PDF version of this entire document
It is important to note that if the only task was representing a set of shapes to some required degree of accuracy, we could select a dense set of landmarks on each shape, with no correspondence between landmarks on different shapes. This however would perform poorly if we used these landmarks to try and describe the variation among shapes. The simplest way of obtaining meaningful corresponding landmarks across a set of shapes is for a human expert to annotate the training shapes with the positions of equivalent points using some computerised special-purpose tools. In recent years, alternatives which are automatic showed great promise [] and they were also extended to 3-D []. A subsequent background chapter on MDL shape models is dedicated purely to that one strand of work, so I will not discuss this further here.
