In this paper, we propose a data-driven approach for automatically estimating a plausible depth map from a single monocular image. Instead of using complicated parametric model, we cast the estimation as a simple yet effective regression problem. We first retrieve semantically similar RGB and depth candidates from database using an activation descriptor. Then, initial estimates are synthesized based on block-matching and robust patch regression. Finally, a weighted median filter (WMF) is adapted to further align depth boundaries to RGB edges. We explicitly take texture-removing technique into consideration for visually plausible results. Experimental results on natural images show that the proposed method outperforms existing approaches in term of both qualitative and quantitative evaluations. © 2016, Society for Imaging Science and Technology (IS&T).