LuJingyi/kinetics400-grt-video-masks

Kinetics-400 GRT Video Masks

Precomputed GRT disocclusion masks for Kinetics-400 videos, released for Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation (ICML 2026).

• Source videos: Kinetics-400
• Depth model: Video Depth Anything Large
• Relative stereo shift scale: 0.06 * video_width
• Format: compressed .npz, one mask sequence per video
• Total masks: 298,337 videos (239,789 train, 19,877 val, 38,671 test)
• Shards: 2,984 tar shards with matching .done markers
• Code: github.com/Visual-AI/GRT
• Project page: visual-ai.github.io/grt
• Paper: arXiv:2607.05354

Masks preserve the source-video relative filename stem: a video at <split>/<relative_stem>.mp4 matches <split>/<relative_stem>.npz.

Citation

@inproceedings{lu2026grt,
  author    = {Jingyi Lu and Kai Han},
  title     = {Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026},
}