grt-davis-eval-masks
收藏资源简介:
DAVIS GRT 评估掩码数据集是为论文《几何互惠性:解锁立体视频生成的自监督学习(ICML 2026)》发布的预计算评估掩码。该数据集基于 DAVIS 2017 480p 视频源,使用 Video Depth Anything Large 深度模型生成,并按照相对立体位移比例(视频宽度的 0.06 倍)进行处理。数据以压缩的 .npz 格式存储,每个视频对应一个掩码序列,掩码文件名与原始视频文件名(不含扩展名)保持一致。该数据集主要用于立体视频生成、视频修复等任务的评估,支持几何互惠性自监督方法的研究与应用。
The DAVIS GRT Evaluation Mask Dataset is a pre-computed evaluation mask set released for the paper *Geometric Reciprocity: Unlocking Self-Supervised Learning for Stereo Video Generation* (ICML 2026). This dataset is based on the DAVIS 2017 480p video source, generated using the Video Depth Anything Large depth model, and processed according to the relative stereo displacement ratio (0.06 times the video width). The data is stored in compressed .npz format, with each video corresponding to a mask sequence, and the mask filenames are consistent with the original video filenames (without file extensions). This dataset is primarily used for evaluating tasks such as stereo video generation and video inpainting, and supports the research and application of geometric reciprocity-based self-supervised learning methods.




