Models of appearance have been presented in this chapter. An associated algorithm, the active appearance model allows to search for instances of faces in an image.
The active appearance model search algorithm is a local optimiser, and is able to match well if initialised sufficiently close to the true optima. In sequences of faces, the difference between frames are usually sufficiently small that using the match to the previous frame as initialisation for the search in the next for satisfying tracking. However, large movements cause the search to fail. It can be recovered by using a global search.
The appearance model is generative. We can synthesise an image of a face given its appearance parameter vector.
The next chapter describes some of the video sequences of faces successfully tracked with the active appearance model. The video sequences constitute the training data our model should mimic.