VNBench
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/joez17/videoniah
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为VNBench，专注于通过诸如检索、计数和排序等针尖任务来评估视频中的细节感知能力。它采用标准化的输入长度，即每秒1帧，并使用合成视频样本进行规模扩展，旨在完成视频细节感知评估的任务。
The dataset, designated VNBench, specializes in evaluating fine-grained detail perception in videos through needle-in-a-haystack tasks including retrieval, counting, and sorting. It employs a standardized sampling strategy of 1 frame per second as input, scales its corpus with synthetic video samples, and is purpose-built for video fine-grained detail perception assessment.
