Plane-based Surface Regularization for Urban 3D Reconstruction

We propose a method for urban 3D reconstruction that is a hybrid between a volumetric 3D reconstruction approach and a plane fitting approach in order to obtain a denoised and compact representation of the scene. In our hybrid approach, a single global optimization, using visibility as main information, defines whether the final reconstructed surface should align with a detected plane or rather follow the details of the input data. Our method is based on an established tetrahedral occupancy labeling approach which we taylor for urban reconstruction by adding the possibility to favor an alignment of the surface with detected planes. We further add novel regularization terms that favor Manhattan-like structures and which allow to control the level of detail of the output model. A variety of experiments demonstrate state-of-the-art performance and show that our approach is suitable for both indoor and outdoor environments.