PDF version of this entire document
In the previous section, I described how to build statistical models that both describe a training set of shapes, as well as encapsulate information about the variation across this set of shapes.
It is possible to extend this method and construct models that encapsulates not just the variation of the shape of objects in images, but also the variation in the appearance of the object itself.
Appearance models were developed by Edwards et al. []. Their greatest contribution, advantage, and essence lie within the fact that they incorporate textural information rather than shape alone. Texture is a made out of grey-level pixel intensities. Incorporation of full colour is possible as well []. Colour can be simply thought of as an extension of the single grey-scale band. It can be divided into bands using the most common separability: red, green, and blue components 3.2.
A shape model can be thought of as providing very limited information about the appearance of an object within an image, in that it describes the shape of an object, where it is implicitly understood that the shape of an object corresponds to strong edges in an image.
Appearance models describe not only the shape of an object, but the image intensities within the outline of the object as well. In the following subsections, three steps are discussed in turn: modelling shape, modelling intensity, and combined models.
