A Non-local Low-rank and Sparsity based Framework for Depth Map Inpainting

Depth is an important cue along with RGB data in many computer vision applications. Depth maps captured by most of the mainstream depth sensors are noisy with a lot of missing depth information causing annoying visual artifacts. In this work, we proposed a joint framework for color guided depth map inpainting. We exploit the redundancy of RGBD data by integrating non-local low-rank patch regularization along with local structural information in an optimization framework. The proposed non-local patch-based regularization prior in addition to a complementary local smoothness prior facilitate robust recovery of the sharp depth discontinuities and the missing depth information.