Section: New Results

Higher Order Graph Matching

Paticipants: Chaohui Wang, Dimitris Samaras, Nikos Paragios

In [42] a generic framework for sparse and dense graph/3D surface matching has been introduced. The framework is endowed with a novel mathematical formulation regarding the matching process along with a novel deformation model. It exploits the power of invariance of higher order clique potentials and through a low to high resolution approach determines optimal correspondences between two sets of 3D points while taking advantage of Mobius tranformation to measure local similarity of shapes/graphs/surfaces. Graph matching of objects undergoing non-rigid deformations along with temporal 3D surface tracking demonstrated the potentials of our method. Inference is solved through an efficient dual decomposition scema.