PDF version of this entire document

Using the Results of NRR

Having established a good non-rigid registration, it was time to use it. The final product of this effort was a framework built along with colleagues []. It transformed a registration, encoded as a set of deformation fields, into a model. The main challenge to address was finding transitions between the representation of warps (including triangulation) into a point-to-point correspondence that could be interpreted and used by the existing model construction framework. An example result, which is an automatically-built model, is shown in Figure .

This key development not only proved that models can be built without manual markup, but it also showed that the resultant models are reasonable. It is important to stress that, due to the scale of this problem, no model-based objective function was used in this 2-D case. In principle, however, as demonstrated in 1-D, it should be possible use any reasonable objective function. There is no reason, other than scalability, why the method adopted for model construction ought not to involve model compactness as its key criterion.