Occlusion handling in depth estimation from multiview video

摘要

This paper presents a novel approach to occlusion handling problem in depth estimation using three views. A solution based on modification of similarity cost function in optimization algorithm (on example of graph cuts) is proposed. During depth estimation via optimization algorithms like graph cuts similarity cost function is constantly updated so that only non-occluded pixels in the side views are considered. For the side views, virtual depth maps are synthesized and occluded regions are detected. Basing on that, similarity cost function is updated for correspondence search only in non-occluded regions of the side views. The experimental results, performed with use of a well-known 3D video test sequences, show that the proposed approach in application for virtual view synthesis for next generation of 3D-television, provides gains of about 1.25 dB of PSNR related to the state-of-the-art technique implemented in MPEG Depth Estimation Reference Software.