Automated detection of surface defects on three-dimensional (3D) parts is vital for ensuring product quality and safety in manufacturing. However, three key challenges hinder reliable detection: geometric context ambiguity across complex part shapes, domain mismatch between generic pretrained features and industrial scans (with their unique noise and reflectivity), and the scarcity of diverse defect examples for training. To overcome these issues, we propose a novel single-forward-pass framework for point cloud anomaly detection, comprising three new modules: (1) Spatial Context Aggregation, which grounds each local patch in a set of learned global prototypes via an optimal-transport alignment to resolve context ambiguity; (2) Feature Adaptor, a lightweight two-layer multilayer perceptron (MLP) that fine-tunes self-supervised Point-MAE embeddings to the specific characteristics of industrial scans; and (3) Selective Anomalous Feature Generator, which synthesizes realistic hard negatives by corrupting random subsets of feature tokens, thus mitigating the need for extensive defect labels. An attention-based discriminator trained with patch-wise supervision learns to distinguish these hard negatives from genuine defect-free patterns. At inference, our pipeline delivers dense per-point anomaly scores in a single pass at up to 13.5 frames per second (FPS). On the Real3D-AD benchmark, we observe point-level improvements of 2.8% in area under the receiver operating characteristic curve (AUROC) and 5.7% in area under the precision-recall curve (AUPR), with object-level gains of 3.0% (AUROC) and 3.5% (AUPR). Evaluated on our newly released Industrial3D-AD dataset, which captures realistic sensor noise and reflective materials, we see similar enhancements (2.9%/5.3% point-level, 2.8%/3.3% object-level).
Hoang et al. (2026) studied this question.