3D point cloud analysis

Abstract

This paper presents a study that investigates the effectiveness of various sampling approaches when combined with the KPConv framework for 3D point cloud segmentation. The focus is mostly on the original grid subsampling strategy employed by the framework. In this study, a series of experiments were conducted to assess and contrast the outcomes derived from the utilization of three distinct techniques: inherent grid subsampling, random sampling, and the Farthest Point Sampling (FPS) approach. Initial results suggest that there are differences in the accuracy and training efficiency of the model. The objective of this study is to provide a thorough examination, elucidating the benefits and possible limitations of each approach. The objective of this study is to provide valuable insights into the optimization of point cloud processing techniques and to establish the superiority of a certain sampling approach in the context of KPConv-based point cloud analysis.