PG-Net: Point-Graph Convolution Neural Network on Semantic Segmentation of Point clouds

Different from the traditional 2D images, 3D point clouds are irregular, disordered, and sparse, which makes it impossible to capture their spatial information well. It is challenging to explore the semantical meaningful context as well as balance such features extracted from local and global scales in a 3D point cloud. The observation on previous studies illustrates that, researchers might pay too much attention to how to obtain local contextual features rather than investigate the importance of well utilizing the global feature. To address this problem, we proposed PG-Net, a simple but effectiveness network, which captures local contextual features and global semantical features by constructing KNN Graph and shared Multilayer Perceptron (MLP), respectively, followed by a merge operation. The experiment on S3DIS benchmark shows that, our model outperforms other competing methods that were recently published with certain margin.