UniGRe‐3D: Unified Geometric Reconstruction for Multi‐category 3D Anomaly Detection

Abstract 3D anomaly detection has shown strong promise for industrial quality inspection. However, most existing methods rely on category‐specific training, resulting in high computational and storage costs as well as limited cross‐category generalization. Therefore, unified modeling is more attractive for practical deployment, especially in industrial scenarios where multiple product categories need to be inspected within the same pipeline. Nevertheless, unified modeling remains challenging, as it must jointly support cross‐category shared representation learning and accurate normal pattern reconstruction. To this end, we propose UniGRe‐3D, a unified geometric reconstruction framework for multi‐category 3D anomaly detection. Specifically, to learn unified representations across different categories, we design a geometry encoder that incorporates positional embeddings, multi‐scale feature interaction, and local‐global feature interaction, thereby improving the discriminative ability of the learned features for geometric anomalies. To enhance normal pattern reconstruction, we further introduce a reconstruction decoder that employs position‐guided decoding and hierarchical progressive reconstruction, thereby enabling more accurate localization of anomalous regions via reconstruction discrepancies. We conducted experiments on multiple datasets, in particular, UniGRe‐3D surpasses prior unified methods by 6.9% in object‐level AUROC and 4.0% in point‐level AUROC on Anomaly‐ShapeNet.