Cross-Domain Learning for Video Anomaly Detection with Limited Supervision.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video Anomaly Detection (VAD) is crucial for identifying unusual events in surveillance footage.
• Real-world VAD requires effective cross-domain performance, handling scenarios underrepresented in training data.
• Current unsupervised cross-domain VAD methods exhibit performance limitations.

Purpose of the Study:

• To develop a novel weakly-supervised framework for cross-domain VAD.
• To enhance cross-domain VAD performance by leveraging cost-effective weak supervision and external unlabeled data.
• To address the limitations of existing unsupervised cross-domain VAD approaches.

Main Methods:

• Introduced a weakly-supervised framework for Cross-Domain Learning (CDL) in VAD.
• Incorporated external unlabeled data during training.
• Estimated prediction bias from external data and adaptively minimized it using predicted uncertainty.

Main Results:

• The proposed CDL framework significantly improved cross-domain VAD performance.
• Achieved an average absolute improvement of 19.6% on the UCF-Crime dataset.
• Achieved an average absolute improvement of 12.87% on the XD-Violence dataset, surpassing state-of-the-art methods.

Conclusions:

• Weakly-supervised cross-domain learning with external data offers a promising approach for VAD.
• The proposed method effectively enhances VAD models for real-world, cross-domain applications.
• Demonstrated significant performance gains over existing state-of-the-art techniques in cross-domain evaluations.