Dynamic Clustering-based Method for Shape Recognition and Retrieval

This paper presents a shape matching framework based on a new shape decomposition approach. A new region-based shape descriptor is proposed to compute the best match between given 2D or 3D shapes. In order to find similar shapes in a database, we first split the interior of each shape into the adequate set of parts, classes, or ellipsoids, then find the corresponding parts between different shapes, and finally compute their similarity. Essentially, we compute the best shape decomposition into k classes using an improved version of the kmeans clustering algorithm without prior fixing of the number of parts. Additionally, we propose a new tool which determines the best ellipsoids packing in order to efficiently represent a shape according to its components. The shape recognition process compares the optimal ellipsoidal partition of the new shape with the different models of a database and extracts the closest shapes. The performances of our shape matching framework are shown through experiments on various data of MPEG-7 and benchmark databases.