Hierarchical Point-Patch Fusion with Adaptive Patch Codebook for 3D Shape Anomaly Detection

2026-04-05

AI Summary

Existing methods for 3D shape anomaly detection exhibit limited generalization, struggling with diverse anomaly types and scales while remaining sensitive to noise or partial point loss. This work proposes a hierarchical point-patch fusion network that leverages self-supervised decomposition for adaptive patch partitioning and jointly models local point and regional part features. Coupled with a codebook-based anomaly scoring mechanism, the approach effectively captures both local structural irregularities and global geometric deviations. The method substantially improves cross-type and cross-scale generalization, achieving absolute gains of 7% and 4% in object-level AUC on Real3D-AD and Anomaly-ShapeNet, respectively, and surpassing prior art by over 40% in point-level AUC-PR on industrial out-of-distribution categories.

Abstract

3D shape anomaly detection is a crucial task for industrial inspection and geometric analysis. Existing deep learning approaches typically learn representations of normal shapes and identify anomalies via out-of-distribution feature detection or decoder-based reconstruction. They often fail to generalize across diverse anomaly types and scales, such as global geometric errors (e.g., planar shifts, angle misalignments), and are sensitive to noisy or incomplete local points during training. To address these limitations, we propose a hierarchical point-patch anomaly scoring network that jointly models regional part features and local point features for robust anomaly reasoning. An adaptive patchification module integrates self-supervised decomposition to capture complex structural deviations. Beyond evaluations on public benchmarks (Anomaly-ShapeNet and Real3D-AD), we release an industrial test set with real CAD models exhibiting planar, angular, and structural defects. Experiments on public and industrial datasets show superior AUC-ROC and AUC-PR performance, including over 40% point-level improvement on the new industrial anomaly type and average object-level gains of 7% on Real3D-AD and 4% on Anomaly-ShapeNet, demonstrating strong robustness and generalization.

Problem

• 3D shape anomaly detection
• generalization
• geometric errors
• noisy point clouds
• anomaly types

Innovation

• hierarchical point-patch fusion
• adaptive patch codebook
• 3D shape anomaly detection
• self-supervised decomposition
• industrial defect detection