MCG-NJU/CaReBench
收藏Hugging Face2025-07-16 更新2025-04-12 收录
下载链接：
https://hf-mirror.com/datasets/MCG-NJU/CaReBench
下载链接
链接失效反馈官方服务：
资源简介：
CaReBench是一个包含1000个高质量视频的细粒度基准数据集，每个视频都伴有详细的人工标注描述，这些描述分为独立的空间和时间描述，用于评估视频语言模型中的时空偏差。数据集为视频检索和标注任务提供了ReBias和CapST指标，并包括一个表现优异的统一基线模型CaRe。
CaReBench is a fine-grained benchmark dataset consisting of 1,000 high-quality videos, each with detailed human-annotated captions that are manually separated into spatial and temporal descriptions for evaluating spatiotemporal bias in video-language models. The dataset provides ReBias and CapST metrics for video retrieval and captioning tasks and includes a competitive unified baseline model called CaRe.
提供机构：
MCG-NJU
