S. Roymoulik, F.P. Ferrie
Earlier research in our laboratory considered the representation and inference of coarse shape descriptors from both static and dynamic images. This work resulted in the Autonomous Exploration framework (Whaite and Ferrie) which is our current basis for computing environmental descriptions from sensor data. It successfully demonstrated how visual feedback could be used to stabalize the inference of model parameters (e.g. superquadrics) to the point where they could serve as a basis for object recognition. However the limited expressiveness of these models becomes a liability when the recognition process must discriminate between objects which are similar in appearance. The goal of this project is to consider richer classes of description within the automous exploration framework, in particular the use of physics-based models. Our approach borrows from Pentland and Terzopoulos in the use of the finite element method (FEM). One can linen 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. Pentland and Sclaroff have shown that the resulting FEM paramaters can serve as useful descriptors of 3-D shape provided that they can be normalized with respect to reference frame, a by-product of the autonomous exploration process. We have incorporated Pentland and Sclaroff's deformable model implementation if the autonomous exploration testbed and have confirmed that the system is capable of automatically determining scene descriptions that are accurate, robust, and stable. We are currently investigating applications of the modeller as part of the visual modeling project (IRIS MSA-2).
[Ferrie_bend.eps].75 Evolution: Raw Data to Articulated Modelferrie:figure7