Optimum Scale Selection for 3D Point Cloud Classification through Distance Correlation

Abstract

Multiple scale machine learning algorithms using handcrafted features are among the most efficient methods for 3D point cloud supervised classification and segmentation. Despite their proven good performance, there are still some aspects that are not fully solved, determining optimum scales being one of them. In thiswork, we analyze the usefulness of functional distance correlation to address this problem. Specifically, we propose to adjust functions to the distance correlation between each of the features, at different scales, and the labels of the points, and select as optimum scales those corresponding to the global maximum of said functions. The method, which to the best of our knowledge has been proposed in this context for the first time, was applied to a benchmark dataset and the results analyzed and compared with those obtained using other methods for scale selection.