CLAWS:基于聚类辅助弱监督学习与常态抑制的异常事件检测
计算机视觉与模式识别
2021-08-05 v4 人工智能
摘要
通过视频级标签学习检测真实世界异常事件是一项具有挑战性的任务,原因在于异常事件发生稀少以及标签中存在噪声。在本工作中,我们提出一种弱监督异常检测方法,其具有多重贡献,包括:1)基于随机批次的训练过程以降低批次间相关性;2)常态抑制机制,通过考虑一个训练批次中可用的整体信息,来最小化视频正常区域的异常分数;3)基于聚类距离的损失,通过鼓励我们的模型生成 distinct 的正常与异常聚类,有助于减轻标签噪声并产生更好的异常表征。所提方法在 UCF Crime 和 ShanghaiTech 数据集上分别取得了 83.03% 和 89.67% 的帧级 AUC 性能,证明了其相对于现有最先进算法的优越性。
引用
@article{arxiv.2011.12077,
title = {CLAWS: Clustering Assisted Weakly Supervised Learning with Normalcy Suppression for Anomalous Event Detection},
author = {Muhammad Zaigham Zaheer and Arif Mahmood and Marcella Astrid and Seung-Ik Lee},
journal= {arXiv preprint arXiv:2011.12077},
year = {2021}
}
备注
Presented in the European Conference on Computer Vision ECCV 2020. (Changes from actual paper: 1) Recently published methods have been added in ShanghaiTech and UCF Crime comparison tabs. 2) Due to some error in arxiv compilation, few references are exceeding the paragraph. Also, word 'normalcy' in the title is misspelling despite being correct in the code. (Contents are intact)