A comparative study of 3D point cloud classification: traditional method vs. deep learning approach

Details

Supervisors

Jungers, Raphaël M. ; Lee, John Aldo

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en science des données, à finalité spécialisée: analyse de données

Abstract

This master thesis investigates 3D point cloud classification by analysing the differences between traditional methods and deep learning approaches on the ModelNet10 Dataset. Before deep learning became the reference of point cloud classification and segmentation with the emergence of PointNet in 2017 that managed to consume the set of points directly in the network, traditional methods with feature descriptors were used to perform the same task. This thesis focuses on a comparison in terms of accuracy, performance and interpretation between 2 methods for a point cloud classification task: a traditional method based on point feature histogram and PointNet, the reference deep learning method. It also describes the theoretical architecture of these methods and a particular feature selection approach using a margin-based iterative algorithm (SIMBA). Finally, it also presents an in-depth analysis of PointNet++, an enhanced version of PointNet, and a comparison of Farthest Point Sampling, the default sampling algorithm used in PointNet++, and K-Means, an alternative used in its place.