Depth image interpolation using confidence-based markov random field

Depth images are essential data for high-quality three-dimensional (3D) video services, but the resolution of depth images captured by commercially available depth cameras is lower than that of the corresponding color images, owing to technical limitations. A depth image up-sampling method that uses a confidence-based Markov random field is proposed for enhancing this resolution. An initial high-resolution depth image and confidence values are generated with consideration of boundaries and textures in the corresponding color images. These are used as the base for a new likelihood and prior model design. The energy function derived from this model is optimized by using a graph cut algorithm, and subsequent experiments show that the proposed algorithm provides sufficiently good up-sampled depth images compared to other state-of-the-art algorithms.