Skip to content
Book Open access

Asymmetric Privileged Distillation for Weakly Supervised Video Anomaly Detection

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 44 references

Abstract

Recent advances in Large Language Models (LLMs) and vision--language models have introduced richer semantic guidance into Weakly Supervised Video Anomaly Detection (WS-VAD). However, video-dependent text generation or encoding during inference introduces substantial computational overhead and complicates deployment in resource-constrained scenarios. We propose APD-VAD, an Asymmetric Privileged Distillation framework that uses LVLM-generated descriptions as privileged information during training and deploys a visual student without dynamic text processing. A Privileged Language Cascade (PLC) supplies video- and snippet-level descriptions to the teacher, and a Background-Debiased Graph (BDG) attenuates scene-dominated relations before temporal graph reasoning. Semantic-distribution, feature, and topology distillation transfer the resulting class, snippet, and relational knowledge to the student. At inference, a dual-branch visual graph and Visual Prototype Memory (VPM) provide category-aware visual context through cached category prototypes and class-routed visual retrieval. Experiments on UCF-Crime and XD-Violence demonstrate competitive anomaly detection and improved fine-grained temporal localization, while deployment profiling confirms efficient inference without recurring LVLM processing. Code is available at https://github.com/cgao-comp/APD-VAD.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.