Cognitive Refined Augmentation for Video Anomaly Detection in Weak Supervision.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised video anomaly detection (WSVAD) methods assess anomaly levels in video frames using labeled data.
• A significant challenge in WSVAD is the high rate of false alarms, often due to poor reflection of frame labels during model training.
• Multiple instance learning (MIL) has been explored to address this by identifying distinct features between normal and abnormal segments.

Purpose of the Study:

• To propose a novel multiple instance learning (MIL) framework for weakly supervised video anomaly detection (WSVAD).
• To enhance feature representation and bridge the gap between normal and abnormal video instances.
• To reduce false alarms and improve the accuracy of anomaly detection in videos.

Main Methods:

• A novel MIL framework incorporating a memory unit for feature augmentation.
• Integration of a multi-head attention feature augmentation module.
• A loss function combining KL divergence and Gaussian distribution estimation for improved distance metrics.

Main Results:

• The proposed framework effectively augments features using memory, bridging the gap between normal and abnormal instances.
• The method successfully identifies distinguishable features and secures inter-instance distances.
• Experiments on XD-Violence and UCF-Crime datasets demonstrate the effectiveness of the proposed WSVAD model.

Conclusions:

• The study presents a novel MIL-based framework for WSVAD with an efficient integration strategy for feature augmentation.
• The proposed approach effectively mitigates false alarms by improving feature representation and distance metrics.
• The model shows strong performance on benchmark datasets, validating its efficacy for video anomaly detection.