FineAction

FineAction是一个大规模的细粒度视频数据集，专为时间动作定位设计。该数据集由中国科学院深圳先进技术研究院创建，包含106个动作类别和103,324个时间实例，分布在17,000个未修剪的视频中。数据集的特点是细粒度的动作类别、多标签和密集注释，以及不同类别的动作同时发生。创建过程包括定义新的分类法、从现有基准和互联网收集视频，并开发定制的注释工具。FineAction的应用领域包括视频内容分析、智能监控和人类计算机交互，旨在解决时间动作定位的挑战。

FineAction is a large-scale fine-grained video dataset specifically designed for temporal action localization. It was developed by the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, and contains 106 action categories and 103,324 temporal instances distributed across 17,000 untrimmed videos. The dataset features fine-grained action categories, multi-label and dense annotations, as well as concurrent actions across different categories. Its creation process involves defining a novel taxonomy, collecting videos from existing benchmarks and the Internet, and developing customized annotation tools. Its application fields include video content analysis, intelligent surveillance, and human-computer interaction, aiming to address the challenges in temporal action localization.