A depth optimization method for 2D-to-3D conversion based on RGB-D images

2D-to-3D conversion is one of the key technologies in stereoscopic video. Its visual qualities depend on the qualities of depth image estimation and virtual viewpoint image rendering. Although there are many methods to get depth information, the depth information still has some errors because of calculation errors and noises, which influences the visual qualities of the generated virtual viewpoint image. This paper proposes a depth image optimization method to improve the qualities of virtual viewpoint image. It utilizes the surface normal information to segment objects and obtain the smooth factor of depth image. The experimental results demonstrate that our method can denoise and smooth the original depth map effectively.