LAUSR

Abstract Plausible depth prediction from a single monocular image is a challenging task in computer vision. This paper presents a new non-parametric learning-based depth recovery framework in the gradient domain. Specifically, the proposed method leverage a global matching-based depth gradient transfer process to generate meaningful depth gradients from training images. Then, a confidence measure-based depth gradient fusion scheme is designed, which allows us to measure the individual contribution of each pixel in the warped depth gradient maps. In addition, an edge-aware depth gradient refinement strategy is proposed to mitigate the depth gradient outliers and generate accurate depth gradient values. In the end, reliable depth maps can be reconstructed based on these refined depth gradients by using the Poisson solver. Qualitative and quantitative evaluation results on the outdoor and indoor datasets demonstrate that the proposed depth estimation algorithm outperforms other data-driven methods, and is very effective for inferring convincing depth maps.