Generate anomalies from normal: a partial pseudo-anomaly augmented approach for video anomaly detection

Video anomaly detection (VAD) aims to identify unexpected behaviors or objects in videos. Due to the lack of available anomaly samples for training, video anomaly detection is often considered as a one-class classification problem. Specifically, an autoencoder is trained only on normal data, expected to produce large reconstruction errors when detecting anomalies. However, autoencoders can often learn to reconstruct anomalies, leading to detection failures. To address this issue, we introduce a partial appearance-based pseudo-anomaly generation method in training. Through this approach, the autoencoder becomes more sensitive to the differences between normal and anomalous data, resulting in superior anomaly discrimination capability. We validated our approach on three widely adopted datasets, and experimental results validate the effectiveness of our proposed method. Our source code is published on https://github.com/OctCjy/GenerateAnomaliesFromNormal.