3D reconstruction with depth prior using graph-cut

In this paper, we propose a novel graph-cut based 3D reconstruction method which is able to take into account partially available depth data as a prior. We formulate the energy as a function written as a sum of terms that can be minimized in two different representations: (1) assignment-based, which yields a standard binary energy, in this approach occlusion and uniqueness is handled naturally; as well as (2) a multi-label representation which yields to a non-binary energy where occlusion is only handled by assigning a special label to occluded pixels, while uniqueness is not handled explicitly. Both representations have its advantages and disadvantages, which are analyzed and discussed in detail through various experimental results on the publicly available Middlebury stereo datasets and on real stereo images. Results show, that the use of depth prior information from different sources produces better 3D reconstructions.