Few-Shot Anomaly Detection on MVTec AD 2: A Systematic Study and a Strong Frozen Baseline

Abstract

Industrial inspection must detect and localise defects when new production lines provide only a few defect-free reference images. On MVTec AD 2, few-shot evidence remains fragmented across reference budgets, seeds, and metrics, obscuring which design choices matter. We close this gap with the first systematic multi-baseline few-shot benchmark on MVTec AD 2—the anomalib reference implementations of PatchCore, WinCLIP, and AnomalyDINO under one fixed protocol (one, two, four, or eight references; five random draws; all eight categories)—and a strong training-free baseline: patches from a frozen DINOv2 transformer form a memory bank, and each test patch is scored by its distance to the closest stored patch. Our study identifies three empirical findings: two deep layers improve localisation over the common four-layer concatenation; pooling the highest-scoring patches outperforms averaging all patches; and the tested greedy coreset reference strategy degrades pixel-level performance at both evaluated reference budgets. With eight references, our baseline reaches 0.621 pAUPRO@0.30 versus 0.426 for PatchCore, an improvement of 19.5 percentage points (category-clustered bootstrap 95% confidence interval: $[{11.6,27.3}]$ points). Because each method keeps its default backbone and the configuration was selected on the same public split, we report a controlled design study, not a backbone-matched or unseen-split comparison. The configuration transfers unchanged to MVTec AD v1 and MVTec LOCO; code and per-run metrics will be released publicly.