Markov Random Field Models for Pose Estimation in Object Recognition

In this paper, we explore theoretical models for pose estimation and object matching based on Markov random elds (MRFs) and the maximum a posteriori (MAP) probability principle. The set of pose estimates as well as matching estimates are considered to be MRFs whose prior distributions are used as the prior constraints. The MAP solution is found from these distributions and an assumed observation model. Two statistical models are derived. The rst model is aimed at pose clustering from corresponding point data. It estimates possibly multiple poses simultaneously and identiies outlier (false) correspondences. The second model attempts to solve matching (correspondence) and pose of 3D objects simultaneously from a 2D image. It gives a parallel and distributed hypothesis-veriication procedure for pose and matching. These models may be used as criteria for evaluating the goodness of matching and pose.