Building CAD Model Reconstruction from Point Clouds via Instance Segmentation, Signed Distance Function, and Graph Cut

Abstract

Although three-dimensional (3D) modeling of buildings is gaining increasing significance across various real-world applications, the concise representation of buildings from point clouds acquired through unmanned aerial vehicles (UAVs) and other means remains a formidable challenge. In this paper, we introduce an innovative framework for the reconstruction of individual 3D building CAD models derived from point clouds generated by UAV-captured photographs. Our framework encompasses four pivotal components: An instance segmentation model designed to extract buildings from UAV-observed point clouds. Estimation of building surfaces through the utilization of neural networks and the signed distance function of point clouds. Edge estimation based on the inferred building surface. Estimation of building polygons derived from the identified edges. Experimental results obtained from the SPLAT3D dataset affirm the capability of our proposed methodology to generate high-quality building models, thereby offering substantial advantages in terms of accuracy, compactness, and computational efficiency. Furthermore, we demonstrate the robustness of our approach against noise and incomplete measurements, thereby showcasing its applicability to point clouds obtained through photogrammetry utilizing UAV-captured photos.