Representing object manifolds by parametrized SOMs

The recognition and pose estimation of three-dimensional objects is a challenging task that requires suitable object representations. In this paper, we propose the "parametrized self-organizing map" (PSOM) as a flexible method for the generation of appearance-based object models. A PSOM in an eigenspace can be used to extend multiple views of an object to a continuous parametrized manifold that describes the object appearance under various conditions. In a computer vision application the distance from an unknown input to the object specific PSOMs can be used for classification, and the projection on the manifold gives additional informations about additional scene parameters like object pose or illumination direction. We illustrate this concept in a benchmark example that is based on the COIL-20 database which consists of 20 different objects in 72 poses.