FCPNet: Method for Rescuing Feature Information Loss in Scaling Change for Urban 3-D Point Cloud Classification

Abstract

The loss of feature information during scale propagation in the deep learning method usually causes a big misclassification rate for many complex urban scenes. For this reason, this article presents a new deep learning method, called “Feature combination and promotion network (FCPNet).” This method consists of an end-to-end feature learning layer for obtaining multiscale depth features of point clouds, an external feature combination module for obtaining more fine-grained point cloud features, and a mutiheaded separable self-attention module for learning connections between features to obtain more globally informative features. When compared with PointNet++, the proposed FCPNet improved OA, MIOU, F1-score, and Kappa in the NPM3D dataset by 1.75%, 17.02%, 2.13%, and 0.0263, respectively. When compared with KpConv, the proposed FCPNet improved OA, mIOU, F1-score, and Kappa in the NPM3D dataset by 0.36%, 12.11%, 0.77%, and 0.0085, respectively. Especially, the proposed FCPNet is able to classify the objects with fewer point cloud data, such as pedestrians and cars, whose OA can reach 88.04% and 96.42%, respectively. These experimental results demonstrated the proposed FCPNet has rescued much lost information that happened in the traditional PointNet++. In addition, the adaptability to point cloud density variations for the proposed method is verified as well. The results demonstrated that when the total density of point cloud data decreases from 731.3 to 52.2 ${\text{pcs/}}{{\text{m}}^{2}}$, the OA of classification with the proposed FCPNet method only decreases by 3.07%. This means that the proposed FCPNet method is capable of being adaptive to the point cloud density changes.

Keywords

• Classification
• deep learning
• information loss
• point cloud data
• separable self-attention
• three-dimensional (3-D)