Segmenting 3D Objects into Geons

We propose a new approach to 3D object segmentation and description. Beginning with multiview range images of a 3D object, we segment the object into parts at deep surface concavities. Motivated by physics, we detect these concavities by locating surface points where the simulated electrical charge density achieves a local minimum. The individual parts are then described by parametric geons. The latter are de-ned as seven distinctive volumetric shapes characterized by constrained superellipsoids, with deformation parameters controlling the tapering and bending. We obtain a unique part model by tting all parametric geons to each part and classifying the tting residuals. The advantage of our classiication approach is its ability to approximate the shape of an object, even one not composed of perfect geon-like parts. The resulting 3D shape description is a prerequisite for generic object recognition.