Abstract

This paper introduces a new definition of multiscale neighborhoods in 3Dpoint clouds. This definition, based on spherical neighborhoods andproportional subsampling, allows the computation of features with a consistentgeometrical meaning, which is not the case when using k-nearest neighbors. Withan appropriate learning strategy, the proposed features can be used in a randomforest to classify 3D points. In this semantic classification task, we showthat our multiscale features outperform state-of-the-art features using thesame experimental conditions. Furthermore, their classification power competeswith more elaborate classification approaches including Deep Learning methods.

 

Quick Read (beta)