Cub3: A Large-Scale Covisibility Dataset

Cub3 is a large-scale dataset comprising 5 million image pairs with dense, pixel-level covisibility annotations derived from the autonomous driving nuScenes dataset and the indoor ScanNet dataset. It was introduced alongside Alligat0R (arXiv:2503.07561), a NeurIPS 2025 Spotlight paper.

Covisibility Classes

Each pixel in a view is classified into one of three categories with respect to the other view:

Class	Label	Description
Covisible	0 (pixel value 255)	The 3D point is visible in both views
Occluded	1 (pixel value 128)	The 3D point is occluded in the other view
Outside FOV	2 (pixel value 0)	The 3D point is outside the field of view of the other view

Dataset Variants

Split	Overlap threshold	nuScenes pairs	ScanNet pairs	Total
Cub3-50	>= 50%	2.5M	2.5M	5M
Cub3-all	>= 5%	2.5M	2.5M	5M

Repository Structure

├── nuscenes/
│   ├── train_samples_50.json        # Pair metadata for Cub3-50
│   ├── train_samples_all.json       # Pair metadata for Cub3-all
│   ├── train_seg_50/                # Segmentation masks (10 shards)
│   │   ├── scene-00.tar.gz
│   │   └── ...
│   └── train_seg_all/               # Segmentation masks (10 shards)
│       ├── scene-00.tar.gz
│       └── ...
└── scannet/
    ├── train_samples_50.json
    ├── train_samples_all.json
    ├── scene00.tar.gz               # Segmentation masks (8 shards)
    └── ...

Pair Metadata Format

nuScenes (train_samples_{50,all}.json)

Each entry is a list: [img1_name, img2_name, scene, relative_pose_4x4, overlap, angle, dist_ratio]

• img1_name / img2_name: nuScenes sweep filenames (e.g. n008-2018-...CAM_FRONT__153565...jpg)
• scene: nuScenes scene id (e.g. scene-0527)
• relative_pose_4x4: 4x4 relative pose matrix from camera 1 to camera 2
• overlap: fraction of covisible pixels
• angle: viewpoint angle between the two cameras (degrees)
• dist_ratio: scale ratio between the two views

ScanNet (train_samples_{50,all}.json)

Each entry is a list: [scene, img1_name, img2_name, overlap, angle, dist_ratio, relative_pose_4x4]

Segmentation Mask Format

Masks are stored as grayscale PNG images named <IMG1>+<IMG2>.png:

• Pixel value 255 = covisible
• Pixel value 128 = occluded
• Pixel value 0 = outside field of view
• Pixel value 42 (ScanNet only) = undefined / ignored during training

Download and Setup

pip install huggingface_hub[cli]

# Download metadata
huggingface-cli download thibautloiseau/Cub3 --include "nuscenes/train_samples_*.json" --local-dir data/Cub3
huggingface-cli download thibautloiseau/Cub3 --include "scannet/train_samples_*.json" --local-dir data/Cub3

# Download segmentation masks
huggingface-cli download thibautloiseau/Cub3 --include "nuscenes/train_seg_50/*.tar.gz" --local-dir data/Cub3
huggingface-cli download thibautloiseau/Cub3 --include "nuscenes/train_seg_all/*.tar.gz" --local-dir data/Cub3
huggingface-cli download thibautloiseau/Cub3 --include "scannet/*.tar.gz" --local-dir data/Cub3

# Extract archives
cd data/Cub3/nuscenes/train_seg_50 && for f in *.tar.gz; do tar xzf "$f"; done && cd -
cd data/Cub3/nuscenes/train_seg_all && for f in *.tar.gz; do tar xzf "$f"; done && cd -
cd data/Cub3/scannet && for f in scene*.tar.gz; do tar xzf "$f"; done && cd -

Expected on-disk layout after extraction:

data/Cub3/
├── nuscenes/
│   ├── train_samples_50.json
│   ├── train_samples_all.json
│   ├── train_seg_50/scene-XXXX/seg_aligned/<IMG1>+<IMG2>.png
│   └── train_seg_all/scene-XXXX/seg_aligned/<IMG1>+<IMG2>.png
└── scannet/
    ├── train_samples_50.json
    ├── train_samples_all.json
    └── sceneXXXX_XX/segs/<IMG1>+<IMG2>.png

Important Notes

• Raw images are NOT included. You must obtain the original images from nuScenes and ScanNet under their respective licenses.
• The nuScenes annotations are generated using monocular depth predictions (UniDepth) and surface normals (Depth Anything V2) combined with COLMAP poses, following the pipeline from RUBIK. Some annotations may contain noise, particularly the distinction between covisible and occluded pixels.
• The ScanNet annotations use ground-truth depth maps and camera poses.

Usage with Alligat0R

See the Alligat0R GitHub repository for training scripts that consume this dataset directly, and demo.py for a quick covisibility visualization example.

Citation

@article{loiseau2026alligat0r,
  title={Alligat0r: Pre-training through covisibility segmentation for relative camera pose regression},
  author={Loiseau, Thibaut and Bourmaud, Guillaume and Lepetit, Vincent},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  pages={13762--13789},
  year={2026}
}

License

The Cub3 annotations are released under the MIT License. The underlying images from nuScenes and ScanNet are subject to their own licenses.