VidCapBench
收藏资源简介:
VidCapBench是一个针对可控文本到视频(T2V)生成的视频字幕评估方案,由中国科学院自动化研究所等机构创建。该数据集包含643个经过丰富标注的视频片段,这些视频片段与关键信息相关,如视频美学、内容、动作和物理定律。VidCapBench将关键信息属性分为可自动评估和手动评估的子集,以满足敏捷开发和彻底验证的需求。数据集适用于T2V模型训练的评估。
VidCapBench is a video captioning evaluation benchmark for controllable text-to-video (T2V) generation, developed by the Institute of Automation, Chinese Academy of Sciences and other institutions. This dataset contains 643 richly annotated video clips, each of which is associated with key attributes such as video aesthetics, content, actions, and physical laws. VidCapBench categorizes these key attribute sets into two subsets: automatically evaluable and manually evaluable, to meet the requirements of agile development and rigorous validation. This dataset is suitable for the evaluation of T2V model training.




