Computer Vision Using Rich Features - Geometry and Systems

Abstract: In this thesis a number of computer vision problems is discussed, in particular we study the problem of view synthesis. The focus is on using complex features to determine the geometry of the scene and the cameras, in contrast to traditional point-based methods. The thesis is divided into two parts: In the fist part geometric relations between different features, corresponding in several images are investigated. Among others geometric relations of images of planes are presented. It is shown how these relations may be used to generate images from novel viewpoints and to create 3D models from images. The feature quiver, defined by a point and a number of directions from this point, is introduced. Three minimal cases of estimating the structure and motion using correspondences of quivers in three images are solved. In the second part of the thesis a number of implemented computer vision systems is presented. First a novel system for automatic generation of images from new view points using the information in a number of given images, is described. Secondly we propose an automatic system for visualisation of fridge contents. 3D models of the objects in the fridge are created when they are inserted. Thirdly a system for detecting windows in a city scene based on support vector machines is presented, and finally we describe a system for estimation of position and orientation from a image, when a model of the surroundings is available.