Multi-level 3DCNN with Min-Max Ranking Loss for Weakly-Supervised Video Anomaly Detection

Snehashis Majhi, Deepak Ranjan Nayak, Ratnakar Dash, Pankaj Kumar Sa

Abstract: Video anomaly detection in real-world surveillance systems is challenging due to the unavailability of large annotated data, visual challenges like partial occlusion and illumination change, and the untrimmed nature of videos. In this paper, we propose a method that mitigates the above challenges. The proposed method adopts a weakly-supervised learning paradigm to address the scarcity of temporally annotated data. In this, only video-level supervision is required for learning, but precise temporal locations of anomalies are detected during testing. To effectively learn from weak supervision, a Min-Max ranking loss is proposed with the objective to maximize the margin of separation between anomaly and normal instances and to minimize the separation among the normal instances simultaneously. Further, to handle the visual challenges in real-world scenarios, a multi-level feature combination strategy from 3DCNN is proposed to extract the fine lower-level representation of the input video sequences. An efficient temporal dependency encoding is utilized further to capture the sharp change in untrimmed surveillance videos. The proposed method is evaluated on a widely used benchmark anomaly detection dataset, UCF-Crime. The results demonstrate that the proposed method achieves competitive performance compared to recently reported anomaly detection methods.