FAVOR-Bench
收藏资源简介:
FAVOR-Bench是由复旦大学等多个机构构建的细粒度视频运动理解综合基准,包含1776个经过精心挑选的视频,涵盖日常生活记录、主观视角视频、电视剧和动画等多种类型。该数据集通过半自动化管道构建了8184个挑战性的问题答案对,并提供了开放式的视频描述评估任务。FAVOR-Bench旨在评估模型在细粒度运动理解和描述方面的能力,包含了闭合式的问题回答和开放式任务,如GPT辅助评估和新型无LLM评估框架。
FAVOR-Bench is a comprehensive fine-grained video motion understanding benchmark developed by Fudan University and multiple other institutions. It contains 1776 carefully selected videos covering diverse categories including daily life footage, first-person perspective videos, TV dramas, and animations. A total of 8184 challenging question-answer pairs are constructed via a semi-automated pipeline for this benchmark, and it also provides open-ended video description evaluation tasks. FAVOR-Bench aims to evaluate models' capabilities in fine-grained video motion understanding and description, encompassing closed-ended question answering and open-ended tasks such as GPT-assisted evaluation and a novel LLM-free evaluation framework.




