Dy-MIL: Dynamic Multiple-Instance Learning Framework for Video Anomaly Detection

Anomaly detection is an extremely challenging task in the field of visual understanding because it involves identifying events that deviate significantly from normal patterns. One of the primary reasons for the difficulty of this task is the diversity and complexity of anomalous events. Therefore, it is impossible for us to collect all types of anomalies and label them. In recent work, weakly supervised methods becomes one of the optimal solutions for anomaly detection. Thus, in this paper, we focus on weakly supervised learning and propose a dynamic multiple-instance learning framework for video anomaly detection, which develops a dynamic ranking method combined the k-max-selection scheme to enlarge the inter-class distance between anomalous and normal instances by only using video-level labels. Experimental results demonstrate that our framework achieves superior improvements on three benchmark datasets, including the ShanghaiTech dataset, UCF Crime dataset and NUT dataset.