A Higher Order MRF-Model for Stereo-Reconstruction

We consider the task of stereo-reconstruction under the following fairly broad assumptions. A single and continuously shaped object is captured by two uncalibrated cameras. It is assumed, that almost all surface points are binocular visible. We propose a statistical model which represents the surface as a triangular (hexagonal) mesh of pairs of corresponding points. We introduce an iterative scheme, which simultaneously finds an optimal mesh (with respect to a certain Bayes task) and a corresponding optimal fundamental matrix (in a maximum likelihood sense). Thus the surface is reconstructed up to a projective transform.