<title>3D object parts inference from range image data</title>

Abstract

We describe an approach for representing an object parts by using its surface curvatures and curve tangent fields. The part representation is based on a set of 12 primitive volumes called geons. The convex edges and the compatibility between the curves guide to infer the geon type. This approach constitutes the first stage of an object recognition systems. In this system, range image data is used as the input, and part-based descriptions are built and matched to 3D object models for recognition. Segmentation and identification of the object parts are based on the RBC theory and 3D properties embedded in the range image. We do not make simplifying assumptions such as the availability of perfect line drawings. Definitions of geometrical constraints are introduced in order to infer the geons from the range image. A method for identifying the parts as one of the twelve 3D part primitives based on differential geometry is then presented. We show that range images are more suitable for geon type recognition than line drawings. The considered features give a unique and natural description to each geon.