A Compact Surface Reconstruction Method for Buildings Based on Convolutional Neural Network Fitting Implicit Representations

Abstract

Three-dimensional building models have a wide range of applications in smart cities, urban planning, and disaster assessment. However, how to efficiently represent 3D building models with fewer facets is still a pressing problem. In this paper, we propose a method that can extend the application scope of convolutional occupancy networks to outdoor unmanned aerial vehicle (UAV) building point cloud and reconstruct 3D architectural models with fewer facets. The method is comprised of three main steps. First, a candidate set of cells is constructed through spatial division. Second, a convolutional occupancy network is employed to recognize the occupancy state of the cells. Last, the graph cut algorithm is used to select a suitable set of cells to form the final surface model. In order to verify the effectiveness of the method, this paper reconstructed complex buildings from noisy point clouds and compared them with several reconstruction methods. The experimental results demonstrate that the proposed method cannot only rapidly reconstruct a single building but also be applied to multibuilding complexes.