Dynamic Erasing Network With Adaptive Temporal Modeling for Weakly Supervised Video Anomaly Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised video anomaly detection uses video-level labels.
• Existing methods struggle with anomaly duration and completeness.
• Prior work often overlooks subtle anomalies by focusing on the most abnormal segments.

Purpose of the Study:

• To develop a novel network for weakly supervised video anomaly detection.
• To address limitations in temporal modeling of anomaly complexity and duration.
• To improve the detection of complete and subtle anomalies.

Main Methods:

• Proposed a dynamic erasing network (DE-Net).
• Introduced adaptive temporal modeling (ATM) for video-specific feature selection and aggregation.
• Implemented a dynamic erasing (DE) strategy to identify and erase prominent anomalies, encouraging discovery of gentler anomalies.

Main Results:

• The DE-Net achieved favorable performance on benchmark datasets.
• The method demonstrated effectiveness in handling varying anomaly durations.
• The dynamic erasing strategy successfully encouraged the discovery of less prominent abnormal segments.

Conclusions:

• The proposed DE-Net effectively enhances weakly supervised video anomaly detection.
• Adaptive temporal modeling and dynamic erasing are crucial for comprehensive anomaly detection.
• The approach shows significant improvements over state-of-the-art methods on XD-Violence, TAD, and UCF-Crime datasets.