Are object shape primitives learnable?

Despite extensive efforts focusing on the recovery of shape primitives (or "geons") in the vision community and despite the large body of psychophysical studies strongly suggesting the importance of such representation for the visual system, no model has yet addressed the development of such representations. We propose a developmental architecture, based on the existing experimental evidence on human visual development, and present an algorithm which illustrates how geons (such as cylinder, cone, cube etc.) may be learned. The algorithm learns 2D projections of geons as partial matches between several images. The learning algorithm is based on topographic matching and cross-correlation methods, and uses no a priori knowledge about the structure of geons or any non-local feature. (C) 1999 Elsevier Science B.V. All rights reserved.