LVBench
收藏资源简介:
LVBench是由清华大学开发的长视频理解基准数据集,旨在评估模型对长视频内容的处理和理解能力。该数据集包含来自公开来源的多样化长视频,平均时长约为4101秒,远超现有数据集。数据集通过精细的人工标注和多阶段质量控制确保高质量,涵盖体育、直播、电视节目等多种类别。LVBench通过定义六种核心能力,设计了一系列复杂问题,以全面评估模型在长视频理解上的表现,适用于解决如长期决策、深度电影评论和体育直播解说等实际问题。
LVBench is a long-form video understanding benchmark dataset developed by Tsinghua University, which aims to evaluate models' capabilities in processing and comprehending long-form video content. The dataset includes diverse long videos sourced from public platforms, with an average duration of approximately 4101 seconds, far exceeding that of existing datasets. It ensures high data quality through rigorous manual annotation and multi-stage quality control, covering various categories such as sports, live streaming, and TV programs. LVBench defines six core capabilities and designs a series of complex questions to comprehensively evaluate models' performance in long-form video understanding, and is applicable to solving real-world problems including long-term decision-making, in-depth film reviews, and sports live commentary.




