BatchNorm-based Weakly Supervised Video Anomaly Detection

In weakly supervised video anomaly detection (WVAD), where only video-levellabels indicating the presence or absence of abnormal events are available, theprimary challenge arises from the inherent ambiguity in temporal annotations ofabnormal occurrences. Inspired by the statistical insight that temporalfeatures of abnormal events often exhibit outlier characteristics, we propose anovel method, BN-WVAD, which incorporates BatchNorm into WVAD. In the proposedBN-WVAD, we leverage the Divergence of Feature from Mean vector (DFM) ofBatchNorm as a reliable abnormality criterion to discern potential abnormalsnippets in abnormal videos. The proposed DFM criterion is also discriminativefor anomaly recognition and more resilient to label noise, serving as theadditional anomaly score to amend the prediction of the anomaly classifier thatis susceptible to noisy labels. Moreover, a batch-level selection strategy isdevised to filter more abnormal snippets in videos where more abnormal eventsoccur. The proposed BN-WVAD model demonstrates state-of-the-art performance onUCF-Crime with an AUC of 87.24%, and XD-Violence, where AP reaches up to84.93%. Our code implementation is accessible athttps://github.com/cool-xuan/BN-WVAD.