Graph-Based Point Cloud Denoising

Abstract

3D Point cloud data has attracted attention in various applications such as free-view rendering, heritage reconstruction and navigation. However, point clouds often suffer from noise, either from hardware or software causes. We propose an efficient point cloud denoising approach, where the geometry of the point cloud is naturally represented on graphs. We first divide noise in the point cloud into two categories: outlier and surface noise according to the distribution, and then remove them separately. Outliers are firstly removed based on the sparsity of the neighborhood. Next, we formulate the surface noise removal as an optimization problem regularized by graph-signal smoothness prior, which essentially tries to reconstruct the underlying geometry of the point cloud. Experimental results show that our approach significantly outperforms five competing methods.