MNAT-Net: Multi-Scale Neighborhood Aggregation Transformer Network for Point Cloud Classification and Segmentation

Accurate understanding of 3D objects in complex scenes plays essential roles in the fields of intelligent transportation and autonomous driving technology. Recent deep neural networks have made significant progress in 3D visual tasks by using point cloud data. However, the acquisition of geometric features and the expression of local fine-grained features in point clouds are still not sufficient for the classification and segmentation tasks. Inspired by the application of transformer structures in 2D and 3D computer vision tasks, in this paper, a multi-scale neighborhood aggregation transformer network (MNAT-Net) is proposed for point cloud classification and segmentation, which captures the global semantic information and local geometric structure features of point clouds by aggregating the receptive field and node weights. MNAT-Net consists of three key components, namely the multi-scale neighborhood feature aggregation module, the global transformer module and the category-weighted focal loss. The neighborhood features learned by the MNAT-Net network is sent to the global transformer module to fully enrich the contextual representation. Experimental results show that MNAT-Net achieves competitive performance on publicly available ModelNet40, ShapeNet, S3DIS and SemanticKITTI data sets in comparison to related methods.