PI Net:基于3D点云与RGB数据的大型室内场景语义分割

\justifying Semantic segmentation for point cloud scene has always been a challenging task. Due to the rich geometric data structure of point cloud, most researchers focus on the spatial structure of point cloud. However, the appearance information of point cloud is not fully utilized. In this paper, we propose a novel deep fusion network, named PI Net, for semantic segmentation. Our proposed network has two branches, P-Net and I-Net, to combine the spatial and the appearance information of point cloud data. P-Net employs the raw point cloud to extract the spatial features. Meanwhile, I-Net transforms point cloud data to corresponding images and then extracts appearance information. Finally, we use projection transform to align and fuse the features of the two branches for semantic segmentation. Our proposed PI-Net achieves state-of-the-art performance on Stanford 2D-3D-Semantics Dataset.