3D-GCNN - 3D Object Classification Using 3D Grid Convolutional Neural Networks

Abstract

In this paper we propose to solve the problem of 3D Object Classification on point cloud data. We propose a 3D CNN architecture which we call 3D-GCNN that consumes point cloud data directly and performs classification. We present a novel method to represent the point cloud data using a margin based density occupancy grids which creates a minimum volume bounding box around the point cloud data. This information is fed to our proposed 3D CNN model which has far lesser trainable parameters and ensures convergence. We demonstrate our results on the ModelNet 10 and ModelNet 40 models and show we achieve better or comparable performance compared to other methods.