Affine invariant regions++

Many problems in Computer Vision require the computation of correspondences between images. In order to cope with large differences in viewing conditions, several affine invariant region detectors have been developed in the last few years. These regions automatically adapt their shape so as to cover the same scene surface in any view. This dissertation builds upon existing detectors and develops various novel techniques which extend the power and functionalities of the regions. The advances relate to different subfields of Vision, and can be summarized as four main contributions. First and foremost, the thesis presents a powerful Object Recognition system capable of working with large amounts of background clutter, severe occlusion, and strong viewpoint and scale changes. It can handle non-rigid deformations, and also finds the contours of the visible parts of the object. The second innovation consists of a method to obtain region correspondences across several images taken from different viewpoints. These multi-view correspondences are important as they enable the automatic reconstruction of a 3D model given only a few still images. In constrast, traditionally this task requires a complete video as input. Another branch of the thesis introduces a real-time algorithm which tracks the full affine shape of a region as it evolves through a video, and its application for markerless Augmented Reality. Most prior works instead rely on adding special markers to the scene. Lastly, a technique to automatically find groups of regions correspondences lying on planar surfaces is presented. This allow to detect planar scene structures and their geometric transformation between views, which in turn can considerably simplify 3D reconstruction procedures, and is useful for robot navigation. Acknowledgements Four years is a long period of time. During this period I have met a number of wonderful people who influenced my work in various ways. First and foremost, I extend my deepest gratitude to Dr. Tinne Tuytelaars, who was always next to me, even thought a thousand kilometers divided our working places. Throughout the whole PhD, her brilliant intellectual support was second only to her amazing capacities to keep up a challenging and exciting working atmosphere, savour the successes, and react positively to defeats. My heartfelt thanks go to my supervisor, Prof. Luc Van Gool, whose enormous drive was a true inspiration. His expert advice, extensive knowledge of the Computer Vision field, and continuous incentives to improve my work were invaluable. I am grateful to my co-referee, Prof. …