Normal-Adaptive Memory Module for Unsupervised Domain Adaptation in Weakly-Supervised Video Anomaly Detection

Video Anomaly Detection (VAD) effectively identifies abnormal events in surveillance systems, utilizing both temporal intervals and pixel-level analysis. However, due to the domain gap between datasets, current VAD approaches are only effective in the in-domain context. These issues hinder adaptation and generalization across multiple source datasets, leading to inadequate detection of abnormal events. To address these challenges, we propose a Domain-adaptive Memory Module (DAMM) designed to utilize across various domains. This module facilitates the definition of normalcy and allows for easy adaptation through unsupervised methods in unseen domains. Our method demonstrates superior performance in VAD tasks, achieving significant results across two benchmark datasets.