Point Cloud Semantic Segmentation with Geometric Contextual Aggregation and Adaptive Feature Enhancement

Most existing methods mainly learn geometric representations with spatial coordinates of the points, while the complementarity between local geometric structures and high-level semantic features is ignored. Therefore, we propose a segmentation network to exploit and adaptively enhance the local geometric features. For each 3D point, we first encode the finegrained geometric structure by fully utilizing the spatial relative position, orientation information, and normal information in the neighborhood. Then, we designed a module to aggregate the learned geometric representation and semantic information, which adaptively enhances local features and reduces information loss. Finally, based on the above two modules, we use residual connections to aggregate geometric context information. The experimental results on the S3DIS public dataset show that the network proposed in this paper achieves competitive performance.