Graph-based point drift: Graph centrality on the registration of point-sets

The problem of point-set registration often arises in Pattern Recognition whenever one needs to match information available in images, such as feature locations, landmarks, or points representing a surface of an object. It is a challenging task and a widely explored topic in stereo vision, image alignment, medical imaging, and other fields. Many of those problems have been addressed using graph theory by taking advantage of the structural information available in graphs. In this paper, graph centralities are explored in the point-set registration problem for the first time. We propose a variant of the Coherent Point Drift (CPD) by integrating the degree, betweenness, closeness, eigenvector, and pagerank centralities. The centrality values bring topological information used during the computation of correspondence between points. We analyse the performance on several datasets and our results indicate that the registration can converge faster when the centrality is combined with the spatial information in the traditional probabilistic framework. Our novel contribution introduces the social network centralities as a good source of prior information for the registration problem and it demonstrates how one can take advantage of such information.