videophysics/videophy2_test

VIDEOPHY2数据集是一个包含视频和与之相关的文本提示的多模态数据集。它用于视频分类任务，特别关注物理规则的一致性。数据集中的每个视频都由一个原始提示生成，并可能有一个上采样版本的提示。人类评估者对每个视频的语义一致性和物理常识进行评分，并记录了视频遵守或违反的物理规则。此外，数据集还标注了动作类型、物理活动或对象交互类别，并区分了容易和难题实例。数据集还提供了物理规则与其物理定律之间的映射信息。

The VIDEOPHY2 dataset is a multimodal dataset containing videos and associated text prompts. It is designed for video classification tasks with a focus on the consistency of physical rules. Each video in the dataset is generated from an original prompt and may have an upsampled version of the prompt. Human evaluators score the semantic adherence and physical commonsense of each video, and record the physical rules followed or violated by the videos. Additionally, the dataset is annotated with the type of action, category of physical activity or object interaction, and distinguishes between easy and hard instances. The dataset also provides a mapping between physical rules and their corresponding physical laws.