Object recognition with structural descriptions and deformable models

This paper presents first results for an object recognition system based on a viewer centered approach. Object models are derived from a self-organizi ng neural net using a variant of the Neural Gas algorithm. Neurons are gene rated and shifted according to spatial locations of geometrical features. A s a result, a graph oriented representation of the neurons' positions and t heir interconnections reflect the structure of the model objects. The match ing process utilizes a deformation strategy to adapt the topology of the mo del graphs to the features in a presented scene. The amount of deformation for the resulting graphs together with an analysis of gray value moments in dicate the objects shown in the scene. (C) 2000 Elsevier Science B.V. All r ights reserved.