Feature-based depth refinement for view synthesis

Chao Yao, Yao Zhao, Chunyu Lin, Jingyu Yang

Abstract: The accuracy of depth map limited the performance of Depth-Image-Based-Rendering (DIBR) system, especially the misalignment of objects between the wrapped image and the target image which locates on the real viewpoint would seriously decreases the quality of the wrapped image. In practice, the misalignment is a major cause that the objective quality is not enough, comparing the wrapped image and the target image. In this paper, we propose a new depth refinement method which aims to align the wrapped positions and the real target positions. More specifically, to align the wrapped image, some features extracted from texture images are used. Then, based on the matched features, an energy function is minimized the depth errors. This energy function is efficient enough to provide high quality rendered image, even comparing with the real target image. Experiments show that the proposed method provides visually pleasing results and high objective quality.