U-shaped network based on Transformer for 3D point clouds semantic segmentation

3D point clouds processing is a significant technical direction of autonomous driving, computer vision, and 3D mapping. However, due to the disorder and irregularity of 3D point clouds, it brings some challenges to its development. In recent years, Transformer, as an important technology in natural language processing, has been successfully applied in 2D image processing and achieved excellent results. Recently, relevant research on the application of Transformer on 3D point clouds has also been published. In this paper, we refer to the self-attention mechanism in the transformer architecture and propose a U-shaped network based on Transformer for 3D point clouds segmentation. And we do semantic segmentation experiments on the Stanford Large-Scale 3D Indoor Spaces Dataset (S3DIS). Experiments show that the performance of our proposed network is better than some semantic segmentation algorithms in common evaluation metrics.