3D-NLM: Voxel-based non-local means for 3D point cloud noise detection and smoothing

Abstract

Point cloud data has been widely used in 3D tasks such as object detection and surface reconstruction, which can be affected by the contained noise and outliers due to sensor limitations. This paper proposes a simple yet effective non-local means 3D point cloud denoising approach (3D-NLM) by leveraging local spatial and global structural information. The proposed method voxelizes the point cloud and formulates the 3D non-local means based on the voxel similarity estimation and an iterative update mechanism to identify noise points. Specifically, 3D-NLM takes the number of points in a voxel as the local spatial feature and utilizes Gaussian Mixture Models (GMM) to estimate the expected number of points from the global structural similarity with other voxel patches. Points that deviate significantly from the majority structure with the updated number of points, characterized using a Euclidean Minimum Spanning Tree (EMST). Consequently, the remaining points are identified as noise and can be directly removed. Alternatively, the intra- and inter-voxel 3D erosion strategy is designed to guide noise points moving towards the underlying surface for smoothing instead of removing. To evaluate the proposed method, we conduct extensive experiments on ModelNet40 and ShapeNet datasets with two types of noise under various metrics. The proposed method consistently outperforms baseline traditional and even learning-based methods, demonstrating it as a promising solution for point cloud denoising.