Filling Kinect Depth Holes Via Position-Guided Matrix Completion.

Zhongyuan Wang,Xiaowei Song,ShiZheng Wang,Jing Xiao,Rui Zhong,Ruimin Hu
DOI: https://doi.org/10.1016/j.neucom.2015.05.146
IF: 6
2016-01-01
Neurocomputing
Abstract:Due to occlusion and measurement errors, there commonly exist holes, inaccurate depth values and noises in depth maps acquired by low-cost Kinect devices. The artifacts seriously affect the practical applicability of depth maps and thus filling holes is a critical pre-processing task for 3D applications. Without the assistance of the accompanied color image, the representative bilateral filtering and inpainting methods hardly provide satisfactory recovery results. Since the depth map containing holes can be naturally regarded as a corrupted low-rank matrix of missing entries, this paper addresses hole filling problem from the perspective of low rank matrix completion. Our method identifies the positions of invalid pixels in hole regions, and then incorporates the known entries into the formulation which considers the low-rank constraint on results and the sparse constraint on residuals. Owning to the well-established principle component pursuit theory, our method substantially boosts the Kinect depth recovery performance in terms of accuracy and reliability.
What problem does this paper attempt to address?