HMDB-AD, HMDB-Violence

本研究引入了两个新的数据集HMDB-AD和HMDB-Violence，旨在评估模型对复杂基于动作的异常检测能力。HMDB-AD包含995个视频片段，主要关注正常活动（如行走和跑步）与异常行为（如攀爬和翻筋斗）的区分。HMDB-Violence则是一个更大、更复杂的数据集，将16种动作类别分为7种暴力（异常）和9种非暴力（正常）活动，要求模型在更广泛的背景下分类事件。这两个数据集的创建是为了解决现有基准数据集在异常定义上的局限性，推动视频异常检测领域的研究进展。

This study introduces two novel datasets, HMDB-AD and HMDB-Violence, aimed at evaluating a model's capacity for complex action-based anomaly detection. HMDB-AD contains 995 video clips, focusing primarily on differentiating between normal activities (e.g., walking and running) and abnormal behaviors (e.g., climbing and somersaults). HMDB-Violence, on the other hand, is a larger and more complex dataset that categorizes 16 action classes into 7 violent (abnormal) and 9 non-violent (normal) activities, requiring models to classify events within a broader contextual framework. These two datasets were developed to address the limitations of existing benchmark datasets regarding anomaly definition, and to advance research progress in the field of video anomaly detection.

搜集汇总

背景与挑战

背景概述

HMDB-AD和HMDB-Violence是两个新引入的视频异常检测数据集，旨在评估模型对复杂动作的异常检测能力。HMDB-AD包含995个视频片段，区分正常活动（如行走、跑步）与异常行为（如攀爬、翻筋斗）；HMDB-Violence则更大更复杂，将16种动作分为7种暴力（异常）和9种非暴力（正常）活动，要求模型在更广泛背景下分类事件。这些数据集解决了现有基准在异常定义上的局限性，推动了视频异常检测领域的研究进展。