2inf/tapvid-davis

TAP-Vid DAVIS

This repository is a community mirror of the official TAP-Vid DAVIS evaluation release. It contains 30 videos from the DAVIS validation set together with point-track and occlusion annotations.

Files

• tapvid_davis.pkl - complete benchmark data.
• SOURCES.md - attribution and source links for the original videos.

The pickle stores a dictionary keyed by DAVIS video name. Each value contains:

• video: frames as a 4-D uint8 array.
• points: tracks as a float32 array with point, time, and (x, y) axes.
• occluded: visibility labels as a boolean point-by-time array.

import pickle

with open("tapvid_davis.pkl", "rb") as f:
    dataset = pickle.load(f)

Only unpickle files obtained from a trusted source.

Provenance and integrity

Official release: https://storage.googleapis.com/dm-tapnet/tapvid_davis.zip

The release archive was checked against its Google Cloud Storage ETag before extraction:

MD5 (tapvid_davis.zip) = a7e786a1a5d26ba941b7c10bed21c20e
SHA256 (tapvid_davis.pkl) = 8d4ed1b232fd27a78bc192e2e1ce943866ff640ab15e43db3bc7b10258bb48f5

Project documentation and evaluation code: https://github.com/google-deepmind/tapnet/tree/main/tapnet/tapvid

License and attribution

The TAP-Vid annotations are released under CC BY 4.0. The bundled video content remains owned by its respective creators and may have source-specific attribution requirements. See SOURCES.md before redistributing or adapting the videos.

Citation

@inproceedings{doersch2022tapvid,
  author    = {Carl Doersch and Ankush Gupta and Larisa Markeeva and
               Adria Recasens Continente and Kucas Smaira and Yusuf Aytar and
               Joao Carreira and Andrew Zisserman and Yi Yang},
  title     = {TAP-Vid: A Benchmark for Tracking Any Point in a Video},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2022}
}

This is a community mirror and is not an official Google or DAVIS repository.