An Improved Point Cloud Segmentation Method Based on Multi-scale Fusion Strategy

With the rapid development of 3D sensors, point cloud semantic segmentation methods have gradually become important components of 3D scene understanding. Considering the limitation of the local receptive field, we design dilate sampling, which improves the segmentation performance without increasing the computation. Furthermore, we fuse information from multiple scales in the decoder to recognize objects of various sizes. After experiments, we obtain comparable results on the S3DIS dataset and Toronto3D dataset.