M. S. Lew, N. Sebe, and J. P. Eakins (Eds.): CIVR 2002, LNCS 2383, pp. 225-234, 2002.  Springer-Verlag Berlin Heidelberg 2002 Spin Images and Neural Networks for Efficient Content-Based Retrieval in 3D Object Databases Pedro A. de Alarcn, Alberto D. Pascual-Montano, and JosØ M. Carazo Biocomputing Unit. Centro Nacional de Biotecnologa (CSIC) Campus Universidad Autnoma de Madrid, 28049 Madrid, Spain {pedro,pascual,carazo}@cnb.uam.es http://www.biocomp.cnb.uam.es Abstract. We describe a system for querying 3D model databases using the spin image representation as a shape signature for objects depicted as triangular meshes. The spin image representation facilitates the task of aligning the query object with respect to matched models (coarse- grain registration). The main contribution of this work is the introduction of a three-level indexing schema based on artificial neural networks. The indexing schema improves significantly the efficiency in matching query spin images against those stored in the database. Our results are suitable for content-based retrieval in 3D general object databases. A particular application to molecular databases is also presented. 1 Introduction Retrieving objects by their content as opposed to keyword indexing or simple browsing has become an important operation, and consequently, an active field of research. After more than a decade of intensive research, content-based image retrieval (CBIR) technology moved out of the laboratory and into the marketplace, in the form of commercial products like QBIC [1] and Virage [2]. CBIR draws many of its methods from the field of image processing, computer vision, pattern recognition, and database technology. As web-based repositories of multi-dimensional scientific data continue to grow, so does the need for content-based retrieval of three-dimensional (3D) scalar data. One of the most prevalent source of volumetric data is medical and biological imaging. During the last decade bioscientists have witnessed an spectacular growth of molecular databases. Molecular databases store in some cases (Protein Data Bank [3]) thousands of molecular complexes described as three dimensional datasets. Public access to these repositories boosts research targeted at the discovery of new drugs and medicines. Certainly, content-based retrieval would provide scientists with a valuable tool that facilitates, for example, the task of finding molecules that are structurally similar to a given one. To do so, different features are used to represent the content of