Understanding Image Registration

Vision researchers have created an incredible range of algorithms and systems to detect, track, recognize, and contextualize objects in a scene, using a myriad of internal models to represent their problem and solution. However in order to effectively make use of these algorithms sophisticated expert knowledge is required to understand and properly utilize the internal models used. Researchers must understand the vision task and the conditions surrounding their problem, and select an appropriate algorithm which will solve the problem most effectively under these constraints. Within this thesis we present a new taxonomy for the computer vision problem of image registration which organizes the field based on the conditions surrounding the problem. From this taxonomy we derive a model which can be used to describe both the conditions surrounding the problem, as well as the range of acceptable solutions. We then use this model to create testbenches which can directly compare image registration algorithms under specific conditions. A direct evaluation of the problem space allows us to interpret models, automatically selecting appropriate algorithms based on how well they perform on similar problems. This selection of an algorithm based on the conditions of the problem mimics the expert knowledge of vision researchers without requiring any knowledge of image registration algorithms. Further, the model identifies the dimensions of the problem space, allowing us to automatically detect different conditions. Extending beyond image registration, we propose a general framework of vision designed to make all vision tasks more accessible by providing a model of vision which allows for the description of what to do without requiring the specification of how the problem is solved. The description of the vision problem itself is represented in such a way that even non-vision experts can understand making the algorithms much more accessible and usable outside of the vision research community.