Towards Open Set Video Anomaly Detection

AbstractOpen Set Video Anomaly Detection (OpenVAD) aims to identify abnormal events from video data where both known anomalies and novel ones exist in testing. Unsupervised models learned solely normal videos are applicable any testing but suffer a high false positive rate. In contrast, weakly supervised methods effective detecting could fail an open world. We develop method for the OpenVAD problem by integrating evidential deep learning (EDL) normalizing flows (NFs) into multiple instance (MIL) framework. Specifically, we propose use graph neural networks triplet loss learn discriminative features training EDL classifier, is capable of identifying unknown quantifying uncertainty. Moreover, uncertainty-aware selection strategy obtain clean anomaly instances NFs module generate pseudo anomalies. Our superior existing approaches inheriting advantages unsupervised weakly-supervised MIL Experimental results on real-world datasets show effectiveness our method. KeywordsVideo detectionWeakly learningOpen set recognitionNormalizing