Exploiting Low Rank Prior for Depth Map Completion

Abstract

Occurrence of regions with missing data in depth maps either captured by active sensors or estimated by different passive computer vision algorithms, is unavoidable due to several reasons. The task of depth inpainting from a single degraded depth map is more challenging as compared to using multiple depth observations or RGB-D data. Recently, low rank techniques have become popular and shown supremacy over several state-of-the-art techniques for image deblurring, denoising, upsampling, etc. Since completion of missing regions in a given degraded depth observation is a severely ill-posed problem, low rank property of the inpainted depth map can be posed as the regularization constraint. We perform several experiments to show the superiority of the proposed method over the state-of-the-art depth inpainting techniques.