Graph-based 3D collaborative disparity map estimation technique

In this paper, a new segment-based stereo matching algorithm is proposed (3DFilt_Gc). Adopting a fast and accurate new energy formulation for the stereo problem in the segment domain makes the proposed 3DFilt_Gc a strong candidate with the top ranked stereo matching techniques according to Middlebury benchmark. Firstly, it introduces a 3D collaborative filter with a new self adapting similarity measure for more reliable initial disparity estimates. Secondly, RANSAC with SVD are employed for a robust disparity plane fitting over the most reliable segments. Finally, a refined accurate disparity map is obtained by the proper disparity plane assignment using graph cuts. In addition, robustness in the performance is shown when it is tested against a blur in one of the stereo images (3DFilt_Gc _Blur) as one of the real image conditions.