Experimental Evaluation of Point Cloud Classification using the PointNet Neural Network

Abstract

Recently, new approaches for deep learning on unorganized point clouds have been proposed. Previous approaches used multiview 2D convolutional neural networks, volumetric representations or spectral convolutional networks on meshes (graphs). On the other hand, deep learning on point sets hasn’t yet reached the “maturity” of deep learning on RGB images. To the best of our knowledge, most of the point cloud classification approaches in the literature were based either only on synthetic models, or on a limited set of views from depth sensors. In this experimental work, we use a recent PointNet deep neural network architecture to reach the same or better level of performance as specialized hand-designed descriptors on a difficult dataset of nonsynthetic depth images of small household objects. We train the model on synthetically generated views of 3D models of objects, and test it on real depth images.