Occlusion size aware multi-viewpoint images generation from 2D plusdepth images

摘要

This paper proposes a new scheme to synthesize multi-view images from 2D plus depth (2D+D) images. A 2D+D image is combined from one single image with its corresponding depth map. The 2D+D image format is one of the most widely used representations of 3D images. To synthesize a new viewpoint image, we perform procedures of calculating the disparity first, and then filling disocclusion regions. For the first step, we calculate the disparity of new images by a geometric model which defines the geometric relation between the scene and synthesized viewpoints. After obtaining new images, we have to fill disocclusion regions. To fill the regions, several methods have been proposed, e.g. (1) linear interpolation of foreground and background image color, background color extrapolation; (2) mirroring of background color information; (3) preprocessing of depth information. To fill the regions with better quality, we classify regions into two different types. Bi-cubic interpolation is then applied while the size of absence region is small. If the region is large, we fill the proper texture by searching in the neighboring area. The described processes improve the visual quality of synthesized multi-view images. Finally this paper presents the comparison between the proposed method and traditional methods.