Studying Platelet-Based Depth Map Coding

The multi-view plus depth (MVD) format is a promising representation approach for 3D and free viewpoint video systems as it allows synthesizing more views at the decoder than those explicitly coded at the encoder. This format represents each view of a visual scene with both texture (array of color pixels) and a depth map which provides information about the 3D scene geometry. Thus, for the transmission of MVD data, both texture and the associated depth maps need to be efficiently coded. For depth maps, coding artifacts should be minimized (especially at the object boundaries) since they can lead to severe geometric distortions in the virtual intermediate views. A popular approach to code depth maps is the so-called platelet coding scheme which uses piece-wise linear functions and quad-tree decomposition to model the depth information. In this context, this paper studies the performance of plateletbased depth map coding, notably through the analysis of several statistics such as the frequency of each modeling function, the selected block sizes and the bits consumed for the function coefficients. The platelet-based depth map coding rate-distortion (RD) performance is evaluated and compared to other standard based depth map coding solutions. This study is important not only to understand the weaknesses and strengths of this coding solution but also to design new techniques to improve the MVD coding efficiency. Keywords— depth map; quad-tree; platelet coding; multiview plus depth coding;