PDF version of this document

Open Questions

Active appearance models are a powerful method of interpreting and synthesising^2.37 images. Nevertheless, they are heavy, complex and they require a long time to train. Active appearance models sometimes serve a purpose which is different from that of active shape models and often they require more time to reach good convergence, mainly due to their additional complexity. In that sense, some implementation issues in appearance models need to be addressed; this can hopefully make them very powerful in more aspects. Furthermore, the accuracy of appearance models is sometimes lower^2.38 than that which is offered by other methods. If synthesis of photo-realistic images is a pre-requisite of the model to be used, then AAM's are indeed a unique and sensational technology that does the job adequately.

An additional valid critique of AAM's speaks of its occasional failure to reach the global minimum when posed with the goal of fitting. It is still not immune to large initial displacements (and hence discrepancies) or target instances that deviate abnormally from the training set. Since AAM's still rely on a good initial placement in a given target, there are possibly pressing issues to be looked at.

It is yet hard to ignore the fact that results of an AAM fitting are sometimes less accurate than those of an ASM^2.39. This brings up the doubts as for whether the extra complexity associated with texture is worthy of being considered. The investment of time and intensive effort, including the need for human intervention, raises some important doubts and scepticism.

A significant drawback that is associated with appearance models is that for automation of model construction, landmark selection [,], or more fundamentally image correspondence [], is necessary and yet somewhat difficult to achieve. It is not obvious how to choose landmarks sensibly and how to judge the optimality of an automatic choice of significant points. Since the efficiency of an appearance model depends greatly on the textures embedded in that model, it is not sufficient to use existing techniques to select landmarks and pseudo-landmarks (additional points between the original anatomical or mathematical landmarks), as quite recently suggested by Davies et al. []. A further explanation of this work is spread throughout some of the following chapters, but primarily Chapter needs to be fully grasped for understanding of this undertaken project and its manifestation.