Mapping synthetic texture on surfaces

Computer Graphics has become an important tool in many fields such as the entertainment industry (movies, games), medicine (simulation), or architecture. High quality surface textures are a significant contribution towards realistic and visually appealing results. Often, the scene designer would like to cover a 3D model with a texture pattern that looks like a given sample, Even nowadays, this is still a difficult undertaking which requires skills and a lot of time. Hence, the goal of this thesis is the following. Given a polygonal 3D model and a small digital sample image of a texture, the system should automatically cover the surface with synthetic texture that looks like the sample texture. Of course, additional user input such as desired local orientations of the pattern may be required. The result should be an ordinary 2D texture image that can be mapped to the surface. Our basic approach consists of transforming the problem from 3D to the plane, perform texture synthesis in that domain, and map the result back to the surface. One advantage of this approach is that one of the many 2D texture synthesis methods (based on planar geometry) can be used with minor extensions. On the other hand, a consequence is that the intrinsic geometrie problems related to the mapping between the plane and the curved surface have to be dealt with explicitly. Our novel 2D texture synthesis algorithm belongs to the family of methods based on global feature matehing and is formulated in terms of a framework with configurable components. There are different choices for a) the statistical features (based on n-dimensional cooccurrences), b) the distance function that measures the deviation of the current image features to the reference ones and c) the minimisation method. We systematically experimented with several combinations of the components and found that two-dimensional cooccurrences and a Bhattacharyya based distance function are in general superior to the other choices. Also, an adequate minimisation method is essential for a good result: very stochastic textures can efficiently be synthesised using the deterministic "greedy" pixel sampler, whereas regular ones definitely require one of the stochastic methods. Additional features include synthesis according to per-pixel orientations and the adaptation of an existing texture analysis strategy to select the relevant features. The main remaining