Depth maps captured by modern depth cameras such as Kinect and Time-of-Flight (ToF) are usually contaminated by missing data, noises and suffer from being of low resolution. In this paper, we present a robust method for high-quality restoration of a degraded depth map with the guidance of the corresponding color image. We solve the problem in an energy optimization framework that consists of a novel robust data term and smoothness term. To accommodate not only the noise but also the inconsistency between depth discontinuities and the color edges, we model both the data term and smoothness term with a robust exponential error norm function. We propose to use Iteratively Re-weighted Least Squares (IRLS) methods for efficiently solving the resulting highly non-convex optimization problem. More importantly, we further develop a data-driven adaptive parameter selection scheme to properly determine the parameter in the model. We show that the proposed approach can preserve fine details and sharp depth discontinuities even for a large upsampling factor ($8\times$ for example). Experimental results on both simulated and real datasets demonstrate that the proposed method outperforms recent state-of-the-art methods in coping with the heavy noise, preserving sharp depth discontinuities and suppressing the texture copy artifacts.