DAUP: Enhancing point cloud homogeneity for 3D industrial anomaly detection via density-aware point cloud upsampling

Abstract

The use of 3D information in industrial anomaly detection tasks has been shown to enhance performance by uncovering unseen abnormal patterns in the RGB modality. Despite the focus on detection pipeline design and multimodal fusion schemes in previous approaches, explorations of dataset characteristics were often overlooked. In contrast to RGB images where pixels form regular grids, point clouds intrinsically lack order and exhibit inhomogeneous densities across regions, thereby adversely affecting the feature extraction process. In this work, we propose a learning-based density-aware point cloud upsampling module (DAUP) to address the inhomogeneous problem. A learning-based neural shape function is developed to generate a local representation of the surface for point upsampling purposes. Utilizing the points generated by the neural shape function, we devise a density-aware resampling mechanism aimed at selecting a diverse number of points from varied regions to facilitate adaptive upsampling within regions of varying densities. DAUP can substantially reducing the misclassification rate for off-the-shelf anomaly detection pipelines. Extensive experiments confirm the effectiveness of our upsampling method on the benchmark dataset MVTec 3D-AD. Notably, our method surpasses previous state-of-the-art methods in terms of image-level AUROC based on the feature bank-based anomaly detection pipeline.