Early Vision Optimization: Parametric Models, Parallelization and Curvature

Abstract: Early vision is the process occurring before any semantic interpretation of an image takes place. Motion estimation, object segmentation and detection are all parts of early vision, but recognition is not. Many of these tasks are formulated as optimization problems and one of the key factors for the success of recent methods is that they seek to compute globally optimal solutions. This thesis is concerned with improving the efficiency and extending the applicability of the current state of the art. This is achieved by introducing new methods of computing solutions to image segmentation and other problems of early vision. The first part studies parametric problems where model parameters are estimated in addition to an image segmentation. For a small number of parameters these problems can still be solved optimally. In the second part the focus is shifted toward curvature regularization, i.e. when the commonly used length and area regularization is replaced by curvature in two and three dimensions. These problems can be discretized over a mesh and special attention is given to the mesh geometry. Specifically, hexagonal meshes are compared to square ones and a method for generating adaptive methods is introduced and evaluated. The framework is then extended to curvature regularization of surfaces. Thirdly, fast methods for finding minimal graph cuts and solving related problems on modern parallel hardware are developed and extensively evaluated. Finally, the thesis is concluded with two applications to early vision problems: heart segmentation and image registration.