Establishing correspondences between different images is one of the key
problems in computer vision. Two classical problems, where this task plays a major role are optic flow computation and stereo reconstruction. While correspondences between two consecutive images in time (optic flow) are not subject to any restrictions, correspondences between two images that differ by location (stereo reconstruction) have to fullfill the geometry of two views (epipolar constraint).
In this talk we show how stereo reconstruction methods can benefit from recent progress in optic flow computation. To this end, we propose a novel variational model that is based on the currently most accurate optic flow technique (Brox et al. 2004). By incorporating the epipolar constraint in the estimation, the proposed model allows to transfer a variety of advantages to the field of variational stereo reconstruction: (i) Robustness under noise. (ii) Preservaton of discontinuites in the solution. (iii) A theoretically justified minimsation strategy.
Experiments with both synthetic and real-world data show the excellent performance and the noise robustness of our approach.