Depth image super-resolution via semi self-taught learning framework

Depth images have recently attracted much attention in computer vision and high-quality 3D content for 3DTV and 3D movies. In this paper, we present a new semi self-taught learning application framework for enhancing resolution of depth maps without making use of ancillary color images data at the target resolution, or multiple aligned depth maps. Our framework consists of cascade random forests reaching from coarse to fine results. We learn the surface information and structure transformations both from a small high-quality depth exemplars and the input depth map itself across different scales. Considering that edge plays an important role in depth map quality, we optimize an effective regularized objective that calculates on output image space and input edge space in random forests. Experiments show the effectiveness and superiority of our method against other techniques with or without applying aligned RGB information