Stereoscopic depth refinement by mid-level hypothesis

In this paper we propose a new technique for refinement of depth maps. The technique exploits a stereoscopic pair of images and full-pixel precision disparity map in order to produce the output disparity map with sub-pixel precision. The technique employs view synthesis for verification of hypotheses on disparity values, which are formed to iteratively improve the disparity map. Results of experiments, presented for still images and video sequences, show that proposed technique increases quality of depth and has low computational cost. Applications of the proposed technique can be found in future real-time 3D video broadcasting systems which require fast and robust disparity estimation algorithm with high precision.