Semantic segmentation of point cloud based on graph neural network

Abstract

As a key technology in the field of computer vision, point cloud semantic segmentation has been widely used in intelligent life and has become a hot spot and a difficult area of research in 3D vision. Current research on the semantic segmentation of 3D point clouds is mainly focused on the field of deep learning, however, there are two problems with such models: insufficient local feature representation and long model training period. To this end, we propose a point cloud semantic segmentation model based on geometric segmentation and graph neural network to improve the performance of semantic segmentation and model training efficiency. In our method, Firstly geometric segmentation is used for sparse and feature representation of the original point cloud. A directed graph is then constructed to partition the point cloud into "superpoints" and "superedges" with specific feature descriptors. followed by a spatial transformer network (STN) as the backbone of embedded network to encode local features from these superpoints. Finally, the graph neural network is employed for the classification of architectural components such as walls, ceilings, floors, and beams. Experiments were conducted on the Stanford University Large Scale 3D Indoor Spatial Dataset (S3DIS) dataset. Compared with the SPG network that also takes the entire large scene as input, the proposed algorithm improves the overall accuracy (OA), mean accuracy (mAcc), and per-class intersection over union (mIoU) by 1.0%, 2.9%, and 3. 0%.