一种探讨点云深度学习决策的PointNet++解析网络  

A PointNet++ Analytic Network that Explores Point Cloud Deep Learning Decision-Making

作　　者：龚国栋 李耀斌[1] 花向红[2,3] 赵不钒 卢荣[1] GONG Guodong;LI Yaobin;HUA Xianghong;ZHAO Bufan;LU Rong(Guangzhou Urban planning&Design Survey Research Institute,Guangzhou 510030,China;School of Geodesy&Geomatics,Wuhan University,Wuhan 430079,Chin;Hazard Monitoring&Prevention Research Center,Wuhan University,Wuhan 430079,Chin)

机构地区：[1]广州市城市规划勘测设计研究院,广东广州510030 [2]武汉大学测绘学院,湖北武汉430079 [3]武汉大学灾害监测与防治研究中心,湖北武汉430079

出　　处：《测绘地理信息》2022年第6期50-54,共5页Journal of Geomatics

基　　金：国家自然科学基金(41674005,41871373)。

摘　　要：针对三维点云数据分类深度学习可解释性研究,提出一种探讨点云深度学习决策的PointNet++解析网络,探索隐藏在PointNet++网络中的特征信息。根据二维图像解译工作中的类激活映射图,提出了三维点云的类激活映射图,并将点云类激活映射图作为探索PointNet++网络分类决策的依据,采用多层感知机取代全连接层,并使用均值池化层来聚合卷积特征。实验数据为ModelNet40数据集,验证了所提出的PointNet++解析网络的可行性。研究结果表明,所提算法达到了较高的分类精度并且能够对PointNet++分类决策进行探讨,提取直接有助于决策制定的特征区域。Aiming at the interpretability of 3D point cloud data classification deep learning,this paper proposes a PointNet++ analytic network that explores point cloud deep learning decision-making,and explores the feature information hidden in the PointNet++ network.According to the class activation map in the 2D image interpretation work,the class activation map of the 3D point cloud is proposed,and the point cloud class activation map is used as the basis for exploring the classification decision of the PointNet++ network,and the multi-layer perceptron is used to replace the full connection Layers and use mean pooling layers to aggregate convolutional features.The experimental data is the ModelNet40 data set,which verifies the feasibility of the proposed PointNet++ parsing network.The research results show that the proposed algorithm achieves high classification accuracy and can discuss PointNet ++ classification decision-making,and extract feature areas that directly contribute to decision-making.

关 键 词：点云 深度学习 PointNet++ 解析网络 

分 类 号：P237[天文地球—摄影测量与遥感]