PDF version of this entire document

A Generic Method for
Evaluating Appearance Models

Abstract:

Generative models of appearance have been studied extensively as a basis for image interpretation by synthesis. Typically, these models are statistical, learnt from sets of training images. Different methods of representation and training have been proposed, but little attention has been paid to evaluating the resulting models. We propose a method of evaluation that is independent of the form of model, relying only on the generative property. The evaluation is based on measures of model specificity and model generalisation ability. These are calculated from sets of distances between synthetic images generated by the model and those in the training set. We have validated the approach using Active Appearance Models (AAMs) of face and brain images, showing that both measures worsen monotonically as the models are progressively degraded. Finally, we compare three distinct automatic methods of constructing appearance models, and show that we can detect significant differences between them.