Sparse-to-Dense Depth Reconstruction using Non-Convex Optimization

Depth reconstruction aims at recovering accurate depth maps from sparse measurements. Recent works reconstruct depth maps based on the assumption that neighboring pixels have similar depth when they have similar color values. However, such methods tend to blur depth boundaries due to color bleeding. We address this problem by a non-convex penalty. First, we formulate depth recovery problem as a non-convex optimization problem in which depth boundaries are preserved by an L-1-L-2 penalty. Second, we solve the problem based on iteratively reweighted least squares. Numerical experiments demonstrate that our method outperforms state-of-the-art algorithms in terms of PSNR and visual quality.