PDF version of this entire document
Statistical models of shape and appearance have been used extensively as a basis for image interpretation by synthesis [,]. The basic idea is that, given a dense correspondence between the set of images of the same class of structures, it is possible to build coupled statistical models of the shape variation and shape-free texture variation across the set1.1. Given the generative `appearance model', it is possible to interpret new images of the same type by finding an instance of the model that matches the image as closely as possible. There are efficient algorithms that achieve this by finding a best match [,,].
