A Probabilistic Approach to Robust Shape Matching and Part Decomposition

Abstract

We present a probabilistic approach to shape matching which is invariant to rotation, translation and scaling. Shapes are represented by unlabeled point sets, so discontinuous boundaries and non-boundary points do not pose a problem. Occlusions, significant dissimilarities between shapes and image clutter are explained by a ‘background model’ and hence, their impact on the overall match is limited. By simultaneously learning a part decomposition of both shapes, we are able to successfully match shapes that differ as a result of independent part transformations – a form of variation common amongst real objects of the same class. The effectiveness of the matching algorithm is demonstrated using the benchmark MPEG-7 data set and real images.