MotionBench
收藏资源简介:
MotionBench是由清华大学和Zhipu AI联合创建的一个视频理解基准测试数据集,旨在评估视频语言模型在细粒度运动理解上的能力。该数据集包含8052个问题,数据来源于网络视频(如Panda70M、Pexels)、公共数据集(如MedVid、SportsSloMo、Ha-ViD)以及通过Unity生成的自合成视频,涵盖了广泛的现实世界应用场景。数据集通过精心设计的注释流程确保了多样性和准确性,主要用于视频理解模型的开发和评估,特别是在细粒度运动分析领域。
MotionBench is a video understanding benchmark dataset jointly created by Tsinghua University and Zhipu AI, which aims to evaluate the fine-grained motion understanding capabilities of video-language models. This dataset contains 8052 questions, with data sourced from online videos (e.g., Panda70M, Pexels), public datasets (e.g., MedVid, SportsSloMo, Ha-ViD), and self-synthesized videos generated via Unity, covering a wide range of real-world application scenarios. The dataset ensures diversity and accuracy through a meticulously designed annotation pipeline, and is primarily used for the development and evaluation of video understanding models, particularly in the field of fine-grained motion analysis.




