An international conference connecting people
in CAD research, education and business
Copyright (C) CAD Solutions, LLC. All rights reserved.
Proceedings of CAD'16, 2016, 142-146
Learned 3D Shape Descriptors for Classifying 3D Point Cloud Models
Abstract. The recent developments in 3D sensing devices that deliver high-quality raw 3D data in real time offer growing opportunities to explore the usage of this data in a variety of 3D perception and reasoning tasks. In this paper, we focus on the problem of classifying 3D point clouds, and we are integrating different supervised machine learning classifiers with several capable yet underexplored shape descriptors based on visual similarity (light-field), angular radial transform (ART) and Zernike moments. Specifically, we investigate the use of 3D Zernike descriptors as well as a combination of 2D ART descriptors with light field techniques to construct and compare the performance of practical descriptors for 3D point cloud classification. We train our classifiers with a database of point clouds corresponding to several common objects obtained by sampling polygonal models obtained from Google’s 3D Warehouse and by post-processing them to attain controlled levels of density and noise. We show that these descriptors provide a promising alternative to the current shape descriptors employed for classifying point clouds in the presence of noise.
Keywords. Point Clouds, Classification, Shape Descriptors, Zernike Moments.