MIL Based Anomaly Detection in Surveillance Videos Using I3D and Clip Feature Extractor

This work presents an efficient anomaly detection method for surveillance videos using I3D and CLIP feature extractors. By leveraging a deep Multiple Instance Learning (MIL) framework with weakly labeled data, our approach avoids the need for time-consuming temporal annotations. The model is trained on the UCF Crime dataset, containing various real-world anomalies. Our method significantly outperforms previous approaches, achieving an AUC of 84.07 with I3D feature extractor, demonstrating its effectiveness in identifying anomalous events. This provides a practical and scalable solution for real-world surveillance applications, requiring minimal annotation effort.