PointSegmentorNet: Incorporating Graph Embeddingbased Convolution for Point Cloud Transformer

Point cloud segmentation is a crucial part of 3D scene understanding. Existing work lacks the capability to integrate local and global features of point clouds. Furthermore, the priority of local correlations and long-term dependencies in unordered point clouds is still unclear. Therefore, we propose a Unet-like model with each building block incorporating a graph embedded based convolution for the Farthest Point Sampling (FPS) along with local feature extraction, and a point transformer facilitates the combination of local correlations and long term dependencies of point cloud. Additionally, an Adjustable Channel Information Selection Module (ACISM) is also proposed for bottleneck feature sparsity, which turns out to increase the generalizability of the proposed model. The model combines the robust ability of graph convolutions to merge spatial and semantic information with the Transformer's global context awareness, reducing the loss of spatial structural information and the topological relationships among points in the feature extraction process. Experiments on the S3DIS and ScanNetv2 datasets turns out that our model shows better performance than the other methods in point cloud segmentation tasks.