3D Point Cloud Denoising Using a Joint Geometry and Color k-NN Graph

Many point cloud acquisition methods, e.g. multi-viewpoint image stereo matching and acquisition of depth data from active light sensors, suffer from significant geometry noise in the data. In the existing literature, denoising of this geometry noise has been performed using only geometry information. In this paper, based on the notion that color attributes are correlated with the geometry, we propose a novel geometry denoising technique that takes advantage of this correlation via a graph-based optimization process. In particular, we construct a graph based on both color and geometry information, and use it for graph-based Tikhonov regularization. Results on synthetic and real-world point clouds show that the proposed denoising method significantly outperforms existing geometry-only techniques.