ACVUBench
收藏资源简介:
ACVUBench是一个专为音频中心视频理解设计的综合基准,包含跨越18个不同领域的2,662个视频,具有丰富的音频信息,以及超过13k的高质量人类注释或验证的问题-答案对。该数据集引入了一系列精心设计的音频中心视频理解任务,全面测试音频内容和音频-视觉交互的理解。数据集分为两部分:AV-Human和AV-Gemini,分别包含人工注释和半自动生成的数据。
ACVUBench is a comprehensive benchmark specifically designed for audio-centric video understanding. It contains 2,662 videos spanning 18 distinct domains, with rich audio information, as well as over 13,000 high-quality human-annotated or human-verified question-answer pairs. This dataset introduces a series of well-designed audio-centric video understanding tasks, which comprehensively evaluate the comprehension of audio content and audio-visual interactions. The dataset is divided into two subsets: AV-Human and AV-Gemini, which respectively contain manually annotated and semi-automatically generated data.




