Researchers expose 'definition blindness' in AI video anomaly detection, propose DeCoS fix

Researchers from the computer vision community have identified a critical flaw in open-world video anomaly detection (OWVAD): models often exhibit 'definition blindness,' meaning they fail to adjust their predictions based on the specific definition of abnormality provided by the user. Current evaluation methods, including standard VAD metrics and dynamic-definition protocols, are dominated by target-versus-normal separation, giving 7.2–26.8× more weight to generic anomaly localization than to definition-conditioned discrimination. As a result, strong baselines can achieve high scores while remaining nearly insensitive to the queried definition, rendering them unreliable for real-world applications where anomaly definitions vary by context.

To address this, the paper introduces three new evaluation metrics—DC-Disc, DC-Det∆, and DC-Sel∆—that progressively remove shortcuts such as normal-frame, generic-anomaly, and multi-event selection biases. Experiments on UCF-Crime, XD-Violence, and MSAD show that even top-performing VAD, OWVAD, and vision-language models have near-zero definition-response margins. The authors also propose DeCoS, a definition-contrastive scoring rule that subtracts anomaly evidence shared across definitions. DeCoS boosts the strongest baseline by 7.3–16.0 AUROC points on DC-Disc and 15.5–28.3 points on DC-Det∆, arguing for a paradigm shift toward definition-conditioned anomaly scoring.