GBA-AD: geometry- and boundary-aware unsupervised anomaly detection for high-resolution 3D point clouds

Abstract

Unsupervised 3D point-cloud anomaly detection trains exclusively on normal samples and must identify anomalous objects and localize anomalous regions at test time. Although high-resolution, complete point clouds encode fine-grained geometric detail, defects typically affect only a small fraction of the object surface; moreover, normal local shape variability and discontinuities at boundaries or from uneven sampling frequently trigger spurious anomaly responses. We introduce GBA-AD, a geometry- and boundary-aware framework for anomaly detection on complete point clouds. GBA-AD augments PointMAE patch features with local PCA curvature, enabling the nearest-neighbor anomaly distance to directly reflect surface shape changes such as protrusions, depressions, and cuts. Full-Split augments the normal reference bank using plane-partitioned subsets extracted from normal training point clouds, diversifying the local contexts represented in memory. At test time, it combines anomaly scores from the curvature-enhanced PointMAE branch and the XYZ-coordinate branch, and applies boundary score suppression exclusively to the PointMAE branch to reduce spurious boundary responses caused by patch-score back-projection. On Real3D-AD, GBA-AD achieves an object-level AUROC of 0.8177 ± 0.0033 and a point-level AUROC of 0.7721 ± 0.0072 under train-only leave-one-out calibration, reported as mean ± standard deviation across three random seeds. Sequential ablations quantify the contributions of curvature enhancement, the Full-Split memory bank, and boundary suppression.