack = torch.meshgrid, torch.stack + ones = lambda *a: torch.ones(*a, device=device) + + tw, th = [arange(o, o + s, **arange_kw) for s, o in zip((W, H), origin)] + grid = meshgrid(tw, th, indexing="xy") + if homogeneous: + grid = grid + (ones((H, W)),) + if unsqueeze is not None: + grid = (grid[0].unsqueeze(unsqueeze), grid[1].unsqueeze(unsqueeze)) + if cat_dim is not None: + grid = stack(grid, cat_dim) + return grid + + +def geotrf(Trf, pts, ncol=None, norm=False): + """Apply a geometric transformation to a list of 3-D points. + + H: 3x3 or 4x4 projection matrix (typically a Homography) + p: numpy/torch/tuple of coordinates. Shape must be (...,2) or (...,3) + + ncol: int. number of columns of the result (2 or 3) + norm: float. if != 0, the resut is projected on the z=norm plane. + + Returns an array of projected 2d points. + """ + assert Trf.ndim >= 2 + if isinstance(Trf, np.ndarray): + pts = np.asarray(pts) + elif isinstance(Trf, torch.Tensor): + pts = torch.as_tensor(pts, dtype=Trf.dtype) + + output_reshape = pts.shape[:-1] + ncol = ncol or pts.shape[-1] + + if ( + isinstance(Trf, torch.Tensor) + and isinstance(pts, torch.Tensor) + and Trf.ndim == 3 + and pts.ndim == 4 + ): + d = pts.shape[3] + if Trf.shape[-1] == d: + pts = torch.einsum("bij, bhwj -> bhwi", Trf, pts) + elif Trf.shape[-1] == d + 1: + pts = ( + torch.einsum("bij, bhwj -> bhwi", Trf[:, :d, :d], pts) + + Trf[:, None, None, :d, d] + ) + else: + raise ValueError(f"bad shape, not ending with 3 or 4, for {pts.shape=}") + else: + if Trf.ndim >= 3: + n = Trf.ndim - 2 + assert Trf.shape[:n] == pts.shape[:n], "batch size does not match" + Trf = Trf.reshape(-1, Trf.shape[-2], Trf.shape[-1]) + + if pts.ndim > Trf.ndim: + + pts = pts.reshape(Trf.shape[0], -1, pts.shape[-1]) + elif pts.ndim == 2: + + pts = pts[:, None, :] + + if pts.shape[-1] + 1 == Trf.shape[-1]: + Trf = Trf.swapaxes(-1, -2) # transpose Trf + pts = pts @ Trf[..., :-1, :] + Trf[..., -1:, :] + elif pts.shape[-1] == Trf.shape[-1]: + Trf = Trf.swapaxes(-1, -2) # transpose Trf + pts = pts @ Trf + else: + pts = Trf @ pts.T + if pts.ndim >= 2: + pts = pts.swapaxes(-1, -2) + + if norm: + pts = pts / pts[..., -1:] # DONT DO /= BECAUSE OF WEIRD PYTORCH BUG + if norm != 1: + pts *= norm + + res = pts[..., :ncol].reshape(*output_reshape, ncol) + return res + + +def inv(mat): + """Invert a torch or numpy matrix""" + if isinstance(mat, torch.Tensor): + return torch.linalg.inv(mat) + if isinstance(mat, np.ndarray): + return np.linalg.inv(mat) + raise ValueError(f"bad matrix type = {type(mat)}") + + +def depthmap_to_pts3d(depth, pseudo_focal, pp=None, **_): + """ + Args: + - depthmap (BxHxW array): + - pseudo_focal: [B,H,W] ; [B,2,H,W] or [B,1,H,W] + Returns: + pointmap of absolute coordinates (BxHxWx3 array) + """ + + if len(depth.shape) == 4: + B, H, W, n = depth.shape + else: + B, H, W = depth.shape + n = None + + if len(pseudo_focal.shape) == 3: # [B,H,W] + pseudo_focalx = pseudo_focaly = pseudo_focal + elif len(pseudo_focal.shape) == 4: # [B,2,H,W] or [B,1,H,W] + pseudo_focalx = pseudo_focal[:, 0] + if pseudo_focal.shape[1] == 2: + pseudo_focaly = pseudo_focal[:, 1] + else: + pseudo_focaly = pseudo_focalx + else: + raise NotImplementedError("Error, unknown input focal shape format.") + + assert pseudo_focalx.shape == depth.shape[:3] + assert pseudo_focaly.shape == depth.shape[:3] + grid_x, grid_y = xy_grid(W, H, cat_dim=0, device=depth.device)[:, None] + + if pp is None: + grid_x = grid_x - (W - 1) / 2 + grid_y = grid_y - (H - 1) / 2 + else: + grid_x = grid_x.expand(B, -1, -1) - pp[:, 0, None, None] + grid_y = grid_y.expand(B, -1, -1) - pp[:, 1, None, None] + + if n is None: + pts3d = torch.empty((B, H, W, 3), device=depth.device) + pts3d[..., 0] = depth * grid_x / pseudo_focalx + pts3d[..., 1] = depth * grid_y / pseudo_focaly + pts3d[..., 2] = depth + else: + pts3d = torch.empty((B, H, W, 3, n), device=depth.device) + pts3d[..., 0, :] = depth * (grid_x / pseudo_focalx)[..., None] + pts3d[..., 1, :] = depth * (grid_y / pseudo_focaly)[..., None] + pts3d[..., 2, :] = depth + return pts3d + + +def depthmap_to_camera_coordinates(depthmap, camera_intrinsics, pseudo_focal=None): + """ + Args: + - depthmap (HxW array): + - camera_intrinsics: a 3x3 matrix + Returns: + pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels. + """ + camera_intrinsics = np.float32(camera_intrinsics) + H, W = depthmap.shape + + assert camera_intrinsics[0, 1] == 0.0 + assert camera_intrinsics[1, 0] == 0.0 + if pseudo_focal is None: + fu = camera_intrinsics[0, 0] + fv = camera_intrinsics[1, 1] + else: + assert pseudo_focal.shape == (H, W) + fu = fv = pseudo_focal + cu = camera_intrinsics[0, 2] + cv = camera_intrinsics[1, 2] + + u, v = np.meshgrid(np.arange(W), np.arange(H)) + z_cam = depthmap + x_cam = (u - cu) * z_cam / fu + y_cam = (v - cv) * z_cam / fv + X_cam = np.stack((x_cam, y_cam, z_cam), axis=-1).astype(np.float32) + + valid_mask = depthmap > 0.0 + return X_cam, valid_mask + + +def depthmap_to_absolute_camera_coordinates( + depthmap, camera_intrinsics, camera_pose, **kw +): + """ + Args: + - depthmap (HxW array): + - camera_intrinsics: a 3x3 matrix + - camera_pose: a 4x3 or 4x4 cam2world matrix + Returns: + pointmap of absolute coordinates (HxWx3 array), and a mask specifying valid pixels. + """ + X_cam, valid_mask = depthmap_to_camera_coordinates(depthmap, camera_intrinsics) + + X_world = X_cam # default + if camera_pose is not None: + + R_cam2world = camera_pose[:3, :3] + t_cam2world = camera_pose[:3, 3] + + X_world = ( + np.einsum("ik, vuk -> vui", R_cam2world, X_cam) + t_cam2world[None, None, :] + ) + + return X_world, valid_mask + + +def colmap_to_opencv_intrinsics(K): + """ + Modify camera intrinsics to follow a different convention. + Coordinates of the center of the top-left pixels are by default: + - (0.5, 0.5) in Colmap + - (0,0) in OpenCV + """ + K = K.copy() + K[0, 2] -= 0.5 + K[1, 2] -= 0.5 + return K + + +def opencv_to_colmap_intrinsics(K): + """ + Modify camera intrinsics to follow a different convention. + Coordinates of the center of the top-left pixels are by default: + - (0.5, 0.5) in Colmap + - (0,0) in OpenCV + """ + K = K.copy() + K[0, 2] += 0.5 + K[1, 2] += 0.5 + return K + + +def normalize_pointcloud( + pts1, pts2, norm_mode="avg_dis", valid1=None, valid2=None, ret_factor=False +): + """renorm pointmaps pts1, pts2 with norm_mode""" + assert pts1.ndim >= 3 and pts1.shape[-1] == 3 + assert pts2 is None or (pts2.ndim >= 3 and pts2.shape[-1] == 3) + norm_mode, dis_mode = norm_mode.split("_") + + if norm_mode == "avg": + + nan_pts1, nnz1 = invalid_to_zeros(pts1, valid1, ndim=3) + nan_pts2, nnz2 = ( + invalid_to_zeros(pts2, valid2, ndim=3) if pts2 is not None else (None, 0) + ) + all_pts = ( + torch.cat((nan_pts1, nan_pts2), dim=1) if pts2 is not None else nan_pts1 + ) + + all_dis = all_pts.norm(dim=-1) + if dis_mode == "dis": + pass # do nothing + elif dis_mode == "log1p": + all_dis = torch.log1p(all_dis) + elif dis_mode == "warp-log1p": + + log_dis = torch.log1p(all_dis) + warp_factor = log_dis / all_dis.clip(min=1e-8) + H1, W1 = pts1.shape[1:-1] + pts1 = pts1 * warp_factor[:, : W1 * H1].view(-1, H1, W1, 1) + if pts2 is not None: + H2, W2 = pts2.shape[1:-1] + pts2 = pts2 * warp_factor[:, W1 * H1 :].view(-1, H2, W2, 1) + all_dis = log_dis # this is their true distance afterwards + else: + raise ValueError(f"bad {dis_mode=}") + + norm_factor = all_dis.sum(dim=1) / (nnz1 + nnz2 + 1e-8) + else: + + nan_pts1 = invalid_to_nans(pts1, valid1, ndim=3) + nan_pts2 = invalid_to_nans(pts2, valid2, ndim=3) if pts2 is not None else None + all_pts = ( + torch.cat((nan_pts1, nan_pts2), dim=1) if pts2 is not None else nan_pts1 + ) + + all_dis = all_pts.norm(dim=-1) + + if norm_mode == "avg": + norm_factor = all_dis.nanmean(dim=1) + elif norm_mode == "median": + norm_factor = all_dis.nanmedian(dim=1).values.detach() + elif norm_mode == "sqrt": + norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2 + else: + raise ValueError(f"bad {norm_mode=}") + + norm_factor = norm_factor.clip(min=1e-8) + while norm_factor.ndim < pts1.ndim: + norm_factor.unsqueeze_(-1) + + res = pts1 / norm_factor + if pts2 is not None: + res = (res, pts2 / norm_factor) + if ret_factor: + res = res + (norm_factor,) + return res + + +def normalize_pointcloud_group( + pts_list, + norm_mode="avg_dis", + valid_list=None, + conf_list=None, + ret_factor=False, + ret_factor_only=False, +): + """renorm pointmaps pts1, pts2 with norm_mode""" + for pts in pts_list: + assert pts.ndim >= 3 and pts.shape[-1] == 3 + + norm_mode, dis_mode = norm_mode.split("_") + + if norm_mode == "avg": + + nan_pts_list, nnz_list = zip( + *[ + invalid_to_zeros(pts1, valid1, ndim=3) + for pts1, valid1 in zip(pts_list, valid_list) + ] + ) + all_pts = torch.cat(nan_pts_list, dim=1) + if conf_list is not None: + nan_conf_list = [ + invalid_to_zeros(conf1[..., None], valid1, ndim=3)[0] + for conf1, valid1 in zip(conf_list, valid_list) + ] + all_conf = torch.cat(nan_conf_list, dim=1)[..., 0] + else: + all_conf = torch.ones_like(all_pts[..., 0]) + + all_dis = all_pts.norm(dim=-1) + if dis_mode == "dis": + pass # do nothing + elif dis_mode == "log1p": + all_dis = torch.log1p(all_dis) + elif dis_mode == "warp-log1p": + + log_dis = torch.log1p(all_dis) + warp_factor = log_dis / all_dis.clip(min=1e-8) + H_W_list = [pts.shape[1:-1] for pts in pts_list] + pts_list = [ + pts + * warp_factor[:, sum(H_W_list[:i]) : sum(H_W_list[: i + 1])].view( + -1, H, W, 1 + ) + for i, (pts, (H, W)) in enumerate(zip(pts_list, H_W_list)) + ] + all_dis = log_dis # this is their true distance afterwards + else: + raise ValueError(f"bad {dis_mode=}") + + norm_factor = (all_conf * all_dis).sum(dim=1) / (all_conf.sum(dim=1) + 1e-8) + else: + + nan_pts_list = [ + invalid_to_nans(pts1, valid1, ndim=3) + for pts1, valid1 in zip(pts_list, valid_list) + ] + + all_pts = torch.cat(nan_pts_list, dim=1) + + all_dis = all_pts.norm(dim=-1) + + if norm_mode == "avg": + norm_factor = all_dis.nanmean(dim=1) + elif norm_mode == "median": + norm_factor = all_dis.nanmedian(dim=1).values.detach() + elif norm_mode == "sqrt": + norm_factor = all_dis.sqrt().nanmean(dim=1) ** 2 + else: + raise ValueError(f"bad {norm_mode=}") + + norm_factor = norm_factor.clip(min=1e-8) + while norm_factor.ndim < pts_list[0].ndim: + norm_factor.unsqueeze_(-1) + + if ret_factor_only: + + return norm_factor + + res = [pts / norm_factor for pts in pts_list] + if ret_factor: + return res, norm_factor + return res + + +@torch.no_grad() +def get_joint_pointcloud_depth(z1, z2, valid_mask1, valid_mask2=None, quantile=0.5): + + _z1 = invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1) + _z2 = ( + invalid_to_nans(z2, valid_mask2).reshape(len(z2), -1) + if z2 is not None + else None + ) + _z = torch.cat((_z1, _z2), dim=-1) if z2 is not None else _z1 + + if quantile == 0.5: + shift_z = torch.nanmedian(_z, dim=-1).values + else: + shift_z = torch.nanquantile(_z, quantile, dim=-1) + return shift_z # (B,) + + +@torch.no_grad() +def get_group_pointcloud_depth(zs, valid_masks, quantile=0.5): + + _zs = [ + invalid_to_nans(z1, valid_mask1).reshape(len(z1), -1) + for z1, valid_mask1 in zip(zs, valid_masks) + ] + _z = torch.cat(_zs, dim=-1) + + if quantile == 0.5: + shift_z = torch.nanmedian(_z, dim=-1).values + else: + shift_z = torch.nanquantile(_z, quantile, dim=-1) + return shift_z # (B,) + + +@torch.no_grad() +def get_joint_pointcloud_center_scale( + pts1, pts2, valid_mask1=None, valid_mask2=None, z_only=False, center=True +): + + _pts1 = invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3) + _pts2 = ( + invalid_to_nans(pts2, valid_mask2).reshape(len(pts2), -1, 3) + if pts2 is not None + else None + ) + _pts = torch.cat((_pts1, _pts2), dim=1) if pts2 is not None else _pts1 + + _center = torch.nanmedian(_pts, dim=1, keepdim=True).values # (B,1,3) + if z_only: + _center[..., :2] = 0 # do not center X and Y + + _norm = ((_pts - _center) if center else _pts).norm(dim=-1) + scale = torch.nanmedian(_norm, dim=1).values + return _center[:, None, :, :], scale[:, None, None, None] + + +@torch.no_grad() +def get_group_pointcloud_center_scale(pts, valid_masks=None, z_only=False, center=True): + + _pts = [ + invalid_to_nans(pts1, valid_mask1).reshape(len(pts1), -1, 3) + for pts1, valid_mask1 in zip(pts, valid_masks) + ] + _pts = torch.cat(_pts, dim=1) + + _center = torch.nanmedian(_pts, dim=1, keepdim=True).values # (B,1,3) + if z_only: + _center[..., :2] = 0 # do not center X and Y + + _norm = ((_pts - _center) if center else _pts).norm(dim=-1) + scale = torch.nanmedian(_norm, dim=1).values + return _center[:, None, :, :], scale[:, None, None, None] + + +def find_reciprocal_matches(P1, P2): + """ + returns 3 values: + 1 - reciprocal_in_P2: a boolean array of size P2.shape[0], a "True" value indicates a match + 2 - nn2_in_P1: a int array of size P2.shape[0], it contains the indexes of the closest points in P1 + 3 - reciprocal_in_P2.sum(): the number of matches + """ + tree1 = KDTree(P1) + tree2 = KDTree(P2) + + _, nn1_in_P2 = tree2.query(P1, workers=8) + _, nn2_in_P1 = tree1.query(P2, workers=8) + + reciprocal_in_P1 = nn2_in_P1[nn1_in_P2] == np.arange(len(nn1_in_P2)) + reciprocal_in_P2 = nn1_in_P2[nn2_in_P1] == np.arange(len(nn2_in_P1)) + assert reciprocal_in_P1.sum() == reciprocal_in_P2.sum() + return reciprocal_in_P2, nn2_in_P1, reciprocal_in_P2.sum() + + +def get_med_dist_between_poses(poses): + from scipy.spatial.distance import pdist + + return np.median(pdist([to_numpy(p[:3, 3]) for p in poses])) + + +def weighted_procrustes(A, B, w, use_weights=True, eps=1e-16, return_T=False): + """ + X: torch tensor B x N x 3 + Y: torch tensor B x N x 3 + w: torch tensor B x N + """ + assert len(A) == len(B) + if use_weights: + W1 = torch.abs(w).sum(1, keepdim=True) + w_norm = (w / (W1 + eps)).unsqueeze(-1) + a_mean = (w_norm * A).sum(dim=1, keepdim=True) + b_mean = (w_norm * B).sum(dim=1, keepdim=True) + + A_c = A - a_mean + B_c = B - b_mean + + H = torch.einsum("bni,bnj->bij", A_c, w_norm * B_c) + + else: + a_mean = A.mean(axis=1, keepdim=True) + b_mean = B.mean(axis=1, keepdim=True) + + A_c = A - a_mean + B_c = B - b_mean + + H = torch.einsum("bij,bik->bjk", A_c, B_c) + + U, S, V = torch.svd(H) # U: B x 3 x 3, S: B x 3, V: B x 3 x 3 + Z = torch.eye(3).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device) + Z[:, -1, -1] = torch.sign(torch.linalg.det(U @ V.transpose(1, 2))) # B x 3 x 3 + R = V @ Z @ U.transpose(1, 2) # B x 3 x 3 + t = b_mean - torch.einsum("bij,bjk->bik", R, a_mean.transpose(-2, -1)).transpose( + -2, -1 + ) + if return_T: + T = torch.eye(4).unsqueeze(0).repeat(A.shape[0], 1, 1).to(A.device) + T[:, :3, :3] = R + T[:, :3, 3] = t.squeeze() + return T + return R, t.squeeze() diff --git a/dust3r/utils/image.py b/dust3r/utils/image.py new file mode 100644 index 0000000000000000000000000000000000000000..12c207daebbeb1848fc5c762700b81032c1689e5 --- /dev/null +++ b/dust3r/utils/image.py @@ -0,0 +1,263 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import os +import torch +import numpy as np +import PIL.Image +from PIL.ImageOps import exif_transpose +import torchvision.transforms as tvf + +os.environ["OPENCV_IO_ENABLE_OPENEXR"] = "1" +import cv2 # noqa + +try: + from pillow_heif import register_heif_opener # noqa + + register_heif_opener() + heif_support_enabled = True +except ImportError: + heif_support_enabled = False + +ImgNorm = tvf.Compose([tvf.ToTensor(), tvf.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) + + +def img_to_arr(img): + if isinstance(img, str): + img = imread_cv2(img) + return img + + +def imread_cv2(path, options=cv2.IMREAD_COLOR): + """Open an image or a depthmap with opencv-python.""" + if path.endswith((".exr", "EXR")): + options = cv2.IMREAD_ANYDEPTH + img = cv2.imread(path, options) + if img is None: + raise IOError(f"Could not load image={path} with {options=}") + if img.ndim == 3: + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) + return img + + +def rgb(ftensor, true_shape=None): + if isinstance(ftensor, list): + return [rgb(x, true_shape=true_shape) for x in ftensor] + if isinstance(ftensor, torch.Tensor): + ftensor = ftensor.detach().cpu().numpy() # H,W,3 + if ftensor.ndim == 3 and ftensor.shape[0] == 3: + ftensor = ftensor.transpose(1, 2, 0) + elif ftensor.ndim == 4 and ftensor.shape[1] == 3: + ftensor = ftensor.transpose(0, 2, 3, 1) + if true_shape is not None: + H, W = true_shape + ftensor = ftensor[:H, :W] + if ftensor.dtype == np.uint8: + img = np.float32(ftensor) / 255 + else: + img = (ftensor * 0.5) + 0.5 + return img.clip(min=0, max=1) + + +def _resize_pil_image(img, long_edge_size): + S = max(img.size) + if S > long_edge_size: + interp = PIL.Image.LANCZOS + elif S <= long_edge_size: + interp = PIL.Image.BICUBIC + new_size = tuple(int(round(x * long_edge_size / S)) for x in img.size) + return img.resize(new_size, interp) + + +def load_images(folder_or_list, size, square_ok=False, verbose=True): + """open and convert all images in a list or folder to proper input format for DUSt3R""" + if isinstance(folder_or_list, str): + if verbose: + print(f">> Loading images from {folder_or_list}") + root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list)) + + elif isinstance(folder_or_list, list): + if verbose: + print(f">> Loading a list of {len(folder_or_list)} images") + root, folder_content = "", folder_or_list + + else: + raise ValueError(f"bad {folder_or_list=} ({type(folder_or_list)})") + + supported_images_extensions = [".jpg", ".jpeg", ".png", ".bmp"] + if heif_support_enabled: + supported_images_extensions += [".heic", ".heif"] + supported_images_extensions = tuple(supported_images_extensions) + + imgs = [] + for path in folder_content: + if not path.lower().endswith(supported_images_extensions): + continue + img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert("RGB") + W1, H1 = img.size + if size == 224: + + img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1))) + else: + + img = _resize_pil_image(img, size) + W, H = img.size + cx, cy = W // 2, H // 2 + if size == 224: + half = min(cx, cy) + img = img.crop((cx - half, cy - half, cx + half, cy + half)) + else: + halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8 + if not (square_ok) and W == H: + halfh = 3 * halfw / 4 + img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh)) + + W2, H2 = img.size + if verbose: + print(f" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}") + imgs.append( + dict( + img=ImgNorm(img)[None], + true_shape=np.int32([img.size[::-1]]), + idx=len(imgs), + instance=str(len(imgs)), + ) + ) + + assert imgs, "no images foud at " + root + if verbose: + print(f" (Found {len(imgs)} images)") + return imgs + + +def load_images_for_eval( + folder_or_list, size, square_ok=False, verbose=True, crop=True +): + """open and convert all images in a list or folder to proper input format for DUSt3R""" + if isinstance(folder_or_list, str): + if verbose: + print(f">> Loading images from {folder_or_list}") + root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list)) + + elif isinstance(folder_or_list, list): + if verbose: + print(f">> Loading a list of {len(folder_or_list)} images") + root, folder_content = "", folder_or_list + + else: + raise ValueError(f"bad {folder_or_list=} ({type(folder_or_list)})") + + supported_images_extensions = [".jpg", ".jpeg", ".png"] + if heif_support_enabled: + supported_images_extensions += [".heic", ".heif"] + supported_images_extensions = tuple(supported_images_extensions) + + imgs = [] + for path in folder_content: + if not path.lower().endswith(supported_images_extensions): + continue + img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert("RGB") + W1, H1 = img.size + if size == 224: + # resize short side to 224 (then crop) + img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1))) + else: + # resize long side to 512 + img = _resize_pil_image(img, size) + W, H = img.size + cx, cy = W // 2, H // 2 + if size == 224: + half = min(cx, cy) + if crop: + img = img.crop((cx - half, cy - half, cx + half, cy + half)) + else: # resize + img = img.resize((2 * half, 2 * half), PIL.Image.LANCZOS) + else: + halfw, halfh = ((2 * cx) // 14) * 7, ((2 * cy) // 14) * 7 + if not (square_ok) and W == H: + halfh = 3 * halfw / 4 + if crop: + img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh)) + else: # resize + img = img.resize((2 * halfw, 2 * halfh), PIL.Image.LANCZOS) + W2, H2 = img.size + if verbose: + print(f" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}") + imgs.append( + dict( + img=ImgNorm(img)[None], + true_shape=np.int32([img.size[::-1]]), + idx=len(imgs), + instance=str(len(imgs)), + ) + ) + + assert imgs, "no images foud at " + root + if verbose: + print(f" (Found {len(imgs)} images)") + return imgs + + +def load_images_512(folder_or_list, size, square_ok=False, verbose=True): + """open and convert all images in a list or folder to proper input format for DUSt3R""" + if isinstance(folder_or_list, str): + if verbose: + print(f">> Loading images from {folder_or_list}") + root, folder_content = folder_or_list, sorted(os.listdir(folder_or_list)) + + elif isinstance(folder_or_list, list): + if verbose: + print(f">> Loading a list of {len(folder_or_list)} images") + root, folder_content = "", folder_or_list + + else: + raise ValueError(f"bad {folder_or_list=} ({type(folder_or_list)})") + + supported_images_extensions = [".jpg", ".jpeg", ".png", ".bmp"] + if heif_support_enabled: + supported_images_extensions += [".heic", ".heif"] + supported_images_extensions = tuple(supported_images_extensions) + + imgs = [] + for path in folder_content: + if not path.lower().endswith(supported_images_extensions): + continue + img = exif_transpose(PIL.Image.open(os.path.join(root, path))).convert("RGB") + img = img.resize((512, 384)) + W1, H1 = img.size + if size == 224: + + img = _resize_pil_image(img, round(size * max(W1 / H1, H1 / W1))) + else: + + img = _resize_pil_image(img, size) + W, H = img.size + cx, cy = W // 2, H // 2 + if size == 224: + half = min(cx, cy) + img = img.crop((cx - half, cy - half, cx + half, cy + half)) + else: + halfw, halfh = ((2 * cx) // 16) * 8, ((2 * cy) // 16) * 8 + if not (square_ok) and W == H: + halfh = 3 * halfw / 4 + img = img.crop((cx - halfw, cy - halfh, cx + halfw, cy + halfh)) + + W2, H2 = img.size + if verbose: + print(f" - adding {path} with resolution {W1}x{H1} --> {W2}x{H2}") + imgs.append( + dict( + img=ImgNorm(img)[None], + true_shape=np.int32([img.size[::-1]]), + idx=len(imgs), + instance=str(len(imgs)), + ) + ) + + assert imgs, "no images foud at " + root + if verbose: + print(f" (Found {len(imgs)} images)") + return imgs diff --git a/dust3r/utils/misc.py b/dust3r/utils/misc.py new file mode 100644 index 0000000000000000000000000000000000000000..fbb3f225ba3b0a007541eb81362cd58e1c54d916 --- /dev/null +++ b/dust3r/utils/misc.py @@ -0,0 +1,127 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import torch + + +def fill_default_args(kwargs, func): + import inspect # a bit hacky but it works reliably + + signature = inspect.signature(func) + + for k, v in signature.parameters.items(): + if v.default is inspect.Parameter.empty: + continue + kwargs.setdefault(k, v.default) + + return kwargs + + +def freeze_all_params(modules): + for module in modules: + try: + for n, param in module.named_parameters(): + param.requires_grad = False + except AttributeError: + + module.requires_grad = False + + +def is_symmetrized(gt1, gt2): + x = gt1["instance"] + y = gt2["instance"] + if len(x) == len(y) and len(x) == 1: + return False # special case of batchsize 1 + ok = True + for i in range(0, len(x), 2): + ok = ok and (x[i] == y[i + 1]) and (x[i + 1] == y[i]) + return ok + + +def flip(tensor): + """flip so that tensor[0::2] <=> tensor[1::2]""" + return torch.stack((tensor[1::2], tensor[0::2]), dim=1).flatten(0, 1) + + +def interleave(tensor1, tensor2): + res1 = torch.stack((tensor1, tensor2), dim=1).flatten(0, 1) + res2 = torch.stack((tensor2, tensor1), dim=1).flatten(0, 1) + return res1, res2 + + +def transpose_to_landscape(head, activate=True): + """Predict in the correct aspect-ratio, + then transpose the result in landscape + and stack everything back together. + """ + + def wrapper_no(decout, true_shape, **kwargs): + B = len(true_shape) + assert true_shape[0:1].allclose(true_shape), "true_shape must be all identical" + H, W = true_shape[0].cpu().tolist() + res = head(decout, (H, W), **kwargs) + return res + + def wrapper_yes(decout, true_shape, **kwargs): + B = len(true_shape) + + H, W = int(true_shape.min()), int(true_shape.max()) + + height, width = true_shape.T + is_landscape = width >= height + is_portrait = ~is_landscape + + if is_landscape.all(): + return head(decout, (H, W), **kwargs) + if is_portrait.all(): + return transposed(head(decout, (W, H), **kwargs)) + + def selout(ar): + return [d[ar] for d in decout] + + if "pos" in kwargs: + kwargs_landscape = kwargs.copy() + kwargs_landscape["pos"] = kwargs["pos"][is_landscape] + kwargs_portrait = kwargs.copy() + kwargs_portrait["pos"] = kwargs["pos"][is_portrait] + l_result = head(selout(is_landscape), (H, W), **kwargs_landscape) + p_result = transposed(head(selout(is_portrait), (W, H), **kwargs_portrait)) + + result = {} + for k in l_result | p_result: + x = l_result[k].new(B, *l_result[k].shape[1:]) + x[is_landscape] = l_result[k] + x[is_portrait] = p_result[k] + result[k] = x + + return result + + return wrapper_yes if activate else wrapper_no + + +def transposed(dic): + return {k: v.swapaxes(1, 2) if v.ndim > 2 else v for k, v in dic.items()} + + +def invalid_to_nans(arr, valid_mask, ndim=999): + if valid_mask is not None: + arr = arr.clone() + arr[~valid_mask] = float("nan") + if arr.ndim > ndim: + arr = arr.flatten(-2 - (arr.ndim - ndim), -2) + return arr + + +def invalid_to_zeros(arr, valid_mask, ndim=999): + if valid_mask is not None: + arr = arr.clone() + arr[~valid_mask] = 0 + nnz = valid_mask.view(len(valid_mask), -1).sum(1) + else: + nnz = arr.numel() // len(arr) if len(arr) else 0 # number of point per image + if arr.ndim > ndim: + arr = arr.flatten(-2 - (arr.ndim - ndim), -2) + return arr, nnz diff --git a/dust3r/utils/parallel.py b/dust3r/utils/parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..5082a85b8c66cdcddc7402c401c0c983c5f1078b --- /dev/null +++ b/dust3r/utils/parallel.py @@ -0,0 +1,87 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +from tqdm import tqdm +from multiprocessing.dummy import Pool as ThreadPool +from multiprocessing import cpu_count + + +def parallel_threads( + function, + args, + workers=0, + star_args=False, + kw_args=False, + front_num=1, + Pool=ThreadPool, + **tqdm_kw +): + """tqdm but with parallel execution. + + Will essentially return + res = [ function(arg) # default + function(*arg) # if star_args is True + function(**arg) # if kw_args is True + for arg in args] + + Note: + the first elements of args will not be parallelized. + This can be useful for debugging. + """ + while workers <= 0: + workers += cpu_count() + if workers == 1: + front_num = float("inf") + + try: + n_args_parallel = len(args) - front_num + except TypeError: + n_args_parallel = None + args = iter(args) + + front = [] + while len(front) < front_num: + try: + a = next(args) + except StopIteration: + return front # end of the iterable + front.append( + function(*a) if star_args else function(**a) if kw_args else function(a) + ) + + out = [] + with Pool(workers) as pool: + + if star_args: + futures = pool.imap(starcall, [(function, a) for a in args]) + elif kw_args: + futures = pool.imap(starstarcall, [(function, a) for a in args]) + else: + futures = pool.imap(function, args) + + for f in tqdm(futures, total=n_args_parallel, **tqdm_kw): + out.append(f) + return front + out + + +def parallel_processes(*args, **kwargs): + """Same as parallel_threads, with processes""" + import multiprocessing as mp + + kwargs["Pool"] = mp.Pool + return parallel_threads(*args, **kwargs) + + +def starcall(args): + """convenient wrapper for Process.Pool""" + function, args = args + return function(*args) + + +def starstarcall(args): + """convenient wrapper for Process.Pool""" + function, args = args + return function(**args) diff --git a/dust3r/utils/path_to_croco.py b/dust3r/utils/path_to_croco.py new file mode 100644 index 0000000000000000000000000000000000000000..7e7ce2d9ffbe8a89a0ddc81e0c1c81f571608fc9 --- /dev/null +++ b/dust3r/utils/path_to_croco.py @@ -0,0 +1,21 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import sys +import os.path as path + +HERE_PATH = path.normpath(path.dirname(__file__)) +CROCO_REPO_PATH = path.normpath(path.join(HERE_PATH, "../../croco")) +CROCO_MODELS_PATH = path.join(CROCO_REPO_PATH, "models") + +if path.isdir(CROCO_MODELS_PATH): + + sys.path.insert(0, CROCO_REPO_PATH) +else: + raise ImportError( + f"croco is not initialized, could not find: {CROCO_MODELS_PATH}.\n " + "Did you forget to run 'git submodule update --init --recursive' ?" + ) diff --git a/dust3r/utils/render.py b/dust3r/utils/render.py new file mode 100644 index 0000000000000000000000000000000000000000..bc61fa8993396c9cd850177c288eb2a798561333 --- /dev/null +++ b/dust3r/utils/render.py @@ -0,0 +1,75 @@ +import torch +from gsplat import rasterization +from dust3r.utils.geometry import inv, geotrf + + +def render( + intrinsics: torch.Tensor, + pts3d: torch.Tensor, + rgbs: torch.Tensor | None = None, + scale: float = 0.002, + opacity: float = 0.95, +): + + device = pts3d.device + batch_size = len(intrinsics) + img_size = pts3d.shape[1:3] + pts3d = pts3d.reshape(batch_size, -1, 3) + num_pts = pts3d.shape[1] + quats = torch.randn((num_pts, 4), device=device) + quats = quats / quats.norm(dim=-1, keepdim=True) + scales = scale * torch.ones((num_pts, 3), device=device) + opacities = opacity * torch.ones((num_pts), device=device) + if rgbs is not None: + assert rgbs.shape[1] == 3 + rgbs = rgbs.reshape(batch_size, 3, -1).transpose(1, 2) + else: + rgbs = torch.ones_like(pts3d[:, :, :3]) + + rendered_rgbs = [] + rendered_depths = [] + accs = [] + for i in range(batch_size): + rgbd, acc, _ = rasterization( + pts3d[i], + quats, + scales, + opacities, + rgbs[i], + torch.eye(4, device=device)[None], + intrinsics[[i]], + width=img_size[1], + height=img_size[0], + packed=False, + render_mode="RGB+D", + ) + + rendered_depths.append(rgbd[..., 3]) + + rendered_depths = torch.cat(rendered_depths, dim=0) + + return rendered_rgbs, rendered_depths, accs + + +def get_render_results(gts, preds, self_view=False): + device = preds[0]["pts3d_in_other_view"].device + with torch.no_grad(): + depths = [] + gt_depths = [] + for i, (gt, pred) in enumerate(zip(gts, preds)): + if self_view: + camera = inv(gt["camera_pose"]).to(device) + intrinsics = gt["camera_intrinsics"].to(device) + pred = pred["pts3d_in_other_view"] + else: + camera = inv(gts[0]["camera_pose"]).to(device) + intrinsics = gts[0]["camera_intrinsics"].to(device) + pred = pred["pts3d_in_other_view"] + gt_img = gt["img"].to(device) + gt_pts3d = gt["pts3d"].to(device) + + _, depth, _ = render(intrinsics, pred, gt_img) + _, gt_depth, _ = render(intrinsics, geotrf(camera, gt_pts3d), gt_img) + depths.append(depth) + gt_depths.append(gt_depth) + return depths, gt_depths diff --git a/dust3r/viz.py b/dust3r/viz.py new file mode 100644 index 0000000000000000000000000000000000000000..f25aa80cca6226d34d9f6002bc927115d0e608ed --- /dev/null +++ b/dust3r/viz.py @@ -0,0 +1,1089 @@ +# Copyright (C) 2024-present Naver Corporation. All rights reserved. +# Licensed under CC BY-NC-SA 4.0 (non-commercial use only). +# +# -------------------------------------------------------- +# modified from DUSt3R + +import PIL.Image +import numpy as np +from scipy.spatial.transform import Rotation +import torch +import cv2 +import matplotlib as mpl +import matplotlib.cm as cm +import matplotlib.pyplot as plt +from dust3r.utils.geometry import ( + geotrf, + get_med_dist_between_poses, + depthmap_to_absolute_camera_coordinates, +) +from dust3r.utils.device import to_numpy +from dust3r.utils.image import rgb, img_to_arr +from matplotlib.backends.backend_agg import FigureCanvasAgg +from matplotlib.figure import Figure + +try: + import trimesh +except ImportError: + print("/!\\ module trimesh is not installed, cannot visualize results /!\\") + + +def float2uint8(x): + return (255.0 * x).astype(np.uint8) + + +def uint82float(img): + return np.ascontiguousarray(img) / 255.0 + + +def cat_3d(vecs): + if isinstance(vecs, (np.ndarray, torch.Tensor)): + vecs = [vecs] + return np.concatenate([p.reshape(-1, 3) for p in to_numpy(vecs)]) + + +def show_raw_pointcloud(pts3d, colors, point_size=2): + scene = trimesh.Scene() + + pct = trimesh.PointCloud(cat_3d(pts3d), colors=cat_3d(colors)) + scene.add_geometry(pct) + + scene.show(line_settings={"point_size": point_size}) + + +def pts3d_to_trimesh(img, pts3d, valid=None): + H, W, THREE = img.shape + assert THREE == 3 + assert img.shape == pts3d.shape + + vertices = pts3d.reshape(-1, 3) + + idx = np.arange(len(vertices)).reshape(H, W) + idx1 = idx[:-1, :-1].ravel() # top-left corner + idx2 = idx[:-1, +1:].ravel() # right-left corner + idx3 = idx[+1:, :-1].ravel() # bottom-left corner + idx4 = idx[+1:, +1:].ravel() # bottom-right corner + faces = np.concatenate( + ( + np.c_[idx1, idx2, idx3], + np.c_[ + idx3, idx2, idx1 + ], # same triangle, but backward (cheap solution to cancel face culling) + np.c_[idx2, idx3, idx4], + np.c_[ + idx4, idx3, idx2 + ], # same triangle, but backward (cheap solution to cancel face culling) + ), + axis=0, + ) + + face_colors = np.concatenate( + ( + img[:-1, :-1].reshape(-1, 3), + img[:-1, :-1].reshape(-1, 3), + img[+1:, +1:].reshape(-1, 3), + img[+1:, +1:].reshape(-1, 3), + ), + axis=0, + ) + + if valid is not None: + assert valid.shape == (H, W) + valid_idxs = valid.ravel() + valid_faces = valid_idxs[faces].all(axis=-1) + faces = faces[valid_faces] + face_colors = face_colors[valid_faces] + + assert len(faces) == len(face_colors) + return dict(vertices=vertices, face_colors=face_colors, faces=faces) + + +def cat_meshes(meshes): + vertices, faces, colors = zip( + *[(m["vertices"], m["faces"], m["face_colors"]) for m in meshes] + ) + n_vertices = np.cumsum([0] + [len(v) for v in vertices]) + for i in range(len(faces)): + faces[i][:] += n_vertices[i] + + vertices = np.concatenate(vertices) + colors = np.concatenate(colors) + faces = np.concatenate(faces) + return dict(vertices=vertices, face_colors=colors, faces=faces) + + +def show_duster_pairs(view1, view2, pred1, pred2): + import matplotlib.pyplot as pl + + pl.ion() + + for e in range(len(view1["instance"])): + i = view1["idx"][e] + j = view2["idx"][e] + img1 = rgb(view1["img"][e]) + img2 = rgb(view2["img"][e]) + conf1 = pred1["conf"][e].squeeze() + conf2 = pred2["conf"][e].squeeze() + score = conf1.mean() * conf2.mean() + print(f">> Showing pair #{e} {i}-{j} {score=:g}") + pl.clf() + pl.subplot(221).imshow(img1) + pl.subplot(223).imshow(img2) + pl.subplot(222).imshow(conf1, vmin=1, vmax=30) + pl.subplot(224).imshow(conf2, vmin=1, vmax=30) + pts1 = pred1["pts3d"][e] + pts2 = pred2["pts3d_in_other_view"][e] + pl.subplots_adjust(0, 0, 1, 1, 0, 0) + if input("show pointcloud? (y/n) ") == "y": + show_raw_pointcloud(cat(pts1, pts2), cat(img1, img2), point_size=5) + + +def auto_cam_size(im_poses): + return 0.1 * get_med_dist_between_poses(im_poses) + + +class SceneViz: + def __init__(self): + self.scene = trimesh.Scene() + + def add_rgbd( + self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None + ): + image = img_to_arr(image) + + if intrinsics is None: + H, W, THREE = image.shape + focal = max(H, W) + intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]]) + + pts3d = depthmap_to_pts3d(depth, intrinsics, cam2world=cam2world) + + return self.add_pointcloud( + pts3d, image, mask=(depth < zfar) if mask is None else mask + ) + + def add_pointcloud(self, pts3d, color=(0, 0, 0), mask=None, denoise=False): + pts3d = to_numpy(pts3d) + mask = to_numpy(mask) + if not isinstance(pts3d, list): + pts3d = [pts3d.reshape(-1, 3)] + if mask is not None: + mask = [mask.ravel()] + if not isinstance(color, (tuple, list)): + color = [color.reshape(-1, 3)] + if mask is None: + mask = [slice(None)] * len(pts3d) + + pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)]) + pct = trimesh.PointCloud(pts) + + if isinstance(color, (list, np.ndarray, torch.Tensor)): + color = to_numpy(color) + col = np.concatenate([p[m] for p, m in zip(color, mask)]) + assert col.shape == pts.shape, bb() + pct.visual.vertex_colors = uint8(col.reshape(-1, 3)) + else: + assert len(color) == 3 + pct.visual.vertex_colors = np.broadcast_to(uint8(color), pts.shape) + + if denoise: + + centroid = np.median(pct.vertices, axis=0) + dist_to_centroid = np.linalg.norm(pct.vertices - centroid, axis=-1) + dist_thr = np.quantile(dist_to_centroid, 0.99) + valid = dist_to_centroid < dist_thr + + pct = trimesh.PointCloud( + pct.vertices[valid], color=pct.visual.vertex_colors[valid] + ) + + self.scene.add_geometry(pct) + return self + + def add_rgbd( + self, image, depth, intrinsics=None, cam2world=None, zfar=np.inf, mask=None + ): + + if intrinsics is None: + H, W, THREE = image.shape + focal = max(H, W) + intrinsics = np.float32([[focal, 0, W / 2], [0, focal, H / 2], [0, 0, 1]]) + + pts3d, mask2 = depthmap_to_absolute_camera_coordinates( + depth, intrinsics, cam2world + ) + mask2 &= depth < zfar + + if mask is not None: + mask2 &= mask + + return self.add_pointcloud(pts3d, image, mask=mask2) + + def add_camera( + self, + pose_c2w, + focal=None, + color=(0, 0, 0), + image=None, + imsize=None, + cam_size=0.03, + ): + pose_c2w, focal, color, image = to_numpy((pose_c2w, focal, color, image)) + image = img_to_arr(image) + if isinstance(focal, np.ndarray) and focal.shape == (3, 3): + intrinsics = focal + focal = (intrinsics[0, 0] * intrinsics[1, 1]) ** 0.5 + if imsize is None: + imsize = (2 * intrinsics[0, 2], 2 * intrinsics[1, 2]) + + add_scene_cam( + self.scene, + pose_c2w, + color, + image, + focal, + imsize=imsize, + screen_width=cam_size, + marker=None, + ) + return self + + def add_cameras( + self, poses, focals=None, images=None, imsizes=None, colors=None, **kw + ): + get = lambda arr, idx: None if arr is None else arr[idx] + for i, pose_c2w in enumerate(poses): + self.add_camera( + pose_c2w, + get(focals, i), + image=get(images, i), + color=get(colors, i), + imsize=get(imsizes, i), + **kw, + ) + return self + + def show(self, point_size=2): + self.scene.show(line_settings={"point_size": point_size}) + + +def show_raw_pointcloud_with_cams( + imgs, pts3d, mask, focals, cams2world, point_size=2, cam_size=0.05, cam_color=None +): + """Visualization of a pointcloud with cameras + imgs = (N, H, W, 3) or N-size list of [(H,W,3), ...] + pts3d = (N, H, W, 3) or N-size list of [(H,W,3), ...] + focals = (N,) or N-size list of [focal, ...] + cams2world = (N,4,4) or N-size list of [(4,4), ...] + """ + assert len(pts3d) == len(mask) <= len(imgs) <= len(cams2world) == len(focals) + pts3d = to_numpy(pts3d) + imgs = to_numpy(imgs) + focals = to_numpy(focals) + cams2world = to_numpy(cams2world) + + scene = trimesh.Scene() + + pts = np.concatenate([p[m] for p, m in zip(pts3d, mask)]) + col = np.concatenate([p[m] for p, m in zip(imgs, mask)]) + pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3)) + scene.add_geometry(pct) + + for i, pose_c2w in enumerate(cams2world): + if isinstance(cam_color, list): + camera_edge_color = cam_color[i] + else: + camera_edge_color = cam_color or CAM_COLORS[i % len(CAM_COLORS)] + add_scene_cam( + scene, + pose_c2w, + camera_edge_color, + imgs[i] if i < len(imgs) else None, + focals[i], + screen_width=cam_size, + ) + + scene.show(line_settings={"point_size": point_size}) + + +def add_scene_cam( + scene, + pose_c2w, + edge_color, + image=None, + focal=None, + imsize=None, + screen_width=0.03, + marker=None, +): + if image is not None: + image = np.asarray(image) + H, W, THREE = image.shape + assert THREE == 3 + if image.dtype != np.uint8: + image = np.uint8(255 * image) + elif imsize is not None: + W, H = imsize + elif focal is not None: + H = W = focal / 1.1 + else: + H = W = 1 + + if isinstance(focal, np.ndarray): + focal = focal[0] + if not focal: + focal = min(H, W) * 1.1 # default value + + height = max(screen_width / 10, focal * screen_width / H) + width = screen_width * 0.5**0.5 + rot45 = np.eye(4) + rot45[:3, :3] = Rotation.from_euler("z", np.deg2rad(45)).as_matrix() + rot45[2, 3] = -height # set the tip of the cone = optical center + aspect_ratio = np.eye(4) + aspect_ratio[0, 0] = W / H + transform = pose_c2w @ OPENGL @ aspect_ratio @ rot45 + cam = trimesh.creation.cone(width, height, sections=4) # , transform=transform) + + if image is not None: + vertices = geotrf(transform, cam.vertices[[4, 5, 1, 3]]) + faces = np.array([[0, 1, 2], [0, 2, 3], [2, 1, 0], [3, 2, 0]]) + img = trimesh.Trimesh(vertices=vertices, faces=faces) + uv_coords = np.float32([[0, 0], [1, 0], [1, 1], [0, 1]]) + img.visual = trimesh.visual.TextureVisuals( + uv_coords, image=PIL.Image.fromarray(image) + ) + scene.add_geometry(img) + + rot2 = np.eye(4) + rot2[:3, :3] = Rotation.from_euler("z", np.deg2rad(2)).as_matrix() + vertices = np.r_[cam.vertices, 0.95 * cam.vertices, geotrf(rot2, cam.vertices)] + vertices = geotrf(transform, vertices) + faces = [] + for face in cam.faces: + if 0 in face: + continue + a, b, c = face + a2, b2, c2 = face + len(cam.vertices) + a3, b3, c3 = face + 2 * len(cam.vertices) + + faces.append((a, b, b2)) + faces.append((a, a2, c)) + faces.append((c2, b, c)) + + faces.append((a, b, b3)) + faces.append((a, a3, c)) + faces.append((c3, b, c)) + + faces += [(c, b, a) for a, b, c in faces] + + cam = trimesh.Trimesh(vertices=vertices, faces=faces) + cam.visual.face_colors[:, :3] = edge_color + scene.add_geometry(cam) + + if marker == "o": + marker = trimesh.creation.icosphere(3, radius=screen_width / 4) + marker.vertices += pose_c2w[:3, 3] + marker.visual.face_colors[:, :3] = edge_color + scene.add_geometry(marker) + + +def cat(a, b): + return np.concatenate((a.reshape(-1, 3), b.reshape(-1, 3))) + + +OPENGL = np.array([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]]) + + +CAM_COLORS = [ + (255, 0, 0), + (0, 0, 255), + (0, 255, 0), + (255, 0, 255), + (255, 204, 0), + (0, 204, 204), + (128, 255, 255), + (255, 128, 255), + (255, 255, 128), + (0, 0, 0), + (128, 128, 128), +] + + +def uint8(colors): + if not isinstance(colors, np.ndarray): + colors = np.array(colors) + if np.issubdtype(colors.dtype, np.floating): + colors *= 255 + assert 0 <= colors.min() and colors.max() < 256 + return np.uint8(colors) + + +def segment_sky(image): + import cv2 + from scipy import ndimage + + image = to_numpy(image) + if np.issubdtype(image.dtype, np.floating): + image = np.uint8(255 * image.clip(min=0, max=1)) + hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV) + + lower_blue = np.array([0, 0, 100]) + upper_blue = np.array([30, 255, 255]) + mask = cv2.inRange(hsv, lower_blue, upper_blue).view(bool) + + mask |= (hsv[:, :, 1] < 10) & (hsv[:, :, 2] > 150) + mask |= (hsv[:, :, 1] < 30) & (hsv[:, :, 2] > 180) + mask |= (hsv[:, :, 1] < 50) & (hsv[:, :, 2] > 220) + + kernel = np.ones((5, 5), np.uint8) + mask2 = ndimage.binary_opening(mask, structure=kernel) + + _, labels, stats, _ = cv2.connectedComponentsWithStats( + mask2.view(np.uint8), connectivity=8 + ) + cc_sizes = stats[1:, cv2.CC_STAT_AREA] + order = cc_sizes.argsort()[::-1] # bigger first + i = 0 + selection = [] + while i < len(order) and cc_sizes[order[i]] > cc_sizes[order[0]] / 2: + selection.append(1 + order[i]) + i += 1 + mask3 = np.in1d(labels, selection).reshape(labels.shape) + + return torch.from_numpy(mask3) + + +def get_vertical_colorbar(h, vmin, vmax, cmap_name="jet", label=None, cbar_precision=2): + """ + :param w: pixels + :param h: pixels + :param vmin: min value + :param vmax: max value + :param cmap_name: + :param label + :return: + """ + fig = Figure(figsize=(2, 8), dpi=100) + fig.subplots_adjust(right=1.5) + canvas = FigureCanvasAgg(fig) + + ax = fig.add_subplot(111) + cmap = cm.get_cmap(cmap_name) + norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax) + + tick_cnt = 6 + tick_loc = np.linspace(vmin, vmax, tick_cnt) + cb1 = mpl.colorbar.ColorbarBase( + ax, cmap=cmap, norm=norm, ticks=tick_loc, orientation="vertical" + ) + + tick_label = [str(np.round(x, cbar_precision)) for x in tick_loc] + if cbar_precision == 0: + tick_label = [x[:-2] for x in tick_label] + + cb1.set_ticklabels(tick_label) + + cb1.ax.tick_params(labelsize=18, rotation=0) + if label is not None: + cb1.set_label(label) + + fig.tight_layout() + + canvas.draw() + s, (width, height) = canvas.print_to_buffer() + + im = np.frombuffer(s, np.uint8).reshape((height, width, 4)) + + im = im[:, :, :3].astype(np.float32) / 255.0 + if h != im.shape[0]: + w = int(im.shape[1] / im.shape[0] * h) + im = cv2.resize(im, (w, h), interpolation=cv2.INTER_AREA) + + return im + + +def colorize_np( + x, + cmap_name="jet", + mask=None, + range=None, + append_cbar=False, + cbar_in_image=False, + cbar_precision=2, +): + """ + turn a grayscale image into a color image + :param x: input grayscale, [H, W] + :param cmap_name: the colorization method + :param mask: the mask image, [H, W] + :param range: the range for scaling, automatic if None, [min, max] + :param append_cbar: if append the color bar + :param cbar_in_image: put the color bar inside the image to keep the output image the same size as the input image + :return: colorized image, [H, W] + """ + if range is not None: + vmin, vmax = range + elif mask is not None: + + vmin = np.min(x[mask][np.nonzero(x[mask])]) + vmax = np.max(x[mask]) + + x[np.logical_not(mask)] = vmin + + else: + vmin, vmax = np.percentile(x, (1, 100)) + vmax += 1e-6 + + x = np.clip(x, vmin, vmax) + x = (x - vmin) / (vmax - vmin) + + cmap = cm.get_cmap(cmap_name) + x_new = cmap(x)[:, :, :3] + + if mask is not None: + mask = np.float32(mask[:, :, np.newaxis]) + x_new = x_new * mask + np.ones_like(x_new) * (1.0 - mask) + + cbar = get_vertical_colorbar( + h=x.shape[0], + vmin=vmin, + vmax=vmax, + cmap_name=cmap_name, + cbar_precision=cbar_precision, + ) + + if append_cbar: + if cbar_in_image: + x_new[:, -cbar.shape[1] :, :] = cbar + else: + x_new = np.concatenate( + (x_new, np.zeros_like(x_new[:, :5, :]), cbar), axis=1 + ) + return x_new + else: + return x_new + + +def colorize( + x, cmap_name="jet", mask=None, range=None, append_cbar=False, cbar_in_image=False +): + """ + turn a grayscale image into a color image + :param x: torch.Tensor, grayscale image, [H, W] or [B, H, W] + :param mask: torch.Tensor or None, mask image, [H, W] or [B, H, W] or None + """ + + device = x.device + x = x.cpu().numpy() + if mask is not None: + mask = mask.cpu().numpy() > 0.99 + kernel = np.ones((3, 3), np.uint8) + + if x.ndim == 2: + x = x[None] + if mask is not None: + mask = mask[None] + + out = [] + for x_ in x: + if mask is not None: + mask = cv2.erode(mask.astype(np.uint8), kernel, iterations=1).astype(bool) + + x_ = colorize_np(x_, cmap_name, mask, range, append_cbar, cbar_in_image) + out.append(torch.from_numpy(x_).to(device).float()) + out = torch.stack(out).squeeze(0) + return out + + +def draw_correspondences( + imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2 +): + """ + draw correspondences between two images + :param img1: tensor [B, H, W, 3] + :param img2: tensor [B, H, W, 3] + :param coord1: tensor [B, N, 2] + :param coord2: tensor [B, N, 2] + :param interval: int the interval between two points + :param color_by: specify the color based on image 1 or image 2, 0 or 1 + :return: [B, 2*H, W, 3] + """ + batch_size = len(imgs1) + out = [] + for i in range(batch_size): + img1 = imgs1[i].detach().cpu().numpy() + img2 = imgs2[i].detach().cpu().numpy() + coord1 = ( + coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2) + ) + coord2 = ( + coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2) + ) + img = drawMatches( + img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True + ) + out.append(img) + out = np.stack(out) + return out + + +def draw_correspondences_lines( + imgs1, imgs2, coords1, coords2, interval=10, color_by=0, radius=2 +): + """ + draw correspondences between two images + :param img1: tensor [B, H, W, 3] + :param img2: tensor [B, H, W, 3] + :param coord1: tensor [B, N, 2] + :param coord2: tensor [B, N, 2] + :param interval: int the interval between two points + :param color_by: specify the color based on image 1 or image 2, 0 or 1 + :return: [B, 2*H, W, 3] + """ + batch_size = len(imgs1) + out = [] + for i in range(batch_size): + img1 = imgs1[i].detach().cpu().numpy() + img2 = imgs2[i].detach().cpu().numpy() + coord1 = ( + coords1[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2) + ) + coord2 = ( + coords2[i].detach().cpu().numpy()[::interval, ::interval].reshape(-1, 2) + ) + img = drawMatches_lines( + img1, img2, coord1, coord2, radius=radius, color_by=color_by, row_cat=True + ) + out.append(img) + out = np.stack(out) + return out + + +def drawMatches(img1, img2, kp1, kp2, radius=2, mask=None, color_by=0, row_cat=False): + + h1, w1 = img1.shape[:2] + h2, w2 = img2.shape[:2] + + img1 = np.ascontiguousarray(float2uint8(img1)) + img2 = np.ascontiguousarray(float2uint8(img2)) + + center1 = np.median(kp1, axis=0) + center2 = np.median(kp2, axis=0) + + set_max = range(128) + colors = {m: i for i, m in enumerate(set_max)} + colors = { + m: (255 * np.array(plt.cm.hsv(i / float(len(colors))))[:3][::-1]).astype( + np.int32 + ) + for m, i in colors.items() + } + + if mask is not None: + ind = np.argsort(mask)[::-1] + kp1 = kp1[ind] + kp2 = kp2[ind] + mask = mask[ind] + + for i, (pt1, pt2) in enumerate(zip(kp1, kp2)): + + if color_by == 0: + coord_angle = np.arctan2(pt1[1] - center1[1], pt1[0] - center1[0]) + elif color_by == 1: + coord_angle = np.arctan2(pt2[1] - center2[1], pt2[0] - center2[0]) + + corr_color = np.int32(64 * coord_angle / np.pi) % 128 + color = tuple(colors[corr_color].tolist()) + + if ( + (pt1[0] <= w1 - 1) + and (pt1[0] >= 0) + and (pt1[1] <= h1 - 1) + and (pt1[1] >= 0) + ): + img1 = cv2.circle( + img1, (int(pt1[0]), int(pt1[1])), radius, color, -1, cv2.LINE_AA + ) + + if ( + (pt2[0] <= w2 - 1) + and (pt2[0] >= 0) + and (pt2[1] <= h2 - 1) + and (pt2[1] >= 0) + ): + if mask is not None and mask[i]: + img2 = cv2.drawMarker( + img2, + (int(pt2[0]), int(pt2[1])), + color, + markerType=cv2.MARKER_CROSS, + markerSize=int(5 * radius), + thickness=int(radius / 2), + line_type=cv2.LINE_AA, + ) + else: + img2 = cv2.circle( + img2, (int(pt2[0]), int(pt2[1])), radius, color, -1, cv2.LINE_AA + ) + if row_cat: + whole_img = np.concatenate([img1, img2], axis=0) + else: + whole_img = np.concatenate([img1, img2], axis=1) + return whole_img + if row_cat: + return np.concatenate([img1, img2], axis=0) + return np.concatenate([img1, img2], axis=1) + + +def drawMatches_lines( + img1, img2, kp1, kp2, radius=2, mask=None, color_by=0, row_cat=False +): + + h1, w1 = img1.shape[:2] + h2, w2 = img2.shape[:2] + + img1 = np.ascontiguousarray(float2uint8(img1)) + img2 = np.ascontiguousarray(float2uint8(img2)) + + center1 = np.median(kp1, axis=0) + center2 = np.median(kp2, axis=0) + + set_max = range(128) + colors = {m: i for i, m in enumerate(set_max)} + colors = { + m: (255 * np.array(plt.cm.hsv(i / float(len(colors))))[:3][::-1]).astype( + np.int32 + ) + for m, i in colors.items() + } + + if mask is not None: + ind = np.argsort(mask)[::-1] + kp1 = kp1[ind] + kp2 = kp2[ind] + mask = mask[ind] + + if row_cat: + whole_img = np.concatenate([img1, img2], axis=0) + else: + whole_img = np.concatenate([img1, img2], axis=1) + for i, (pt1, pt2) in enumerate(zip(kp1, kp2)): + if color_by == 0: + coord_angle = np.arctan2(pt1[1] - center1[1], pt1[0] - center1[0]) + elif color_by == 1: + coord_angle = np.arctan2(pt2[1] - center2[1], pt2[0] - center2[0]) + + corr_color = np.int32(64 * coord_angle / np.pi) % 128 + color = tuple(colors[corr_color].tolist()) + rand_val = np.random.rand() + if rand_val < 0.1: + if ( + (pt1[0] <= w1 - 1) + and (pt1[0] >= 0) + and (pt1[1] <= h1 - 1) + and (pt1[1] >= 0) + ) and ( + (pt2[0] <= w2 - 1) + and (pt2[0] >= 0) + and (pt2[1] <= h2 - 1) + and (pt2[1] >= 0) + ): + + whole_img = cv2.circle( + whole_img, + (int(pt1[0]), int(pt1[1])), + radius, + color, + -1, + cv2.LINE_AA, + ) + + if row_cat: + whole_img = cv2.circle( + whole_img, + (int(pt2[0]), int(pt2[1] + h1)), + radius, + color, + -1, + cv2.LINE_AA, + ) + cv2.line( + whole_img, + (int(pt1[0]), int(pt1[1])), + (int(pt2[0]), int(pt2[1] + h1)), + color, + 1, + cv2.LINE_AA, + ) + else: + whole_img = cv2.circle( + whole_img, + (int(pt2[0] + w1), int(pt2[1])), + radius, + color, + -1, + cv2.LINE_AA, + ) + cv2.line( + whole_img, + (int(pt1[0]), int(pt1[1])), + (int(pt2[0] + w1), int(pt2[1])), + color, + 1, + cv2.LINE_AA, + ) + return whole_img + if row_cat: + return np.concatenate([img1, img2], axis=0) + return np.concatenate([img1, img2], axis=1) + + +import torch +import os +import time +import viser + + +def rotation_matrix_to_quaternion(R): + """ + :param R: [3, 3] + :return: [4] + """ + tr = np.trace(R) + Rxx = R[0, 0] + Ryy = R[1, 1] + Rzz = R[2, 2] + q = np.zeros(4) + q[0] = 0.5 * np.sqrt(1 + tr) + q[1] = (R[2, 1] - R[1, 2]) / (4 * q[0]) + q[2] = (R[0, 2] - R[2, 0]) / (4 * q[0]) + q[3] = (R[1, 0] - R[0, 1]) / (4 * q[0]) + return q + + +class PointCloudViewer: + def __init__(self, pc_dir, device="cpu"): + self.server = viser.ViserServer() + self.server.set_up_direction("-y") + self.device = device + self.tt = lambda x: torch.from_numpy(x).float().to(device) + self.pc_dir = pc_dir + self.pcs, self.all_steps = self.read_data() + self.num_frames = len(self.all_steps) + + self.fix_camera = False + self.camera_scale = self.server.add_gui_slider( + "camera_scale", + min=0.01, + max=1.0, + step=0.01, + initial_value=0.1, + ) + + self.camera_handles = [] + + def read_data(self): + pc_list = os.listdir(self.pc_dir) + pc_list.sort(key=lambda x: int(x.split(".")[0].split("_")[-1])) + pcs = {} + step_list = [] + for pc_name in pc_list: + pc = np.load(os.path.join(self.pc_dir, pc_name)) + step = int(pc_name.split(".")[0].split("_")[-1]) + pcs.update({step: {"pc": pc}}) + step_list.append(step) + return pcs, step_list + + def parse_pc_data(self, pc, batch_idx=-1): + idx = batch_idx + ret_dict = {} + for i in range(len(pc.keys()) // 2): + pred_pts = pc[f"pts3d_{i+1}"][idx].reshape(-1, 3) # [N, 3] + color = pc[f"colors_{i+1}"][idx].reshape(-1, 3) # [N, 3] + ret_dict.update({f"pred_pts_{i+1}": pred_pts, f"color_{i+1}": color}) + return ret_dict + + def add_pc(self, step): + pc = self.pcs[step]["pc"] + pc_dict = self.parse_pc_data(pc) + + for i in range(len(pc_dict.keys()) // 2): + self.server.add_point_cloud( + name=f"/frames/{step}/pred_pts_{i+1}_{step}", + points=pc_dict[f"pred_pts_{i+1}"], + colors=pc_dict[f"color_{i+1}"], + point_size=0.002, + ) + + if not self.fix_camera: + raise NotImplementedError + + R21, T21 = find_rigid_alignment_batched( + torch.from_numpy(pc_dict["pred_pts1_2"][None]), + torch.from_numpy(pc_dict["pred_pts1_1"][None]), + ) + R12, T12 = find_rigid_alignment_batched( + torch.from_numpy(pc_dict["pred_pts2_1"][None]), + torch.from_numpy(pc_dict["pred_pts2_2"][None]), + ) + R21 = R21[0].numpy() + T21 = T21.numpy() + R12 = R12[0].numpy() + T12 = T12.numpy() + pred_pts1_2 = pc_dict["pred_pts1_2"] @ R21.T + T21 + pred_pts2_1 = pc_dict["pred_pts2_1"] @ R12.T + T12 + self.server.add_point_cloud( + name=f"/frames/{step}/pred_pts1_2_{step}", + points=pred_pts1_2, + colors=pc_dict["color1_2"], + point_size=0.002, + ) + + self.server.add_point_cloud( + name=f"/frames/{step}/pred_pts2_1_{step}", + points=pred_pts2_1, + colors=pc_dict["color2_1"], + point_size=0.002, + ) + img1 = pc_dict["color1_1"].reshape(224, 224, 3) + img2 = pc_dict["color2_2"].reshape(224, 224, 3) + self.camera_handles.append( + self.server.add_camera_frustum( + name=f"/frames/{step}/camera1_{step}", + fov=2.0 * np.arctan(224.0 / 490.0), + aspect=1.0, + scale=self.camera_scale.value, + color=(1.0, 0, 0), + image=img1, + ) + ) + self.camera_handles.append( + self.server.add_camera_frustum( + name=f"/frames/{step}/camera2_{step}", + fov=2.0 * np.arctan(224.0 / 490.0), + aspect=1.0, + scale=self.camera_scale.value, + color=(0, 0, 1.0), + wxyz=rotation_matrix_to_quaternion(R21), + position=T21, + image=img2, + ) + ) + + def animate(self): + with self.server.add_gui_folder("Playback"): + gui_timestep = self.server.add_gui_slider( + "Train Step", + min=0, + max=self.num_frames - 1, + step=1, + initial_value=0, + disabled=True, + ) + gui_next_frame = self.server.add_gui_button("Next Step", disabled=True) + gui_prev_frame = self.server.add_gui_button("Prev Step", disabled=True) + gui_playing = self.server.add_gui_checkbox("Playing", False) + gui_framerate = self.server.add_gui_slider( + "FPS", min=1, max=60, step=0.1, initial_value=1 + ) + gui_framerate_options = self.server.add_gui_button_group( + "FPS options", ("10", "20", "30", "60") + ) + + @gui_next_frame.on_click + def _(_) -> None: + gui_timestep.value = (gui_timestep.value + 1) % self.num_frames + + @gui_prev_frame.on_click + def _(_) -> None: + gui_timestep.value = (gui_timestep.value - 1) % self.num_frames + + @gui_playing.on_update + def _(_) -> None: + gui_timestep.disabled = gui_playing.value + gui_next_frame.disabled = gui_playing.value + gui_prev_frame.disabled = gui_playing.value + + @gui_framerate_options.on_click + def _(_) -> None: + gui_framerate.value = int(gui_framerate_options.value) + + prev_timestep = gui_timestep.value + + @gui_timestep.on_update + def _(_) -> None: + nonlocal prev_timestep + current_timestep = gui_timestep.value + with self.server.atomic(): + frame_nodes[current_timestep].visible = True + frame_nodes[prev_timestep].visible = False + prev_timestep = current_timestep + self.server.flush() # Optional! + + self.server.add_frame( + "/frames", + show_axes=False, + ) + frame_nodes = [] + for i in range(self.num_frames): + step = self.all_steps[i] + frame_nodes.append( + self.server.add_frame( + f"/frames/{step}", + show_axes=False, + ) + ) + self.add_pc(step) + + for i, frame_node in enumerate(frame_nodes): + + frame_node.visible = i == gui_timestep.value + + prev_timestep = gui_timestep.value + while True: + if gui_playing.value: + gui_timestep.value = (gui_timestep.value + 1) % self.num_frames + for handle in self.camera_handles: + handle.scale = self.camera_scale.value + time.sleep(1.0 / gui_framerate.value) + + def run(self): + self.animate() + while True: + time.sleep(10.0) + + +from sklearn.decomposition import PCA + + +def colorize_feature_map(x): + """ + Args: + x: torch.Tensor, [B, H, W, D] + Returns: + torch.Tensor, [B, H, W, 3] + """ + device = x.device + x = x.cpu().numpy() + + out = [] + for x_ in x: + x_ = colorize_feature_map_np(x_) + out.append(torch.from_numpy(x_).to(device)) + out = torch.stack(out).squeeze(0) + return out + + +def colorize_feature_map_np(x): + """ + Args: + x: np.ndarray, [H, W, D] + """ + pca = PCA(n_components=3) + pca_features = pca.fit_transform(x.reshape(-1, x.shape[-1])) + + pca_features = (pca_features - pca_features.min()) / ( + pca_features.max() - pca_features.min() + ) + pca_features = pca_features.reshape(x.shape[0], x.shape[1], 3) + return pca_features diff --git a/requirements.txt b/requirements.txt index 0565c7c98fb56760c0f8b59adb98f526b832bf62..0322696cf31018a37674560472a567701807e313 100644 --- a/requirements.txt +++ b/requirements.txt @@ -16,4 +16,5 @@ onnxruntime requests trimesh matplotlib -gradio_client \ No newline at end of file +gradio_client +transformers \ No newline at end of file