Asymmetric Network Based on Feedback and Transformer for Multispectral LiDAR Point Cloud Semantic Segmentation

3D point cloud semantic segmentation extends the development of computer vision. Accurate point cloud semantic segmentation is a fundamental problem in point cloud applications. However, effective point cloud semantic segmentation is still a competitive problem due to the disorder and irregularity of point clouds. Witnessing the success of the Transformer structure in natural language processing and 2D computer vision, researchers have introduced the Transformer into point cloud semantic segmentation and conducted many helpful explorations. The Transformer is characterized by global modeling, and it is worth exploring how to obtain the local spatial inductive bias. In this paper, we design an asymmetric network that simultaneously extracts local and global features. The network structure consists of two branches: a pointwise convolutional network with a feedback mechanism and a Transformer structure based on multi-scale pooling. The cascaded asymmetric network structures are used for point cloud semantic segmentation. We validate the effectiveness of the proposed network on a multispectral LiDAR point cloud dataset. In addition, we also conduct a series of experiments to explore the effectiveness of different structures. The proposed method achieves state-of-the-art performance compared with previously designed point cloud semantic segmentation networks.