Abstract
Industrial quality inspection is critical for manufacturing. Single modality methods using 3D point clouds or 2D images miss complementary cues, and existing multimodal fusion struggles to integrate unimodal scores. This paper proposes Geometry-Guided Score Fusion (G2SF), which interprets memory bank anomaly scores as local distances and learns an anisotropic metric for fusion. A Local Scale Prediction Network provides direction-aware scaling, and the loss and aggregation design improve discrimination. Experiments on MVTec-3D AD and Eyecandies show state-of-the-art performance.
Juan Du
Assistant Professor
My research interests include knowledge-infused data science for quality improvement, industrial data analytics and machine learning, and system informatics and control for manufacturing applications.