Injecting Text Clues for Improving Anomalous Event Detection From Weakly Labeled Videos

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Video anomaly detection (VAD) aims to identify unusual events in videos.
• Weakly supervised VAD (WS-VAD) uses video-level labels, balancing performance and annotation costs.
• Existing WS-VAD methods struggle with false alarms and incomplete localization due to limited snippet-level data.

Purpose of the Study:

• To improve weakly supervised video anomaly detection by incorporating text clues.
• To address detection errors like false alarms and incomplete localization in WS-VAD.
• To propose a novel dual-branch framework for enhanced VAD performance.

Main Methods:

• A dual-branch framework integrating text clues for WS-VAD.
• Text-guided Anomaly Discovering (TAG) branch with hierarchical matching for discriminative snippet identification.
• Anomaly-Conditioned Text Completion (ATC) branch for complete anomaly localization via generative tasks.
• Mutual learning strategy with consistency constraint for cross-branch knowledge sharing.

Main Results:

• The proposed method significantly improves WS-VAD performance.
• Experimental results on two public benchmarks demonstrate superior detection accuracy.
• The dual-branch framework effectively suppresses normal context and enhances anomaly localization.

Conclusions:

• Injecting text clues into WS-VAD via a dual-branch framework is effective.
• The TAG and ATC branches, along with mutual learning, enhance anomaly detection and localization.
• The proposed method offers a promising advancement in weakly supervised video anomaly detection.