Our existing shape models rely on every landmark location existing on each object we are modelling (though we can deal with some being invisible due to occlusion). This means that the techniques can only deal with classes of objects which do not change dramatically, which would require either more or less landmarks to define their form. For instance, the methods are unable to cope with objects with varying numbers of protrusions, such as trees. Grenander and Miller  describe a statistical model which can deal with varying numbers of objects in images. We would like to investigate this and apply similar techniques to build flexible shape models which can represent objects with differing numbers of sub-parts (for instance, trees with arbitrary numbers of branches). If a point based approach is adopted, statistical models would control both the position of the landmark points and their number. In some cases it may be possible to exploit the self-similarity of the shapes, using results from fractal geometry. An important consideration would be how such a model could be used to locate objects in new images.
To summarise, the intention is to