Characteristic Analysis of Data Preprocessing for 3D Point Cloud Classification Based on a Deep Neural Network: PointNet

Abstract

Laser scanning is a noncontact and nondestructive technique that captures the three-dimensional (3D) shape of objects as point clouds. Deep neural networks have been widely used to classify the 3D shapes of point clouds. In applying deep learning on point clouds, point cloud preprocessing is the first step. This study was conducted to analyze 3D shape classification characteristics using a deep neural network, PointNet, with a point cloud dataset, ModelNet40, for four preprocessing cases: random, scaling, zero-mean, and normalization. For each preprocessing case, the minimum and maximum coordinates of the point clouds and 3D shape classification performance are investigated. The results show that normalization preprocessing exhibits the most significant improvement in classification performance, and the zero-mean method is particularly effective. The findings indicate that proper preprocessing, such as normalization, should be performed before deep learning when the mean coordinates and scale of the point clouds differ significantly.