RTime
收藏资源简介:
RTime数据集是由中国人民大学研究团队创建的一个强调时间理解的视频-文本检索数据集。该数据集包含21,000个视频,每个视频配有10个描述,总计210,000条数据,视频总时长约122小时。数据集通过收集具有显著时间性的视频,并反转这些视频以创建更难的负样本,进一步通过人工标注和GPT-4扩展描述来构建。RTime数据集旨在解决现有视频-文本检索数据集在时间理解能力评估上的不足,特别适用于评估模型在处理时间性更强的视频内容时的表现。该数据集的应用领域包括视频搜索引擎和视频推荐系统等。
The RTime dataset is a video-text retrieval dataset emphasizing temporal understanding, developed by a research team from Renmin University of China. It contains 21,000 videos, each paired with 10 captions, totaling 210,000 data entries, with an overall video duration of approximately 122 hours. The dataset is constructed by collecting videos with prominent temporal attributes, reversing these videos to generate more challenging negative samples, and further expanding the captions through manual annotation and GPT-4. The RTime dataset aims to address the deficiencies of existing video-text retrieval datasets in evaluating temporal understanding capabilities, and is particularly suited for assessing model performance when processing highly temporal video content. Its application scenarios include video search engines, video recommendation systems, and other related fields.




