PDF version of this entire document
Given a collection of images with common properties, it is possible to model visual form (or shape) of the objects therein. A model is a simplified representation of structures that vary across the images, and can be built in a way that makes it independent from subtle changes in view-point, object position, and size []. The models used here are statistical, using the notion of a probability density to describe shape variation.
The model learns about the variation of shape from a collection of shape examples (the training set).
Suppose we have such a set of shapes. The first issue is that we need to remove variation such as position, orientation, and scale of the entire shape which has nothing to do with the actual variation of shape itself. So, we are treating a shape as being the same shape, whatever position it appears in an image, whatever size it happens to be, or whatever its orientation. This is the process of shape alignment.
The second stage then involves constructing a way to represent the variation of the set of aligned shapes, and to do this we need a useful way to represent the shapes themselves.
