PDF version of this document
Non-rigid registration (NRR) and model-based image analysis were previously believed to possess some commonality - a premise on which current research is still based. There is a growing belief that the best of both can be exploited to construct a unified framework. This modified framework might be more robust and offer higher utility and functionality when compared with the other two approaches working solely.
It is claimed that warping of the images, as is already done in medical imaging in particular, can be used to find correspondences that are optimal in some respects. Pair-wise (and sometimes group-wise) image registration using non-rigid transformations was the way in which previous (local) research by Smith had planned to build good models of appearance. Furthermore, non-rigid registration made it possible to achieve better correspondence in images and maybe supercede other methods. By constructing models from the transformation parameters, one could also highlight and describe successful registration trajectories, i.e. ideal warping sequences or a legal range of warps. As a result of the process in its entirety, active appearance models could be constructed automatically (for identification of correspondence no longer requires any human intervention) and non-rigid registration could guided by appearance models rather than similarity measures which are pair-wise.