3D Reconstruction of Buildings with Automatic Facade Refinement

3D reconstruction and texturing of buildings have a large number of applications and have therefore been the focus of much attention in recent years. One aspect that is still lacking, however, is a way to reconstruct recessed features such as windows and doors. These may have little value when seen from a frontal viewpoint. But when the reconstructed model is rotated and zoomed the lack of details will leap out. In this work we therefore aim at reconstructing a 3D model with refined details. To this end we apply a structure from motion approach based on bottom up bundle adjustment to first estimate a 3D point cloud of a building. Next, a rectified texture of the facade is extracted and analyzed in order to detect recessed features and their depths, and enhance the 3D model accordingly. For evaluation we apply the method to a number of different buildings.