PDF version of this document
Autonomous construction of robust statistical models has for some time been an arduous goal to achieve. This is particularly due to the need to establish dense correspondence across the training set. Numerous attempts have been made to automate the formation of good active appearance models, but none has yet been very successful. Potential exists, however, in the unification of model construction and image registration. The evaluation of non-rigid registration, which is based on non-linear warps, has been another subject exhibiting great difficulties and automatic selection of good warps is far from trivial.
In active appearance models, the main problem is the inability to select good landmark points without human judgement, as well as the difficulty in location and annotation of these landmarks using brute-force only. Non-rigid registration is a quickly emerging technique that can be used to warp multiple data instances and produce a group-wise optimal model. Contrariwise, past attempts sought a model which is derived from pair-wise registration and therefore depended on an arbitrary choice of a reference.
These arguments highlight the benefits summoned by the combination of these two techniques - active appearance model can aid the selection of good warps in non-rigid registration and the functionality of non-rigid registration can help obtain more compact and robust models of appearance or deformation, as well as diminish the necessity of manual annotation. This report outlines some previous work in the field and a summary of the current successful progress. It also explains some concepts that bear potential or whose realisation can contribute to future endeavours and strengthening of the current algorithm.
Work throughout the year made registration using appearance model practical and powerful. The selection of warps is driven purely by the quality of a model and the warping space itself describes the range of legal deformations. It is safe to state that the goal of automatic construction of deformation models, based on registration, has been comfortably approached.