Anomaly-score calibration under severe class imbalance

Improve anomaly-score calibration for industrial anomaly detection and pixel-level localization under severe class imbalance, where relatively low average precision remains an issue.

Background

NC-TFAD achieves strong AUROC-based detection and localization results on MVTec AD and VisA, but its average precision is comparatively low in settings with substantial normal–anomaly or pixel-level class imbalance. The paper attributes this limitation to false-positive predictions receiving a disproportionate effect on precision–recall performance, particularly for small or sparse defects.

The authors explicitly identify finer anomaly-score calibration as unresolved. Such calibration is intended to suppress high-confidence false positives and improve precision-oriented evaluation without undermining the method’s ability to rank anomalous images or pixels effectively. The issue is presented as relevant both to image-level anomaly detection and to pixel-level localization.

References

Finally, the pixel-level comparison is restricted to methods that natively support anomaly localization, and the relatively low AP observed under severe class imbalance indicates that finer anomaly-score calibration remains an open issue.

— Neural-Collapse-guided Task-Free Continual Anomaly Detection  (2609.03406 - Kong et al., 3 Sep 2026) in Section 5, Limitations

This comparison is qualitative; a quantitative evaluation of the baseline generators as calibration sources under the proxy calibration protocol is left to future work.

— FLASH: A "Generate Once, Synthesize Many" Framework for Synthetic Anomaly Generation in Industrial Anomaly Detection  (2609.37314 - Das et al., 29 Sep 2026) in Supplementary Material, Section F, subsection “Visual qualitative comparison”