City-scale building instance segmentation from LiDAR point clouds via structure-aware method

• A novel pipeline to achieve both efficient building semantic segmentation and robust building instance segmentation. • VPBE-Net for automatic building points segmentation in city scenes by leveraging the voxel-point cloud fused features. • A graph-based building instance clustering algorithm SI-DVDC that resists interference from misclassified points and noise. • We provide a City-BIS dataset that totally encompasses more than 7,000 building annotations. Building instance segmentation from city-scale point cloud is of great significance to urban planning management, disaster response and recovery, and land resource management. However, due to the complexity of urban scenes and sparse nature of LiDAR data, existing methods are often limited by the problems of obscured building boundaries and incomplete building structures, particularly in densely populated urban areas with diverse architectural styles. To address these challenges, we propose a novel method that automatically extracts building instances from airborne LiDAR data and is especially aware of the building structures. The proposed method encompasses two main stages, building points semantic segmentation and individual building extraction. First, we design a point cloud semantic segmentation network, VPBE-Net, that innovatively utilizes voxel-point cloud fused features to efficiently extract building points from large-scale point cloud. Second, building instances are automatically and robustly extracted using a graph-based algorithm SI-DVDC, which comprehensively considers both object-level building structure property and point-level density accessibility. We evaluate the semantic segmentation performance on the DALES and Toronto datasets and the building instance segmentation performance on the UrbanBIS and City-BIS datasets. For the semantics, Overall Accuracy (OA) and mean Intersection over Union (mIoU) metrics reach 88.96 % and 70.28 % on DALES dataset, and 89.26% and 75.40% on Toronto dataset, which is 2.22 % and 3.25 % higher than the state-of-the-art methods. For the building instance extraction, the instance-level quality metric reach 88.65 % on UrbanBIS dataset and 76.97 % on City-BIS dataset, respectively. The experiments verify that the proposed method can extract individual buildings from complex urban and rural environments, while being aware of diverse building structures, thereby demonstrating the remarkable generalization ability. To facilitate future research, we make source code and dataset available at https://github.com/Lijp411/City-BIS.