Self-similarity matching with predictive linear upsampling for depth map

We propose a real-time upsampling scheme for depth maps. The proposed scheme contains two upsampling stages; one is self-similarity matching (SSM), and the other is predictive linear upsampling (PLU). SSM accelerates cost volume filtering by using a variant of joint bilateral upsampling, which utilizes high-dimensional vectors, which is neighborhoods of an RGB image and a depth map. The high-dimensional upsampling suppresses edge blurring and scattering problems. PLU generates smooth surfaces with keeping edges guided by the results of SSM. Experimental results show that the proposed scheme has higher accuracy than the state-of-the-art upsampling. Additionally, the proposed method has real-time performance on a multi-core CPU.