Disparity Estimation in Real-Time 3D Acquisition and Reproduction System

The paper concerns the method of disparity estimation and refinement based on iterative filtration of raw disparity estimate. Raw disparity estimate is obtained by conventional stereomatching methods. In suggested algorithm, 6-8 iterations are sufficient to compute high-quality disparity map, suitable for virtual views rendering, which is essential step of any 3D reproduction system. The method is suitable for highly-parallel processing on modern GPU. In contrast to previous methods of stereo matching implemented on GPU, the proposed approach provides correct disparity computation for “problem areas”, i.e. large uniform areas, occlusion areas and areas with repetitive patterns. Presented method was implemented in a real-time 3D acquisition and reproduction system. For GPU programming, an extension of C programming language provided by NVIDIA Compute Unified Device Architecture (CUDA) was used. Experimental results confirm the usefulness and robustness of the method.