摘要
基于深度学习方法在城市激光点云语义分割任务中的应用效果缺乏客观的对比与评价,该文选取当前4种代表性点云语义分割深度网络(PointNet、PointNet++、PointCNN、SPG)以及一种基于特征描述子的层次化点云语义分割方法,采用3组开放点云数据集(Semantic 3D、Oakland及TerraMobilita/iQmulus3Durban)对不同方法的语义分割质量进行对比分析,结果发现:1)层次化点云语义分割方法的语义分割质量优于另外4种深度学习方法;2)考虑局部信息的深度网络(PointNet++、PointCNN、SPG)的表现优于仅考虑点云全局特征的方法(Point-Net);3)在基于深度学习的方法中,基于超点图的SPG网络在测试数据中的效果优于其他几种网络。研究结果对于实际应用选择点云语义分割方法以及点云语义分割深度网络的设计优化具有借鉴意义。
In recent years,deep-learning methods have been developed for semantic segmentation of laser point clouds in urban areas.However,the application effect of deeplearning methods needs to be evaluated.This study presents a comparative analy-sis of four typical approaches(PointNet,PointNet++,PointCNN,SPG)and a hierarchical semantic segmentation method based on feature descriptors.Three benchmark data sets,i.e.,Semantic 3D,Oakland and TerraMobilita/iQmulus 3D urban,are used to evaluate the performance of these approaches.The results show that:1)the hierarchical method performs better than the four typical deep-learning approaches;2)the approaches(PointNet++,PointCNN,SPG)considering local information of point clouds perform better than that only considering global features of point clouds(PointNet);3)SPG performs the best among the four deeplearning approaches.The findings of this paper will be helpful for selecting point clouds semantic segmentation methods and designing deeplearning semantic segmentation methods in practice.
作者
杨柳
刘启亮
袁浩涛
YANG Liu;LIU Qi-liang;YUAN Hao-tao(Department of Geo in.formatics,School of Geosciences and Info physics,Central South University,Changsha 410083,China)
出处
《地理与地理信息科学》
CSCD
北大核心
2021年第1期17-25,I0001,共10页
Geography and Geo-Information Science
基金
国家自然科学基金项目(41971353)
湖南省自然科学基金项目(2020JJ4695)。
关键词
激光点云
城市三维信息
语义分割
深度学习
特征描述子
laser point clouds
3D information of city
semantic segmentation
deep learning
feature descriptors
