Part segmentation method of point cloud considering optimal allocation and optimal mask based on deep learning

Abstract

In order to enhance the generalization ability of the network and improve the precision of the part segmentation, a point cloud part segmentation method which takes into account the optimal allocation and the optimal mask is proposed. Firstly, the optimal allocation between two point clouds is defined according to earth mover's distance. Then the farthest point sampling is used to group the point cloud, and the saliency of each point in the group is calculated. Finally, a new mixed sample is generated by replacing a partial subset of another point cloud sample with a local neighborhood in one point cloud sample. In this paper, the ShapeNet dataset was verified and the enhanced data was transferred to PointNet, Pointnet ++ and DGCNN models using this method. The mIoU increased from 83.7%, 85.1% and 85.1% to 84.6%, 85.9% and 85.7%, respectively. Effectively improve the effect of part segmentation.