Revisiting Single Image Depth Estimation: Toward Higher Resolution Maps with Accurate Object Boundaries

We revisit the problem of estimating depth of a scene from its single RGB image. Despite the recent success of deep learning based methods, we show that there is still room for improvement in two aspects by training a deep network consisting of two sub-networks; a base network for providing an initial depth estimate, and a refinement network for refining it. First, spatial resolution of the estimated depth maps can be improved using skip connections among the sub-networks which are trained in a sequential fashion. Second, we can improve estimation accuracy of boundaries of objects in scenes by employing the proposed loss functions using depth gradients. Experimental results show that the proposed network and methods improve depth estimation performance of baseline networks, particularly for reconstruction of small objects and refinement of distortion of edges, and outperform the state-of-the-art methods on benchmark datasets.