PDF version of this document

A Generic Method for Evaluating Appearance Models

First Author
Institution1
Institution1 address
firstauthor@i1.org - Second Author
Institution2
First line of institution2 address
http://www.author.org/~second

Abstract:

Generative models of appearance have been studied extensively as a basis for image interpretation by synthesis. Typically, these models are learnt from sets of training images and are statistical in nature. 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. We define the specificity and generalisation ability of a model in terms of distances between synthetic images generated by the model and those in the training set. We validate the approach, using Active Appearance Models (AAMs) of face and brain images, and show that specificity and generalisation degrade monotonically as the models are progressively degraded. We compare two different inter-image distance metrics, and show that shuffle distance performs better than Euclidean distance. We then compare three different automatic methods of constructing appearance models, and show that we can detect significant differences between them. Finally, we contend that model construction is analogous to the task of non-rigid registration. The former requires correspondence across images, whereas the other attain to find that correspondence. We then compare our method against another method that is based on ground truth and show that both are in tight agreement.