Depth map compression for 3DTV services

Abstract

Advances in technology have contributed to the development of new applications such as 3D Television (3DTV) and Free Viewpoint Video (FVV), which have recently started to attract interest within the marketplace. The multiview video plus depth (MVD) format is generally employed by these technologies to obtain a compact representation of data. This is possible, because depth maps enable the rendering of videos on additional intermediate views at the receiver. However, the quality of the rendered views is significantly dependent on the quality of the depth map available at the decoder. Traditional image and video compression standards produce artefacts along depth discontinuities, which negatively affect the Quality of Experience (QoE) when rendering virtual views in 3DTV and FVV applications. Their poor performance is attributed to the different characteristics of depth and natural images. The research community has recently identified the need to design depth map compression algorithms which preserves depth discontinuities. In this dissertation, a depth coding algorithm was developed which exploits the correlation between texture and depth video sequences to predict depth discontinuities. The aim is to minimise the side information which is generally required by a number of depth map encoders to preserve edges. In this manner, compression efficiency is improved while still maintaining acceptable levels of quality. This was accomplished using a number of techniques which include contour detection by surround suppression, spatial and temporal predictors and depth estimation. The latter is computationally expensive and its implementation depends on the resources available. The concept of displacement prediction was also introduced. These are used by the algorithm to predict edge blocks. The algorithm displays good performance, achieving a PSNR gain in terms of depth quality of around 4.9 - 6.6 dB relative to JPEG and outperforms the Zanuttigh-Cortelazzo compression scheme by around 3 dB. Moreover, the proposed predictions improve the original system were the lower bound was shifted to 0.37 bpp from the 0.39 bpp previously required in terms of rendered quality.