VidDiffBench

该数据集名为VidDiffBench，是一个包含549个视频对的基准数据集，其中有人类标注的4,469个细粒度动作差异以及2,075个定位时间戳，指出了这些差异发生的位置。该数据集旨在识别同一动作视频中微妙的差异。此外，该数据集涵盖了与技能学习和绩效反馈相关的广泛动作，分为五大类：健身、球类运动、跳水、音乐和手术。每个类别都包含具有结构化差异分类的具体动作，以确保标注的一致性。该数据集的任务是识别同一动作视频之间的微妙差异。

The dataset named VidDiffBench is a benchmark dataset comprising 549 video pairs, which contain 4,469 human-annotated fine-grained action differences and 2,075 localization timestamps indicating the locations where these differences occur. This dataset aims to identify subtle differences within videos of the same action. Additionally, it covers a wide range of actions related to skill learning and performance feedback, and is divided into five major categories: fitness, ball sports, diving, music, and surgery. Each category includes specific actions with structured difference classifications to ensure annotation consistency. The task of this dataset is to recognize subtle differences between videos of the same action.