Rethinking Open-World Video Anomaly Detection: Diagnosing Definition Blindness

2026-07-22 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

Open-world video anomaly detection (OWVAD) aims to identify events matching a user-specified definition of abnormality, a stronger requirement than generic anomaly localization. However, current OWVAD evaluation largely fails to isolate this conditional behavior, leading to "definition blindness" where models score high but are insensitive to the queried definition. This failure stems from standard VAD metrics and dynamic-definition protocols overemphasizing target-versus-normal separation, which receives 7.2-26.8x more weight. To address this, new definition-conditioned evaluation metrics, DC-Disc, DC-DetΔ, and DC-SelΔ, are introduced. Experiments on UCF-Crime, XD-Violence, and MSAD show existing baselines localize anomalies but exhibit weak definition following. The paper validates this issue is actionable by introducing DeCoS, a definition-contrastive scoring rule, which improves the strongest baseline by 7.3-16.0 AUROC points on DC-Disc and 15.5-28.3 points on DC-DetΔ.

Key takeaway

For Computer Vision Engineers developing Open-World Video Anomaly Detection (OWVAD) systems, you must move beyond standard VAD metrics. Your evaluation should incorporate definition-conditioned metrics like DC-Disc and DC-DetΔ to accurately assess a model's ability to follow specific anomaly definitions. Consider implementing definition-contrastive scoring rules such as DeCoS to significantly improve definition-following capabilities, achieving up to 28.3 AUROC points higher.

Key insights

Current OWVAD evaluation suffers from "definition blindness," requiring new metrics and a definition-contrastive scoring rule for true conditional anomaly detection.

Principles

• OWVAD requires definition-conditioned anomaly scoring.
• Evaluation must isolate conditional behavior, not just separation.
• Definition blindness arises from imbalanced metric weighting.

Method

DeCoS, a definition-contrastive scoring rule, subtracts anomaly evidence shared across definitions to improve definition following in Open-World Video Anomaly Detection models.

In practice

• Use DC-Disc, DC-DetΔ, DC-SelΔ for OWVAD evaluation.
• Implement DeCoS for improved definition following.
• Test models on UCF-Crime, XD-Violence, MSAD.

Topics

• Open-World Video Anomaly Detection
• Definition Blindness
• Evaluation Metrics
• Computer Vision
• DeCoS
• Anomaly Scoring

Best for: Research Scientist, AI Scientist, Computer Vision Engineer

Related on AIssential

• An Insight on Evaluation Metrics Under the Imbalanced Case of Anomaly Detection — Anomaly detection metric interpretation requires understanding their behavior across varying class imbalance ratios.
• Why Domain Matters: A Preliminary Study of Domain Effects in Underwater Object Detection — A new framework defines underwater domains using measurable image, scene, and acquisition properties for robust object detection evaluation.
• H2VLR: Heterogeneous Hypergraph Vision-Language Reasoning for Few-Shot Anomaly Detection — H2VLR uses hypergraphs for high-order visual-semantic reasoning in few-shot anomaly detection.
• Decodable but Not Detectable: A Leakage Fingerprint for Near-OOD Benchmarks — Benchmark leaks in near-OOD evaluation can invert detector scores, requiring a specific diagnostic to identify contamination.
• Universal Transformation of One-Class Classifiers for Unsupervised Anomaly Detection — A dataset folding method converts one-class classifiers into unsupervised anomaly detectors without modifying their core.
• Rarity-Gated Context Conditioning for Offline Imitation Learning-Based Maritime Anomaly Detection — Explicitly accounting for context rarity reduces false alarms and improves robustness in contextual anomaly detection.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.