minsikj/kubric-cotracker-wai

CoTracker3 Kubric, converted to WAI scene shards

A redistribution of facebook/CoTracker3_Kubric (the Kubric data CoTracker3 was trained on), converted to the WAI scene format and packed into one uncompressed tar per sequence for byte-range reading. The images, depth and tracks are the release's own; the changes are listed below. Licensed, like the original, under Apache-2.0.

Content

Path	Content
kubric_cotracker/<seq>.shard-00000.tar	5,869 sequences, 108-147 MB each, 757.5 GB in total.
dataset_metadata/{train,val,test}/kubric_cotracker_scene_list_*.npy	5,805 train and 64 test sequences (the last 64 in sorted order); val is a copy of test.
dataset_metadata/covisibility/kubric_cotracker.covisibility.tar	Pairwise frame covisibility from ground-truth depth, per sequence.
dataset_metadata/wai-storage.json	Storage manifest; frame order natural.
dataset_metadata/kubric_cotracker_conversion_report.json	Conversion policy and per-sequence checks.
SHA256SUMS	Checksums of all files above (it also lists a catalog file that is not uploaded: build the catalog where the tars live).

Each tar holds one 120-frame, 512x512 sequence:

scene_meta.json                        cameras (OpenCV camera-to-world, metric) and intrinsics per frame
images/<seq>-<t:05d>.png               RGB
depth/<seq>-<t:05d>.exr                z-depth in metres
point_tracks/trajs_2d.npy              (120, N, 2) float32 pixel coordinates
point_tracks/trajs_3d.npy              (120, N, 3) float32 world coordinates, metres
point_tracks/visibs.npy                (120, N) bool, visible = not occluded and inside the image
point_tracks/valids.npy                (120, N) bool
point_tracks/instance_ids.npy          (N,) segmentation id the track mostly lies on, 0 = background
covisibility/v0/pairwise_covisibility--120x120.npy

N is up to 32,768 per sequence (median 32,768, minimum 32,610). Arrays are time-major and C-ordered; frame <seq>-<t> is row t of the track arrays.

Changes from the original release

• Converted to the WAI format and packed one tar per sequence.
• 3D tracks are unprojected from the release's pixel coordinates and Euclidean camera distance (coords_depth) with each frame's camera, so depth is z in the camera frame. Points behind the camera, which the release stores mirrored, are un-mirrored; 2D tracks are re-projected from the 3D.
• Visibility is inverted to True = visible (the release stores True = occluded) and restricted to points inside the image.
• Tracks never visible in any frame are dropped.
• Camera poses are the inverse of the release's world-to-camera extrinsics, with rotations orthonormalised.
• A test split (the last 64 sequences) is added; the release has no splits.

Citation

Please cite the original works:

@inproceedings{karaev24cotracker3,
  title     = {CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos},
  author    = {Nikita Karaev and Iurii Makarov and Jianyuan Wang and Natalia Neverova and Andrea Vedaldi and Christian Rupprecht},
  booktitle = {Proc. {arXiv:2410.11831}},
  year      = {2024}
}

@article{greff2021kubric,
    title = {Kubric: a scalable dataset generator},
    author = {Klaus Greff and Francois Belletti and Lucas Beyer and Carl Doersch and
              Yilun Du and Daniel Duckworth and David J Fleet and Dan Gnanapragasam and
              Florian Golemo and Charles Herrmann and Thomas Kipf and Abhijit Kundu and
              Dmitry Lagun and Issam Laradji and Hsueh-Ti (Derek) Liu and Henning Meyer and
              Yishu Miao and Derek Nowrouzezahrai and Cengiz Oztireli and Etienne Pot and
              Noha Radwan and Daniel Rebain and Sara Sabour and Mehdi S. M. Sajjadi and Matan Sela and
              Vincent Sitzmann and Austin Stone and Deqing Sun and Suhani Vora and Ziyu Wang and
              Tianhao Wu and Kwang Moo Yi and Fangcheng Zhong and Andrea Tagliasacchi},
    booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2022},
}