Shape Retrieval Using Hierarchical Total Bregman Soft Clustering

摘要: In this paper, we consider the family of total Bregman divergences (tBDs) as an efficient and robust “distance” measure to quantify dissimilarity between shapes. We use tBD-based l1-norm center representative a set shapes, call it t-center. First, briefly present analyze properties tBDs t-centers following our previous work in [1]. Then, prove that for any tBD, there exists distribution which belongs lifted exponential (lEF) statistical distributions. Further, show finding maximum posteriori (MAP) estimate parameters is equivalent minimizing tBD find t-centers. This leads new clustering technique, namely, soft algorithm. evaluate t-center, algorithm on shape retrieval applications. Our framework composed three steps: 1) extraction boundary points, 2) affine alignment shapes Gaussian mixture model (GMM) [2], [3], [4] represent aligned boundaries, 3) comparison GMMs using best matches given query shape. To further speed up algorithm, perform hierarchical enables us compare with small subset are chosen be cluster method various public domain 2D 3D databases, demonstrate comparable or better results than state-of-the-art techniques.