Piecewise Planar and Non-Planar Segmentation of Large Complex 3D Urban Models

Abstract

Advancements in computing power via Multi Core processors and GPUs have made large scale reconstruction modeling and real-time photorealistic rendering possible. However, in urban areas flat surfaces with little texture still challenge multiview algorithms. We present a method for planar area recognition and model correction while avoiding deformation of non-planar areas such as domes, pillars and plant matter. Our method works in object space, allows a global solution that is not affected by individual range map inaccuracies or poorly matched range maps. We describe a segmentation of the model into bounded planar and non-planar areas driven by a global error function incorporating model shape and original images texture. The error is minimized iteratively using locally restricted graph cuts and the model is corrected accordingly. The algorithm was run on various complex and challenging real-world urban scenes and synthetic photo-realistic images are created from novel viewpoints without noticeable deformities that are common to typical reconstructions.