Reducing computation redundancy for high-efficiency view synthesis

Abstract

To facilitate the realization of free-viewpoint 3D video systems, view synthesis and disparity matching are two of the most significant operations. However, these two operations demand high computation complexity which motivates the development of the proposed techniques. In this paper, we propose a shape-adaptive low-complexity (SALC) technique for removing spatial computation redundancy of view synthesis. The proposed idea takes advantage of that depth values of pixels inside the same object are either the same or gracefully changed, which implies that the operations of view synthesis may be reused, and are not necessarily computed pixel by pixel. Instead, the pixels with the same depth value should be treated as a group which becomes the basic unit in computing view synthesis. The proposed technique can cooperate with existing works to further enhance their performance. From the experimental results, the proposed SALC technique saves as much as 96% computation complexity for view synthesis compared with original version with even better visual quality than previous works.