Indexing and retrieval of 3D models by unsupervised clustering with hierarchical SOM

A hierarchical indexing structure for 3D model retrieval based on the Hierarchical Self Organizing Map (HSOM) is proposed. The proposed approach organizes the database into a hierarchy so that head models are partitioned by coarse features initially and finer scale features are used in lower levels. The aim is to traverse a small subset of the database during retrieval. This is made possible by exploiting the multi-resolution capability of spherical wavelet features to successively approximate the salient characteristics of the head models, which are encoded in the form of weight vectors associated with the nodes at different levels (from coarse to fine) of the HSOM. To avoid premature commitment to a possibly erroneous model class, search is propagated from a subset of nodes at each level, which is selected based on a fuzzy membership measure between the query feature vector and weight vector, instead of taking the winner-take-all approach. Experiments show that, in addition to efficiency improvement, model retrieval based on the HSOM approach is able to achieve a much higher accuracy compared with the case where no indexing is performed.