The first of these is how mesh representatons of 3D surfaces can be fitted to noisy range data using a bias-variance criterion. The second part of the talk will describe work aimed at building deformable shape models based around such meshes. Here the novel idea is to use a variant of the EM algorithm to match shapes abstracted in terms of triangulated point-sets.
The talk will conclude with some examples furnished by cartographic alignment, estimating facial pose and estimating perspective geometry.