TIGER-Lab/HRVideoBench

HRVideoBench是一个用于评估视频大型语言模型对高分辨率视频理解能力的全面基准数据集。它包含200个多项选择题，旨在评估模型对视频中小区域和细微动作的感知和理解。测试视频至少为1080p分辨率，包含10种不同类型的视频，这些视频都是考虑到现实世界的应用而收集的，例如自动驾驶和视频监控。数据集中的问题全部为人工注释，可以分为物体和动作相关任务两大类。

HRVideoBench is a comprehensive benchmark designed to assess the high-resolution video understanding capabilities of video large language models (LMMs). It consists of 200 multiple-choice questions aimed at evaluating the models perception and understanding of small regions and subtle actions within videos. The test videos are at least 1080p resolution and include 10 different types of videos collected with real-world applications in mind, such as autonomous driving and video surveillance. The benchmark features questions that are all manually annotated and can be broadly categorized into object and action-related tasks.