Local segment-based dense depth reconstruction from very sparsely sampled data

Abstract

In this paper, we propose two relatively simplistic and efficient methods for depth reconstruction from very sparsely sampled random depth data. Both the proposed approaches exploit the segmentation cue from a registered colour image of the same scene. The first approach which we term as plane fitting depth reconstruction (PFitDR), involves cost computations on plane-fitted depth values over local segments. The second approach, which we call median filled depth reconstruction (MFillDR) is an even simpler method, wherein the reconstruction is carried out using computation of median of depth values over local segments. We demonstrate dense reconstruction from very less number of available depth points (e.g. as low as 1%). Our methods favorably compare with a recent related state-of-the-art method, both qualitatively as well as quantitatively in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indices. � 2017 IEEE.