GeoSep-PointNet++: A 3D Point Cloud Semantic Segmentation Algorithm Based on Geometric Information Enhancement and Feature Separation Encoding

To enhance the information encoding of point cloud geometry and multi-feature utilization, thereby further improving the accuracy of 3D point cloud semantic segmentation, this paper proposes a GeoSep-PointNet++ network based on geometric information enhancement and feature separation encoding. The GEO module is used to enhance the geometric information encoding of the local space, introducing the coordinates of the sampling points in the local space, their own coordinates, and Euclidean distances, which increases the information of each point in the space by 7 dimensions. The SEP module is used to achieve independent encoding of point cloud additional features and geometric features, separating geometric features from additional features and processing them separately with MLP. Experiments on the semantic segmentation of the S3DIS public dataset show that the proposed network's OA and mIoU are 82.4% and 51.96%, respectively, which are 1.3% and 2.57% higher than PointNet++; compared with the segmentation results of other models, the performance of the proposed model has been improved to varying degrees.