The behaviour of objects in the natural world is rarely static. Cheeks bulge and stretch, body parts twist and bend, trees sway in the breeze. Visual models used to represent such objects should be able to reflect such behaviour. Clearly we need models that can accomodate deformation, nonplanarity, inexact symmetry and a whole range of localized irregularities. There is a need for a hybrid representation which can bridge the gap between the sometimes conflicting requirements imposed by reconstruction and recognition.
Consideration of the above leads to a physically based modeling framework for shape and motion reconstruction of flexible objects from their images. Our approach is based on the finite element method (FEM) and parametric solid modelling using implicit functions. One can liken this representation to the force/process metaphor of modeling clay, i.e. shape can be thought of as the result of pushing, pinching and pulling on a lump of elastic material. As Pentland has shown the resulting FEM paramaters can serve as useful descriptors of 3-D shape provided that these parameters can be normalized with respect to reference frame. Our strategy is based on a two-stage approach in which a part-oriented description of an object is first computed using superellipsoid primitives. This serves as a coarse characterization of shape from which a more refined description is computed using deformable models. The basic idea is to use the superellipsoid as a starting point and then calculate the deformation parameters which result in a best fit of the model surface to data. These additional parameters serve to further qualify the shape of each part where more detailed shape information is required, e.g. for purposes of recognition. The resulting shape modeler has recently been incorporated as part of our Autonomous Exploration system. As information is acquired, fused, interpreted, and parameterized using superellipsoids, the modeler asynchronously produces a more refined description using deformable solid primitives. The resulting description is currently used for applications in recognition and visualization (e.g. virtual reality).
S. Roymoulik, F.P. Ferrie
Evolution: Raw Data to Articulated Model