Retrieval by shape using multidimensional indexing structures

In modern visual information retrieval systems, visual content is directly addressed by features such as color, texture, shape and spatial relationship. While a large amount of work is being done around the perceptual soundness of models and around the automatic extraction of features, only a limited amount of attention is being placed on the combination of useful representations and similarity models with efficient index structures for shape. In this paper, we discuss retrieval by shape similarity using local features and metric indexing. Shape is partitioned into tokens in correspondence with its protrusions, following curvature analysis. Each token is modeled by a set of perceptually salient attributes, and two distinct distance functions are used to model token similarity and shape similarity. Shape indexing is obtained by arranging shape tokens into a M-tree index structure, suitably modified. Examples from a prototype system are expounded with considerations about the effectiveness of the approach and a comparative performance analysis.