FunAILab/TVBench
收藏Hugging Face2025-06-09 更新2025-04-12 收录
下载链接：
https://hf-mirror.com/datasets/FunAILab/TVBench
下载链接
链接失效反馈官方服务：
资源简介：
TVBench是一个专门设计用于评估视频问答中的时间理解的基准。它针对现有数据集的三个主要问题进行了改进，包括静态信息足以解决任务、问题及候选答案的文本信息过于丰富、仅凭世界知识即可回答许多问题。TVBench定义了10个具有挑战性的时间任务，包括动作计数、移动对象属性、时间定位、时间顺序排序和区分时间上困难的动作反义词等。
TVBench is a benchmark specifically created to evaluate temporal understanding in video QA. It addresses three main issues in existing datasets, including static information from single frames being often sufficient to solve tasks, overly informative text in questions and candidate answers, and many questions being answerable by world knowledge alone. TVBench defines 10 temporally challenging tasks such as action counting, properties of moving objects, temporal localization, temporal sequential ordering, and distinguishing between temporally hard action antonyms.
提供机构：
FunAILab
