M2P-AD: Memory-to-Prototype Learning with Boundary-aware Score Refinement for 3D Anomaly Detection

Existing 3D anomaly detection methods often struggle with two persistent issues: excessive anomaly scores in normal regions due to poor representation of structural data, and high false-positive rates near object boundaries. The authors seek to improve the reliability of unsupervised 3D anomaly detection by better modeling normal data distributions and explicitly accounting for geometric boundary characteristics.

The authors propose Memory-to-Prototype Anomaly Detection (M2P-AD), which consists of two primary innovations. First, the Memory-to-Prototype (M2P) module replaces standard distance-based memory banks with a set of learned prototype embeddings. By clustering normal features and optimizing these prototypes using both Euclidean distance and cosine similarity, the model preserves structural information that is typically lost during coreset subsampling. Second, the authors introduce a Boundary Extraction (BE) module that projects 3D point clouds onto canonical planes to identify boundary regions. A Boundary-aware score refinement (BSR) strategy then recalibrates initial anomaly scores by comparing test patches against a boundary-specific memory bank, effectively filtering out false positives that occur at structural edges.

Evaluations on the Real3D-AD, Anomaly-ShapeNet, and MulSen-AD benchmarks demonstrate that M2P-AD achieves state-of-the-art performance in both object-level and pixel-level anomaly detection. Qualitative results show that the model produces significantly cleaner anomaly maps compared to baseline methods, particularly for objects with intricate geometries or thin structures. By suppressing boundary-induced artifacts and improving the representation of normal structural features, the proposed framework provides a more stable and reliable solution for industrial quality control.