Authors: [tex2html_wrap4108]K. Wu, M.D. Levine
Investigator username: levine
Subcategory: computer vision
Object part-based descriptions are interesting because they reflect the natural structures of the real world and support efficient object recognition. In this work, we propose a new model called parametric geons as volumetric primitives for object parts. Parametric geons are seven qualitative shape types regarded as basic volumetric forms in art of sculpture. They are constructed using globally deformed superellipsoid equations which control size and global deformation. Model recovery is achieved by fitting all of the seven models to multiview range data using a stochastic global optimization method (Very Fast Simulated Re-annealing) and then selecting the best model according to the minimum fitting residual. The strength of parametric geons is that it provides a global shape constraint which protects model recovery from the influence of noise and minor shape variations. Thus, volumetric shape approximation to distinct shape types can be robustly obtained. The resulting models can be used as coarse descriptions of object parts for qualitative object recognition.