Texture map reconstruction from multiple images via optimized blending

Abstract

In image-based modeling and rendering, once the geometrical model is obtained, it is a key step to recover the information about the surface properties, such as color, shininess or roughness, represented by the texture map. The realism of a 3D model strongly depends on the quality of its texture map. Even with a rough approximation of the geometry of the reconstructed object, it has been proved that texture maps effectively enhance the visual realism of the model. In the construction of a texture map, multiple observations of a same region of the object surface are available; correctly combining them to retrieve all the information contained in those observations is key to the final realism of the model. In this thesis, the issues involved in texture recovery are presented and an effective method addressing the problem of texture recovery from multiple images is proposed. The proposed method consists in an effective weight optimization method that iteratively improves the weights in order for the estimated texture to match the observed images better. During the process, image resampling is required in several steps. In order to alleviate the aliasing effect caused by undersampling, anti-aliasing filtering must be considered. The need to process multiple images and patches demands for an important speed rate, but not at the cost of quality sacrifice. A fast resampling methods based on a Gaussian kernel is proposed in order to reduce the computational burden without penalizing the quality of the resampled images. In the experiments performed on several datasets, the proposed method shows the ability to correctly recombine the multiple inputs preserving the details available in each of the images and smoothly blending them to avoid boundary discontinuities.