WildDet3D Stereo4D Benchmark

3D object detection benchmark with ground-truth stereo depth from Stereo4D.

Images extracted from Stereo4D videos (1 frame per video from the test set, representing 7% of all Stereo4D videos).

Note: The test set is held out for hidden evaluation and is not publicly available.

Download

pip install huggingface_hub

# Download everything
huggingface-cli download weikaih/WildDet3D-Stereo4D-Bench --repo-type dataset --local-dir WildDet3D-Stereo4D-Bench

After downloading, extract the packed archives:

cd WildDet3D-Stereo4D-Bench
tar xzf packed/images.tar.gz
tar xzf packed/depth.tar.gz
tar xzf packed/camera_parameters.tar.gz

Directory Structure

WildDet3D-Stereo4D-Bench/
├── annotations/
│   ├── Stereo4D_val.json
│   ├── Stereo4D_test.json
│   └── Stereo4D_val_class_map.json
├── images/                          # 7,704 images (512x512)
├── depth/                           # 7,704 .npy depth maps (float32, meters)
└── camera_parameters/               # 7,704 .json camera intrinsics

Annotation Format (COCO3D)

Same format as WildDet3D-Data. Each annotation JSON contains:

• images: image metadata with file_path, K (intrinsics), width, height
• annotations: 3D bounding boxes with center_cam, dimensions [W,H,L], R_cam, bbox3D_cam, bbox2D_proj, valid3D
• categories: category list

Depth Format

Each .npy file is a float32 2D array at image resolution (512x512). Values are in meters.

License

CC BY-NC

This dataset is licensed under CC BY-NC. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.