融合几何特征与全局关系的室内点云语义分割

Semantic segmentation of indoorpoint clouds by fusing geometric features and global relations

为充分提取3D点云的深层特征以提高复杂室内点云场景的语义分割精度,提出一种结合局部特征和全局特征的室内点云语义分割网络GSFNet.在局部特征部分,加入几何特征信息,并设计几何与语义特征信息编码模块,以更好地捕获室内点云局部信息.对全局特征部分,在编码解码器结构中间层加入全局关系依赖模块,构建不同邻域对象之间的关系提取有效分割信息.使用斯坦福大规模室内数据集(S3DIS)进行实验验证,在测试数据集上测试的总体精度(OA)和平均交并比(mIoU)分别为87.2%和61.1%,实验结果表明,GSFNet对复杂室内环境有较好的语义分割效果.

In order to fully extract the deep features of 3D point clouds and improve the semantic segmentation accuracy of complex indoor point cloud scenes,this paper proposes an indoor point cloud semantic segmentation network combining local features and global features,named GSFNet.In the local features part,geometric feature information is added to the semantic feature information encoding module to better capture the local information of indoor point clouds.In the global feature part,a global relationship dependency module is added to the middle layer of the encoder-decoder structure to construct relationships between different neighbourhood objects and extract effective segmentation information.Experimental validation was conducted using the Stanford Large-Scale Indoor 3D Space Dataset(S3DIS).The OA and mIoU tested on the test dataset were 87.2%and 61.1%respectively,and the experimental results showed that GSFNet has better semantic segmentation for complex indoor environments.