Influence of Preprocessing and Augmentation on 3D Point Cloud Classification Based on a Deep Neural Network: PointNet

Three-dimensional (3D) laser scanning is widely used to acquire the structural information of a target as a point cloud and reconstruct its shape. Recently, deep learning has shown good performance for 3D point cloud shape classification. The preprocessing of the point cloud is a primary step of deep learning. This study presents the performance of 3D shape classification via PointNet with a point cloud dataset, ModelNet40, with respect to three preprocessing cases: Random, zero mean, and normalization. The minimum and maximum values of the point cloud are compared according to the preprocessing method. In training, the number of points as an input was 1024. In addition, the influence of two augmentation methods (i.e., resampling and zero filling) was investigated. For this, the number of points was increased to 2048. Of the 2048 points, 1024 points were used the same as in the previous experiment, while the remaining 1024 points were added by resampling or zero filling. The results show that the zero mean method is effective for deep learning and normalization is better, whereas increasing the input size with the resampling or zero filling rather degrades the performance and increases unnecessary training costs.