3D model based stereo reconstruction using coupled Markov random fields

Projet SYNTIMA lot of methods have been proposed in the field of stereo-reconstruction. We address here the problem of model-based tridimensional reconstruction from an alreadysegmented and matched stereo pair. This work is a continuation of the work presented, concerning the reconstruction problem. We study here a method based on Markov random fields, which allows the a priori segmentation and matching to be refined during the reconstruction of the 3D surfaces. A new segmentation and matching is then produced which respects the 3D coherence (or equivalently the disparity coherence) of each segmented region-pair. In this first approach, we use simple segmentation energies for each image (without line processes), plus a coupling term between the left and right images, associating planes (as surface primitives) with each region pair. This is the justification for using coupled Markov random fields. We present results on synthetic and real images. These preliminary results allow us to assess the feasability of a hierarchical stereo reconstruction method with no a priori segmentation