MFF-M3AD: A unified reconstruction method with multi-scale feature fusion for multi-category 3D anomaly detection.

Abstract

3D anomaly detection has evolved as an effective method for identifying anomalies from high-resolution industrial products. However, most existing methods are designed for single-category anomaly detection, and thus multi-class 3D anomaly detection faces major challenges in generalizing across different object categories. In this study, we propose a unified reconstruction model with Multi-scale Feature Fusion (MFF) for Multi-Category 3D Anomaly Detection (MFF-M3AD), which consists of three carefully designed modules. First, the Multi-scale Feature Fusion (MFF) module is introduced to fuse features across multiple scales with cross-scale interaction. Then, the Multi-scale Feature Refinement (MFR) module is designed to capture global and local features to enhance the representation for multiple categories. Finally, the Multi-scale Reconstruction (MSR) module is developed to progressively reconstruct features for anomaly detection. Our method achieves state-of-the-art O-AUROC and P-AUROC performance on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD datasets, surpassing existing multi-class and single-class 3D anomaly detection methods by an average improvement of 5.5% and 2.1%, respectively. All code is available at https://github.com/hzzzzzhappy/MFF-M3AD.