3D视频中的深度视频编码分析.pdfVIP

ABSTRACT With the development of 3D display and interactive multimedia systems, 3D video applications, such as 3DTV and Free Viewpoint TV, are attracting more and more academic and industrial research interests. To efficiently represent the 3D scene, new data format - multi-view video plus depth has been proposed and widely studied. Multi-view video plus depth is composed of texture videos as well as the corresponding depth videos from multiple views. The characteristics of the depth video are very different from those of the conventional texture video. Depth video is composed of large homogeneous areas bounded by sharp edges, which can be defined as piece-wise smooth. In addition, depth video and texture video both represent the same scene, only from a different perspective. Hence, motion similarity and structure similarity can be expected between them. In the conventional video coding, the prediction for the pixels around the edge is generally poor, resulting in large residual errors at these locations. In the transform-based video coding framework, such large residual errors degrade the coding efficiency significantly. Therefore, high-performance depth coding method based on the specific characteristics of the depth video is highly demanded. This paper first thoroughly studies the motion similarity and structure similarity between the depth video and the corresponding texture video. Based on the statistic, a new depth coding framework is proposed, together with some new coding techniques, namely weighted averaging based omnidirectional prediction, depth-texture cooperative clustering based prediction and detection and rectification of the misalignment