PDF version of this document
Automatic landmark generation, or more broadly landmark selection, has been an issue of great exploration in the past few years. As one of the ultimate goals of image analysis tasks is complete automation and a precise deterministic approach to selection, older techniques such as manual annotation of an image by experts is a task that ought to be emulated in a reliable way by machine intelligence. Brute-force has been used to enable complex learning tasks for quite some time and investigation points of interest en masse in an image, e.g. lines of high curvature, is certain to lead to automatic annotation of some quality. The level of accuracy of such process, however, which is strongly dependent on the algorithms used, still appears to be a major hindrance. Methods of landmark selection and full automation have been described in [2,3] and more recently a good solution have been discovered by Davies et al. [4,5,6] for landmark selection in statistical shape models.
Problems associated with dimensionality have been pointed out in the literature above. It is vital to ensure that methods work regardless of the number of dimensions dealt with as one of the strengths of manual annotation only becomes apparent when analysing 3-D data. This is primarily due to the impossibility of annotating a large number of slices manually, as in the case of medical imaging.
The rest of Section 2 attempts to objectively explain some of the more fundamental concepts that build up towards the later developments and proposal of new methods.