kinetics400-grt-video-masks
收藏资源简介:
Kinetics-400 GRT Video Masks 是一个用于立体视频生成自监督学习的预计算遮挡掩码数据集,为ICML 2026论文《Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation》而发布。该数据集基于Kinetics-400源视频,使用Video Depth Anything Large深度模型生成几何互易性(GRT)遮挡掩码。掩码通过应用相对立体位移(比例为视频宽度的0.06倍)计算得出,以模拟立体视觉中的视差效应。数据集包含总计298,337个视频的掩码序列,具体划分为239,789个训练样本、19,877个验证样本和38,671个测试样本。数据以压缩的.npz格式存储,每个视频对应一个掩码序列,并按2,984个tar分片进行组织,每个分片附带.done标记以指示完整性。掩码文件与原始Kinetics-400视频的文件名保持对应关系(例如,<split>/<relative_stem>.mp4 对应 <split>/<relative_stem>.npz),便于对齐使用。该数据集主要适用于视频修复、单目到立体视频转换、立体视频生成及几何互易性自监督学习等计算机视觉任务。
Kinetics-400 GRT Video Masks is a pre-computed occlusion mask dataset for self-supervised learning in stereoscopic video generation, released alongside the ICML 2026 paper *Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation*. Built upon Kinetics-400 source videos, this dataset provides Geometric Reciprocity (GRT) occlusion masks generated via the Video Depth Anything Large depth model. The masks are calculated by applying relative stereoscopic displacement scaled to 0.06 times the video width to simulate the disparity effect in stereoscopic vision. The dataset contains mask sequences for a total of 298,337 videos, specifically split into 239,789 training samples, 19,877 validation samples, and 38,671 test samples. The data is stored in compressed .npz format, with each video corresponding to one mask sequence, and organized into 2,984 tar shards, each accompanied by a .done marker to indicate completeness. The mask filenames maintain correspondence with the original Kinetics-400 video filenames (e.g., <split>/<relative_stem>.mp4 corresponds to <split>/<relative_stem>.npz), facilitating aligned usage. This dataset is primarily applicable to computer vision tasks including video inpainting, monocular-to-stereoscopic video conversion, stereoscopic video generation, and self-supervised learning for geometric reciprocity.




