HourVideo/HourVideo

HourVideo是一个用于长时间视频语言理解的基准数据集，包含500个手动策划的自我中心视频，视频时长从20分钟到120分钟不等，并包含12,976个高质量的五项选择题。该数据集的任务包括总结、感知（回忆、跟踪）、视觉推理（空间、时间、预测、因果、反事实）和导航（房间到房间、对象检索）等。HourVideo旨在推动能够真正理解无尽视觉数据流的高级多模态模型的发展。

HourVideo is a comprehensive benchmark dataset for long-duration video-language understanding, consisting of 500 manually curated egocentric videos ranging from 20 to 120 minutes in duration. The dataset features a variety of tasks including summarization, perception, visual reasoning, and navigation, and includes 12,976 high-quality five-way multiple-choice questions, aimed at spurring the development of advanced multimodal models.