Learning Causal Temporal Relation and Feature Discrimination for Anomaly Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised anomaly detection lacks frame-level labels during training.
• Existing methods often overlook temporal cues and feature discrimination, relying solely on current information.

Purpose of the Study:

• To propose a novel method for weakly supervised anomaly detection that leverages temporal context and feature discrimination.
• To address the limitations of current approaches by incorporating historical observations and enhancing feature separability.

Main Methods:

• A four-module method is proposed: Causal Temporal Relation (CTR) for temporal dependencies, Classifier (CL) using causal convolution, Compactness (CP) for intraclass feature compactness, and Dispersion (DP) for interclass feature dispersion.
• The CTR module captures local-range temporal dependencies to enhance features.
• The CP and DP modules aim to learn discriminative features by ensuring intraclass compactness and interclass dispersion.

Main Results:

• The proposed method demonstrates the significance of causal temporal relations and feature discrimination in anomaly detection.
• Extensive experiments on three public benchmarks show the superiority of the proposed method over existing approaches.

Conclusions:

• Incorporating temporal context and feature discrimination significantly improves weakly supervised anomaly detection.
• The proposed four-module method effectively enhances feature representation and achieves superior performance on benchmark datasets.