Depth-based representations: Which coding format for 3D video broadcast applications?

Abstract

3D Video (3DV) delivery standardization is currently ongoing in MPEG. Now time is to choose 3DV data representation format. What is at stake is the final quality for end-users, i.e. synthesized views' visual quality. We focus on two major rival depth-based formats, namely Multiview Video plus Depth (MVD) and Layered Depth Video (LDV). MVD can be considered as the basic depth-based 3DV format, generated by disparity estimation from multiview sequences. LDV is more sophisticated, with the compaction of multiview data into color- and depth-occlusions layers. We compare final views quality using MVD2 and LDV (both containing two color channels plus two depth components) coded with MVC at various compression ratios. Depending on the format, the appropriate synthesis process is performed to generate final stereoscopic pairs. Comparisons are provided in terms of SSIM and PSNR with respect to original views and to synthesized references (obtained without compression). Eventually, LDV outperforms significantly MVD when using state-of-the-art reference synthesis algorithms. Occlusions management before encoding is advantageous in comparison with handling redundant signals at decoder side. Besides, we observe that depth quantization does not induce much loss on the final view quality until a significant degradation level. Improvements in disparity estimation and view synthesis algorithms are therefore still expected during the remaining standardization steps.