PointNet ++ suffers from the problems that the network learning rate cannot be dynamically adjusted, poor robustness to noise and outliers, insufficient local feature extraction and lack of capture of global contextual feature information in the feature learning process, and insufficient efficiency of feature transfer. This paper aims to improve the network for the above problems. The main improvement includes three major parts: training strategy, feature learning, and feature propagation, and then the experimental environment and evaluation indexes carry out separate ablation experiments on these three parts to illustrate the enhancement effect on the network, the enhancement effect on the network segmentation effect after the combination of all the modules, and finally carry out the comparison between before and after the improvement of the network. The results show the improvement effect is obvious.