Weakly Supervised Video Anomaly Detection via Self-Guided Temporal Discriminative Transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised video anomaly detection commonly uses multiple instance learning (MIL).
• Existing methods struggle with modeling temporal dynamics and distinguishing between normal and anomalous video segments.
• Lack of discriminative features limits the performance of current anomaly detection systems.

Purpose of the Study:

• To develop a novel weakly supervised temporal discriminative (WSTD) paradigm for video anomaly detection.
• To address the limitations of existing methods in modeling temporal relationships and feature discrimination.
• To improve the accuracy and robustness of anomaly detection in videos.

Main Methods:

• Proposed a transformer-styled temporal feature aggregator (TTFA) to capture temporal relationships across video segments.
• Introduced a self-guided discriminative feature encoder (SDFE) to enhance feature separability.
• Employed clustering of normal snippets and maximization of separability in embedding space within the SDFE.

Main Results:

• The WSTD paradigm demonstrated superior performance compared to state-of-the-art methods.
• Experiments on three public benchmarks validated the effectiveness of the proposed approach.
• The method successfully improved frame-level anomaly score generation under weak supervision.

Conclusions:

• The proposed WSTD paradigm effectively leverages temporal relations and feature discrimination for video anomaly detection.
• The TTFA and SDFE components significantly contribute to mitigating the drawbacks of previous methods.
• WSTD offers a superior approach for weakly supervised video anomaly detection.