MHNet: Multiscale Hierarchical Network for 3D Point Cloud Semantic Segmentation

Abstract

Point cloud semantic segmentation is a challenging task in 3D understanding due to its disorder, unstructured and nonuniform density. Currently, most methods focus on network design and feature extraction. However, it is difficult to capture the point cloud features of complex objects comprehensively and accurately. In this paper, we propose a multiscale hierarchical network (MHNet) for 3D point cloud semantic segmentation. First, a hierarchical point cloud feature extraction structure is constructed to learn multiscale local region features. Then, these local features are subjected to feature propagation to obtain the features of the entire point set for pointwise label prediction. To take fully advantage of the correlations of propagated information between the different scale coarse layers and the original points, the local features of each scale are characterized by feature propagation to obtain the features of the original point clouds at the corresponding scale. The global features propagated from different scales are integrated to constitute the final features of the input point clouds. The concatenated multiscale hierarchical features, including both local features and global features, can better predict the segmentation probability of each point cloud. Finally, the predicted segmentation results are optimized using the conditional random filed (CRF) with a spatial consistency constraint. The efficiency of MHNet is evaluated on two 3D datasets (S3DIS and ScanNet), and the results show performance comparable or superior to the state-of-the-art on both datasets.