PDF version of this entire document
A very key step in construction of combined appearance models is that of identifying dense correspondence across a given set of training images. This is often achieved by marking up the training set by hand, simply identifying significant points in the images and interpolating between these points. In recent years, automation of this process was a problem of great interest. Denser correspondence, which is also accurate, builds a better model. However, that dense correspondence is arduous to obtain. In 3-D, identification of correspondences is hard to obtain objectively. More points of correspondence must be identified as well.
One approach to solving this problem automatically is to use NRR and bring the images to alignment by optimising a similarity measure [11,17]. A different approach refines initial estimates of the correspondence so as to code the set of images in themost efficient way [1]. We have recently outlined an approach which is based on optimising the total description length of the training set, using its model [25]. A model will be most concise when its training set is fully correspondent.
In Section IV our approach is validated by deliberately perturbing the correspondence in models, i.e. decreasing the registration. Such models were built using manual annotation that establishes a reliable correspondence. In Section V our approach is used to compare common registrations methods [11,17], as well as our minimum description length approach.