PDF version of this document
A significant step of any registration algorithm involves the transformation of a set of images which possess commonalities and constrained variations. Unlike existing registration schemes, where a single reference image is chosen for comparison with each of the other images in a pair-wise manner, we take a global scope of the problem. We believe that in order to find the global optimum, we ought to construct an appearance model from all the given images and transform them, one image at a time, while aspiring to minimise the resulting model's complexity. This means that instead of a simple reference image, a richer entity of reference is used to define our target and this entity is dynamic.
It is clear that the model has a representative quality with respect to the images it encapsulates. Undoubtedly it results in a reference that is based on all images, generalisable to all images and should be superior to other reference types. For instance, if a mean reference is used, real data variation is lost and the reference becomes overly blurry - both problems that appearance models avert. The only drawback of the approach we take is that appearance models are quite slow to construct; time is dependent on the size of the set to be registered and yet such constructions form the very heart of the optimisation. In practical terms it means that model re-computation is expected to be frequent and for large numbers of images, the process will slow down considerably. When a multi-scale approach is applied to the model and the images, this discouraging aspect can be partially tackled.