Prior surface model guided dense reconstruction via graph-cuts

A fronto-parallel surface model is implied by the conventional smoothing constraints used in most stereo vision algorithms. It results in systematic errors in natural scene reconstruction. In this paper, we propose a novel method incorporating a non- planar surface obtained through sparse stereo reconstruction as the prior model. The depth map is recovered by solving a multi-label Markov Random Filed (MRF) optimization via graph-cuts. Our energy minimization formulation exploits prior information by incorporating surface expectation in la- bel definition. It results in a depth map with piecewise con- tinuity along the prior surface while maintains efficiency of original algorithm. Moreover, with the aid of the prior surface model, our method can directly generate an approximate 3D patch in depth refinement which is a more accurate description of the correlation region. We demonstrate our algorithm on real scene and show the promising experimental results.