Texture mapping for 3D building models

Abstract

The need and attempt for creating mathematically-defined three dimensional (3D) models of real world objects has a long history. While, in the past, creating the model was a problem itself, with recent developments in 3D reconstruction, creating accurate, photorealistic and photogrammetric 3D models of the objects has become the focus point. Starting from the architectural models, an interactive texture mapping system was developed concentrating on the visual appearance of the predefined models. The system presents a semi-automatic way of extracting, correcting, and mapping the appropriate textures to the given 3D building models using the images obtained by standard consumer-level digital cameras. Projective geometry takes place in extraction and rectification of texture candidates while popular graph cut optimization approach was utilized to create seamless texture composites using these candidates. A refinement phase was adopted for this procedure with a series of refinement tools including Poisson image editing in terms of seamless cloning. For testing purposes some publicly available datasets were used besides the imagesets that were created by photographing real world objects. It was shown that most buildings could be textured in an acceptable photorealistic quality without any predefined information about the datasets. Furthermore, it was observed that the texture mapping times, even for detailed building models, were quite low. This work mainly focuses on close-range or ground level imagery since the aim is to create detailed and high quality photorealistic view of the models. However, the approach could easily be extended towards the needs of aerial imagery or large scale reconstruction.