Abstract

In this paper, we present a novel framework for findingthe kinematic structure correspondence between two objectsin videos via hypergraph matching. In contrast to priorappearance and graph alignment based matching methodswhich have been applied among two similar static images,the proposed method finds correspondences between twodynamic kinematic structures of heterogeneous objects invideos. Our main contributions can be summarised as fol-lows: (i) casting the kinematic structure correspondenceproblem into a hypergraph matching problem, incorporat-ing multi-order similarities with normalising weights, (ii)a structural topology similarity measure by a new topology constrained subgraph isomorphism aggregation, (iii) a kinematic correlation measure between pairwise nodes, and (iv) a combinatorial local motion similarity measure using geodesic distance on the Riemannian manifold. We demonstrate the robustness and accuracy of our method through anumber of experiments on complex articulated synthetic and real data.