A key step in building a combined appearance model is that of establishing a dense correspondence across the set of training images. In practice, this is often achieved by marking up the training set manually with a set of key landmarks and interpolating between them. Recently there has been considerable interest in automating this process. One approach is to use non-rigid registration methods developed for use in medical image analysis, to align the images by optimising a measure of image similarity [14,11]. An alternative approach refines an initial estimate of correspondence so as to code the training set of images as efficiently as possible . We have recently described an approach based on optimising the total description length of the training set, using the model .
In section 4.1 we validate our approach to model evaluation by deliberately perturbing the correspondences in models built using manual annotation to establish correspondence. In section 4.3 we use a method of evaluation to compare models built using non-rigid registration [14,11] and the minimum description length groupwise registration approach of Twining et al.