PDF version of this document
Appearance model, which are further discussed in §4, can be used for interpretation by synthesis. Given a set of images, models can be built that capture the observed variability in the set. Any set of images, from which an appearance model is built, has an important trait: the better-registered the set, the better its model. In other words, when images are aligned properly and look similar, the model itself improves.
Since models as such are generative, they can also be used 'in reverse', synthesising images that they describe. As explained in §6, there is an empirically-justified method for evaluating models, transitively inferring the quality of registration. As a matter of fact, it is our contention that modelling and registration are inherently an identical problem. The remainder of this report explains how models can be evaluated, thereby helping us evaluate non-rigid registration as well.
Looking at our recent work more comprehensively, registration can properly be evaluated by creating models and evaluating these. Furthermore, evaluation of such models can assist registration or even serve as the sole criterion that drives registration. More on this individual aspect of our work is described briefly in §10.