ImageNet3D

We present ImageNet3D, a large dataset for general-purpose object-level 3D understanding. ImageNet3D augments 200 categories from the ImageNet dataset with 2D bounding box, 3D pose, 3D location annotations, and image captions interleaved with 3D information.

Refer to github.com/wufeim/imagenet3d for the full documentation and sample preprocessing code for ImageNet3D.

Download Data

ImageNet3D-v1.0: Directly download from the HuggingFace WebUI, or on a server, run

wget https://huggingface.co/datasets/ccvl/ImageNet3D/resolve/main/imagenet3d_v1.zip

Future updates: We are working on annotating more object categories and improving the quality of current annotations. The next update is planned to be released by the end of Jan 2025. Please let us know if you have any suggestions for future updates.

Example Usage

from PIL import Image
import numpy as np

img_path = 'imagenet3d/bed/n02818832_13.JPEG'
annot_path = 'imagenet3d/bed/n02818832_13.npz'

img = np.array(Image.open(img_path).convert('RGB'))
annot = dict(np.load(annot_path, allow_pickle=True))['annotations']

# Number of objects
num_objects = len(annot)

# Annotation of the first object
azimuth = annot[0]['azimuth']  # float, [0, 2*pi]
elevation = annot[0]['elevation']  # float, [0, 2*pi]
theta = annot[0]['theta']  # float, [0, 2*pi]
cad_index = annot[0]['cad_index']  # int
distance = annot[0]['distance']  # float
viewport = annot[0]['viewport']  # int
img_height = annot[0]['height']  # numpy.uint16
img_width = annot[0]['width']  # numpy.uint16
bbox = annot[0]['bbox']  # numpy.ndarray, (x1, y1, x2, y2)
category = annot[0]['class']  # str
principal_x = annot[0]['px']  # float
principal_y = annot[0]['py']  # float

# label indicating the quality of the object, occluded or low quality
object_status = annot[0]['object_status']  # str, one of ('status_good', 'status_partially', 'status_barely', 'status_bad')

# label indicating if multiple objects from same category very close to each other
dense = annot[0]['dense']  # str, one of ('dense_yes', 'dense_no')