3D point cloud denoising via Gaussian processes regression

Accurate three-dimensional reconstruction of complex scenes is paramount to further advancements in shape modeling, autonomy, and computer vision. A key aspect is noise removal, whereas other challenges include limited view angles, noise interference, and the presence of outliers. Outliers and noise can significantly compromise 3D reconstructed models and jeopardize downstream applications, especially those requiring near-real-time processing like proximity operations and autonomous driving. In this study, a novel method for point cloud denoising and outlier detection, Stochastic Regression for Denoising with Random Sampling (STORED), is presented. STORED integrates concepts from Random Sample Consensus and Gaussian Process Regression to address prevalent challenges in existing algorithms. By iteratively scoring outliers based on prediction confidence, STORED identifies and removes outliers, and performs denoising with weighted predictions. The algorithm demonstrates robust performance across synthetic datasets with various shapes and outlier compositions, both qualitatively and quantitatively. While challenges remain in handling discontinuities and complex surface features, STORED shows promising potential for enhancing stereo reconstruction and spatial analysis tasks. The source code is available at https://github.com/ickbumk/STORED.