Graph CNN For 3D Semantic Segmentation Of Point Clouds

In this paper,we design a noval graph convolutional neural network(CNN)called PGNet for the semantic segmentation of point cloud data.Firstly,the characteristics of irregular point cloud data are elaborated and related deep learning models for point cloud are reviewed.By comparing these models,we choose the graph CNN as the basic model of PGNet and we proposed two new operators for irregular point cloud called edge-conditioned diffusion and structure-preserved pooling.The PGNet is built by combination of these two operators and we apply it for semantic segmentation of point cloud.The PGNet is a compact point-based deep model which can extract roubust features using relative low computational complexity.To validate the performance of the proposed model,we test PGNet on three benchmark datasets which are S3DIS,ScanNet and Semantic3D.Net respectively.On all these datasets,PGNet achieves comparable or superior performance compared with state-of-the-arts.The experiments result show that PGNet can learn effective and generalizable features from point cloud for semantic segmentation task.