PDF version of this document

Active Appearance Models

Deformable appearance models of shape and intensity are described in a reasonable level of detail by Cootes et al. in [7] and in [8] although they were first introduced by Edwards et al. in [9]. The basic idea was that a measure of similarity between image intensities guided a progressive search. Minimisation of the difference between a couple of images using the sum of a pixel-wise comparison brought the statistical appearance model and its target to convergence¹. During this process, parameters were repetitively being re-evaluated so that they better described the target object - that is the object in the image that resembled the model (see Figure 1). Likewise, free-hand manipulation of these parameters allowed synthesis of new realistic images², meaning that a set of assignments for a collection of parameters $b_{1},$$b_{2},...,$ $b_{n}$ would describe a legal (earthly from an anatomical perspective) type of variation for the model. The following figure shows an appearance model hovering over some target. The model is highlighted in red and labelled 'M'.

Figure 1: Model and target

The following segment expresses in more detail the full process involved. It comprises:

1. Appearance model construction.
2. Correlations learning in active appearance model.
3. Active appearance model search, based on resolved correlations.

The first and the second are concerned with training and learning whereas the third makes use of these previous two. It is somehow possible to learn from experience while searching, but it is not worthy of any further discussion.