Consistent Multi-View Texturing of detailed 3D surface models

Kurzfassung

Texture mapping techniques are used to achieve a high degree of realism for computer generated large-scale and detailed 3D surface models by extracting the texture information from photographic images and applying it to the object surfaces. The photo-realistic appearance of the large-scale and detailed surface models is important for many applications: from environmental and physical simulations to 3D city models. Due to the fact that a single image cannot capture all parts of the scene, a number of images should be taken. However, texturing the object surfaces from several images leads to lightning variations between the neighboring texture fragments. Some techniques blend the textures from several images per face to build a texture map. Within the scope of this master thesis, we describe a creation of a “seamless” texturing of a 3D scene from overlapping areal images, applying a Markov Random Field energy minimization framework. As an input data we use a triangulated mesh of the city center of Munich and multiple camera views of the scene from different directions. We start with the projection of the 3D mesh faces into the input images in order to find the image with the “best” texture for these faces. Afterwards, by back-projecting the images onto the 3D mesh surface and using the Markov Random Field energy minimization framework we aim to maximize the quality of the generated texture mosaic, preserving the resolution from the original images, and at the same time to minimize the seam visibilities between adjacent fragments. This is achieved due to the method’s ability to search for a global solution in one optimization framework based on prior knowledge of the system. Finally, validation of the approach is done on synthetically generated models to demonstrate and test the method’s performance on different sites.