MHIBS-Net: Multiscale hierarchical network for indoor building structure point clouds semantic segmentation

Abstract Efficient and accurate segmentation of indoor building structure (IBS) from complex indoor scenes is the primary task of indoor 3D modeling. At present, the research on semantic segmentation of indoor scenes mainly focuses on multiple types of objects, while there are few studies on semantic segmentation of IBS. In this paper, we introduce an efficient and lightweight multiscale hierarchical network MHIBS-Net for semantic segmentation of IBS point clouds. On the one hand, normal vector information (NVI) is added and a relative position encoding (RPE) unit is designed in feature encoding (FE) to better capture the local geometric structure information of the IBS point clouds. On the other hand, the traditional feature decoding (FD) process is improved by a newly designed multilevel hierarchy FD method. In this way, the MHIBS-Net can capture the local and global features of IBS point clouds comprehensively so as to realize automatic and efficient semantic segmentation of IBS point clouds. Finally, several groups of experiments are designed to compare and analyze the proposed MHIBS-Net with some classical semantic segmentation networks, and the Stanford large-scale 3D indoor spaces (S3DIS) dataset is used for experimental verification. Experimental results show that MHIBS-Net can achieve high-precision and high-efficiency automatic semantic segmentation of IBS point clouds in complex indoor environments.