Reconstructing a 3-D structure with multiple deformable solid primitives

This paper develops a sophisticated approach to a method for fitting of complex 3-D shapes to unsegmented sets of data. The approach features dynamic models consisting of parametrically defined solid primitives (such as superquadrics), global geometric transformations, global deformations and local finite element deformations. Our technique deals with a deformable model consisting of multiple primitives (parts) that can be used to represent even more complex shapes than are common in the natural world. The deformable model is extended to the case of unsegmented data, incorporating an extended version of a fuzzy clustering technique. The image or range data are transformed into forces acting on the model that deform and conform to the given data set while continually synthesizing nonrigid motion. An adaptive procedure updates a damping factor in accordance with the largest possible step in an implemented iteration process. We demonstrate the fitting process in experiments involving the extraction of the shape of the mouse embryo's organs from microscopic sections.