Recovering Total Recall: Multi-Scale Feature Fusion for High-Resolution Industrial Anomaly Localization

Abstract

PatchCore, a memory-bank-based anomaly detector, is near-saturated on the original MVTec AD benchmark but deteriorates sharply on the recently released MVTec AD 2, where its mean Area-Under-the-Per-Region-Overlap at 5% FPR (AU-PRO@5%) drops to single digits in our reproduction. We investigate this regression on a high-resolution factory-inspection workload through three systematic methodological studies. First, we falsify a natural hypothesis: re-implementing PatchCore's official overlapping sliding-window tiling aligns train and test preprocessing but leaves localization essentially unchanged, so input resolution is not the primary bottleneck. Second, we trace the failure to the granularity of the feature embedding and restore the multi-scale feature-fusion design of the original PatchCore (combining ResNet-50 layer2 and layer3), which restores localization (mean AU-PRO@5% of 76.35%, compared with 8.87% for the single-layer configuration) at no extra training cost. Third, we report a negative result on a Vision Transformer backbone: a Global Average Pooling step collapses the spatial token grid and degrades localization to a spatially uninformative map, with MoBY pretraining unable to recover what pooling discards. On AD 2's public test split, the fused pipeline exceeds the PatchCore reference numbers that the dataset paper reports for the private split, without additional supervision or architectural changes, and remains within the inference budget measured on a single commodity GPU.