Distilling Privileged Knowledge for Anomalous Event Detection From Weakly Labeled Videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised video anomaly detection (WS-VAD) typically requires multimodal data (RGB, optical flow, audio) for robust performance.
• Multimodal approaches are computationally expensive and storage-intensive for long video sequences.
• Existing methods face limitations in practical applications due to high resource demands.

Purpose of the Study:

• To develop an efficient WS-VAD framework that leverages multimodal information without requiring it during inference.
• To address the computational and storage burdens of traditional multimodal WS-VAD methods.
• To enable effective anomaly detection using only video-level text annotations.

Main Methods:

• Proposed a privileged knowledge distillation (KD) framework for WS-VAD.
• Developed a multimodal teacher network using text cross-modal interactive learning and an anomaly-normal discrimination loss.
• Implemented representation- and text logits-level distillation to transfer knowledge from the teacher to a unimodal student network.
• Utilized a snippet-to-video distillation strategy.

Main Results:

• The privileged KD framework successfully trains lightweight and effective WS-VAD detectors.
• The proposed methods achieved strong performance on three public benchmarks.
• The framework effectively distills privileged knowledge, enabling unimodal inference.

Conclusions:

• Privileged knowledge distillation offers an efficient solution for WS-VAD.
• The proposed framework balances performance with reduced computational and storage requirements.
• This approach facilitates practical deployment of WS-VAD systems.