BPMF: Bidirectional Prediction by Multi-Scale Features for Multimodal Industrial Anomaly Detection

Multimodal industrial anomaly detection aims to simultaneously localize visual and geometric defects. Existing cross-modal prediction methods often rely solely on global semantic features, neglecting the fine-grained local details essential for accurate defect localization, leading to suboptimal performance, especially for small and multi-scale anomalies. To address this challenge, we propose a multi-scale feature bidirectional prediction network (BPMF) with an asymmetric architecture. The 2D-to-3D branch utilizes target-specific optimization and residual alignment mechanisms, leveraging multi-scale RGB features to simultaneously reconstruct FPFH handcrafted features and PointMAE learned features, encouraging the network to learn comprehensive representations. The 3D-to-2D branch directly reconstructs multi-scale RGB features from the fused point cloud features, providing auxiliary supervision. Additionally, we introduce a non-anomalous region suppression module (NRSM), which refines the initial anomaly map by reducing false positives in non-defective areas. Experiments show that BPMF outperforms state-of-the-art methods on both MVTec 3D-AD and Eyecandies datasets.