PDF version of this document
The existing extensions to shape and appearance models are numerous and their purpose varies. Wavelet compression techniques are used to mitigate the troublesome space requirements (especially in 3-D, e.g. for analysis of brain volumes). These can also make active appearance models far more compact2.35.
There are also some application-specific extensions such as the implementation of view-based models [] and coupled-view models [] for face recognition purposes. The principal idea is that 5 different models can express full appearance irrespective of the wide range of viewing angles around the head. Due to the symmetry of a human head, only 3 models are used in practice (two side views can be mirrored; frontal remains as is). The most appropriate model can then be chosen in a real-time dynamic sequence. The choice of the model to be used depends on the estimated rotation of the head and by estimating that rotation successfully, models do not break down2.36 when introduced with a high degree of freedom (e.g. angular freedom). This idea can undoubtedly be exploited in applications other than faces, but it appears to have a limited demand in industry and it has not been pursued much lately. It is the switching between different models in real-time and the selection of the most suitable model that makes this study challenging and for medical imaging, where the viewing degree of freedom is very limited, this extension is merely irrelevant.
The effect that different facial expressions and aging factors have on statistical model was another intriguing aspect that was mainly pursued by Lanitis et al. [,]. Lanitis has recently worked on synthesis of faces and the analysis of facial attributes [].
For a greater level of detail, as well as information on further extensions and applications, see Appendix cha:Appendix-AElaboration-on.