SWiM: Spacecraft With Masks

A large-scale instance segmentation dataset of nearly 64k annotated spacecraft images created using real spacecraft models, superimposed on a mixture of real and synthetic backgrounds generated using NASA's TTALOS pipeline. To mimic camera distortions and noise in real-world image acquisition, we added different types of noise and distortion.

Dataset Summary

The dataset contains over 63,917 annotated images with instance masks for varied spacecraft. It's structured for YOLO and segmentation applications, and chunked to stay within Hugging Face's per-folder file limits.

Directory Structure Note

Due to Hugging Face Hub's per-directory file limit (10,000 files), this dataset is chunked: each logical split (like train/labels/) is subdivided into folders (000/, 001/, ...) containing no more than 5,000 files each.

Example Structure:

  Baseline
    ├──train/
         ├──images/
              ├── 000/
              │   ├── img_0.png
              │   └── ...
              ├── 001/
              └── ...

If you're using models/tools like YOLO or others that expect a flat directory, kindly use the --flatten argument provided in our utils/download_swim.py script.

YOLO Example Structure:

  Baseline
    ├──train/
         ├──images/
              ├── img_0.png
              ├── ...
              └── imag_99.png

How to Load

Without downloading it entirely, you can stream the SWiM-SpacecraftWithMasks dataset directly using the HuggingFace datasets library. This is particularly useful when working with limited local storage.

Use the following code to load and iterate over the dataset efficiently:


from datasets import load_dataset

dataset = load_dataset(
    "RiceD2KLab/SWiM-SpacecraftWithMasks",
    data_dir="Baseline", # for augmented, data_dir="Augmented"
    streaming=True
)

Note: The directory/folder structure obtained from HuggingFace's load_dataset API includes chunks(eg, 000, 001, etc). Hence, it does not support YOLO training. For YOLO training or CIFAR-10 based directory structure, you can use the use of the utils/download_swim.py script.

How to Download

For local use, if you'd like to either sample a small portion or download the entire dataset to your filesystem, we provide two utility scripts under the utils/ folder:

sample_swim.py for quick sampling from a single chunk
download_swim.py downloads the full dataset, optionally flattening the directory structure.

These scripts let you work offline or run faster experiments by controlling what and how much data you fetch. Furthermore, the download_swim.py helps download the data in a flattened YOLO/CIFAR-10 supported folder structure.

Utility Scripts

The following scripts help you download this dataset. Due to the large nature of the data and the custom directory structure, it is recommended to use the following scripts to either sample or to download the entire dataset. Note, the scripts are in the utils subdirectory.

1. Setup

Create your virtual environment to help manage dependencies and prevent conflicts:

  python -m venv env
  
  source env/bin/activate # On Windows: env\Scripts\activate
  
  pip install -r requirements.txt

2. Sample 500 items from a specific chunk:

This script is helpful for quick local inspection, prototyping, or lightweight evaluation without downloading the full dataset.

Usage:

python3 utils/sample_swim.py --output-dir ./samples --count 100

Arguments:

    --repo-id          Hugging Face dataset repository ID
    --image-subdir     Path to image subdirectory inside the dataset repo
    --label-subdir     Path to corresponding label subdirectory
    --output-dir       Directory to save downloaded files
    --count            Number of samples to download

Example Usage with all args:

  python3 utils/sample_swim.py
  --repo-id RiceD2KLab/SWiM-SpacecraftWithMasks
  --image-subdir Baseline/images/val/000
  --label-subdir Baseline/labels/val/000
  --output-dir ./Sampled-SWiM
  --count 500

3. Download the entire dataset (optionally flatten chunks for YOLO format):

Streams and downloads the full paired dataset (images + label txt files) from a Hugging Face Hub repository. It recursively processes all available chunk subfolders (e.g., '000', '001', ...) under given parent paths.

Features:

Recursively discovers subdirs (chunks) using HfFileSystem
Optionally flattens the directory structure by removing the deepest chunk level
Saves each .png image with its corresponding .txt label

This script can download the complete dataset for model training or offline access.

Usage:

# Download all chunks (flattened/ YOLO format)

python utils/download_swim.py --output-dir ./SWiM --flatten


# Download specific chunks

python3 utils/download_swim.py --chunks 000 001 002

Arguments:

    --repo-id          Hugging Face dataset repository ID
    --images-parent    Parent directory for image chunks (e.g., Baseline/images/train)
    --labels-parent    Parent directory for label chunks (e.g., Baseline/labels/train)
    --output-dir       Where to save the downloaded dataset
    --flatten          Run with "--flatten" to flatten directories. If you want hierarchical chunk folders, omit --flatten
    --chunks           Specific chunk names (e.g., 000 001); omit to download all

Example usage with all args:

  python3 utils/download_swim.py
  --repo-id RiceD2KLab/SWiM-SpacecraftWithMasks
  --images-parent Baseline/images
  --labels-parent Baseline/labels
  --output-dir ./SWiM
  --flatten             # flattens directories

Arguments are all configurable—see --help for details.

Code and Data Generation Pipeline

All dataset generation scripts, preprocessing tools, and model training code are available on GitHub:

GitHub Repository: https://github.com/RiceD2KLab/SWiM

Citation

If you use this dataset, please cite:

@misc{sam2025newdatasetperformancebenchmark, title={A New Dataset and Performance Benchmark for Real-time Spacecraft Segmentation in Onboard Flight Computers}, author={Jeffrey Joan Sam and Janhavi Sathe and Nikhil Chigali and Naman Gupta and Radhey Ruparel and Yicheng Jiang and Janmajay Singh and James W. Berck and Arko Barman}, year={2025}, eprint={2507.10775}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2507.10775}, }