PDF version of this entire document

The Model Construction Algorithm

The registration algorithm was used to establish a set of deformations so that a model can be made aware of variations in the image data. The complexity of this registration algorithm and its type of approach are not as relevant as the results produced for the model to be built from. Cootes led this effort which produced a program that can be used to build a default generative model of the set of deformations. The more accurate the deformations, the better the model will be and here fits my personal role.

The space of this model can be explored and searched. This approach, however, suffers from a couple of drawbacks:

1. The use of a default model will assume that Gaussian distributions are the ones which best fit the pattern of the data. This means that in the process of building the model, there would be a tendency to choose deformations which lead to Gaussian distributions.
2. A computational limitation is introduced because when deformation fields, from which models get built, are changed, their effect on the entire model leads to a problematic situation, as described in [] and also in Chapter 4 on shapes.

Figure: first mode ($\pm2.5$ standard deviations) of an appearance model built automatically by groupwise registration (figure from []).