Niv Cohen, Yedid Hoshen
Nearest neighbor (kNN) methods utilizing deep pre-trained features exhibit very strong anomaly detection performance when applied to entire images. A limitation of kNN methods is the lack of segmentation map describing where the anomaly lies inside the image. In this work we present a novel anomaly segmentation approach based on alignment between an anomalous image and a constant number of the similar normal images. Our method, Semantic Pyramid Anomaly Detection (SPADE) uses correspondences based on a multi-resolution feature pyramid. SPADE is shown to achieve state-of-the-art performance on unsupervised anomaly detection and localization while requiring virtually no training time.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Anomaly Detection | MVTec AD | Detection AUROC | 85.5 | SPADE |
| Anomaly Detection | MVTec AD | FPS | 1.5 | SPADE |
| Anomaly Detection | MVTec AD | Segmentation AUROC | 96.5 | SPADE |
| Anomaly Detection | VisA | Detection AUROC | 82.1 | SPADE |
| Anomaly Detection | VisA | Segmentation AUPRO (until 30% FPR) | 65.9 | SPADE |
| Anomaly Detection | MVTec LOCO AD | Avg. Detection AUROC | 68.9 | SPADE |
| Anomaly Detection | MVTec LOCO AD | Detection AUROC (only logical) | 70.9 | SPADE |
| Anomaly Detection | MVTec LOCO AD | Detection AUROC (only structural) | 66.8 | SPADE |
| Anomaly Detection | MVTec LOCO AD | Segmentation AU-sPRO (until FPR 5%) | 45.1 | SPADE |
| Anomaly Detection | GoodsAD | AUPR | 68.7 | SPADE |
| Anomaly Detection | GoodsAD | AUROC | 64.1 | SPADE |
| 2D Classification | GoodsAD | AUPR | 68.7 | SPADE |
| 2D Classification | GoodsAD | AUROC | 64.1 | SPADE |
| Anomaly Classification | GoodsAD | AUPR | 68.7 | SPADE |
| Anomaly Classification | GoodsAD | AUROC | 64.1 | SPADE |