Efficient roof reconstruction from a single aerial image

Abstract

We present a novel and fast LoD2 (level of detail) modeling method for generating high-quality 3D polygonal meshes from single aerial images. Our approach initially detects line segments to extract polygons as vectorized contours for target roofs. Subsequently, the topological structures of 2D roofs are generated and then optimized by solving geometric constraints. Finally, refined 2D roof structures are lifted to 3D meshes using a geometry-preserving extrusion algorithm. The proposed approach achieves superior quality compared to existing point cloud-based algorithms. In addition, our pipeline allows for interactions between steps. With simple user adjustments, our method can handle complicated roof cases, significantly improving efficiency compared to manual modeling software.