ECVA
收藏资源简介:
ECVA数据集是由北京邮电大学移动网络技术国家工程研究中心创建的一个用于视频异常因果理解的综合基准。该数据集包含2240个来自现实世界的视频,涵盖21个主要类别和100个子类别的异常类型,每个视频平均长度为141秒,包含621.1个Tokens。数据集通过人工注释详细描述了异常的类型、原因和影响,旨在解决视频异常检测和定位中的因果关系理解问题。ECVA数据集的创建过程包括从多个视频平台爬取数据,并通过人工注释生成详细的异常描述和因果关系解释。该数据集主要应用于视频异常理解领域,旨在提升模型对视频中异常事件的因果推理能力。
The ECVA dataset is a comprehensive benchmark for video anomaly causal understanding, created by the National Engineering Research Center for Mobile Network Technology at Beijing University of Posts and Telecommunications. It contains 2,240 real-world videos, covering 21 major categories and 100 sub-categories of anomaly types. The average duration of each video is 141 seconds, with 621.1 Tokens per video. The dataset features detailed annotations of anomaly types, causes and impacts via manual annotation, aiming to address the challenge of causal relationship understanding in video anomaly detection and localization. The construction of the ECVA dataset involves crawling data from multiple video platforms and generating detailed anomaly descriptions and causal explanations through manual annotation. This dataset is primarily applied in the field of video anomaly understanding, with the goal of enhancing the causal reasoning capability of models for anomalous events in videos.




