Language-Guided Decision Override for Adaptive and Retraining-Free Video Anomaly Detection


Ryo Moriyama (Aoyama Gakuin University), Shin Suzuki (Aoyama Gakuin University), Naoshi Kaneko (Tokyo Denki University), Kazuhiko Sumi (Aoyama Gakuin University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

One-class video anomaly detection (OC-VAD) detects all events outside the learned distribution as anomalies and requires costly model retraining to redefine what is considered anomalous. Moreover, most existing methods lack the flexibility to incorporate the prior knowledge that human surveillance operators often have about specific anomalies they wish to flag or ignore. To address these challenges, we propose LinGuard (Language-guided anomaly guard), a method that lets users dynamically redefine anomaly/normality judgments at inference time by providing an Event Prompt—a language description of events that should be treated as normal or anomalous—without retraining. LinGuard fuses the anomaly score from a pre-trained OC-VAD backbone with a similarity-based score computed between the Event Prompt and event captions generated by a large vision language model (LVLM). We evaluate LinGuard on standard OC-VAD and general action recognition datasets under challenging conditions where normal/anomalous event set changes during test time. Experiments demonstrate that LinGuard can flexibly redefine user-specified events as normal or anomalous, effectively handling both unknown and known anomalies while eliminating retraining costs. Our code and model are available at https://github.com/MRagusl/LinGuard.

Citation

@inproceedings{Moriyama_2025_BMVC,
author    = {Ryo Moriyama and Shin Suzuki and Naoshi Kaneko and Kazuhiko Sumi},
title     = {Language-Guided Decision Override for Adaptive and Retraining-Free Video Anomaly Detection},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_1030/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection