Reducing Prompt Sensitivity in Video Anomaly Understanding via Conditional Prompt Learning
Abstract
Video Anomaly Understanding (VAU) requires not only detecting irregular events but also providing interpretable explanations. Recent MLLM-based approaches show promise but face two limitations: high sensitivity to prompt variations and poor generalization to unseen anomaly types. We propose AnoCoCo (Anomaly Conditional Context), a prompt learning method that appends learnable instance-conditional context tokens to video-based MLLM prompts. A lightweight conditioning network generates per-instance tokens from pooled spatio-temporal video features, steering the frozen LLM toward anomalous visual cues without modifying pretrained weights. Experiments demonstrate improved prompt robustness and generalization to unseen anomaly types compared to both zero-shot and finetuned baselines, with applicability to safety-critical settings such as public surveillance, smart city monitoring, and industrial anomaly detection.
Authors 3
-
Affiliation as printed
Leiden Institute of Advanced Computer Science, Leiden University, Leiden, Zuid-Holland, Netherlands
-
Affiliation as printed
Leiden Institute of Advanced Computer Science, Leiden University, Leiden, Zuid-Holland, Netherlands
-
Affiliation as printed
Leiden Institute of Advanced Computer Science, Leiden University, Leiden, Zuid-Holland, Netherlands
Cited by 0 stored of 0
No patents citing this paper on Lens.org (checked 2026-10-11).