Multiscale Bilateral Aggregation Network for Point Cloud Analysis

Point cloud is unordered, irregular, and scattered and inherently lacks topological information, and it is very challenging for semantic segmentation and understanding of point cloud. Existing point cloud segmentation methods usually ignore the structural information between points, which leads to the problem of high boundary segmentation error in the results of point cloud segmentation. This article introduces a multiscale bilateral aggregation method that improves the segmentation performance of boundary regions by learning semantic and geometric features from point clouds and establishing stronger connectivity and adjacency relationships. We conduct extensive qualitative and quantitative evaluations to evaluate our proposed method's effectiveness. The experimental result shows that our method has significant advantages in terms of accuracy and robustness in segmentation tasks on several benchmark datasets, including ShapeNetPart and Stanford Large 3-D Indoor Space (S3DIS).