Bi-perspective Fisher discrimination for single depth map upsampling: A self-learning classification-based approach

摘要

Mostly and differently, the recovery of high-resolution (HR) depth map has been demonstrated under the guidance of its corresponding color image. In this paper, without color guidance, we propose a single depth map upsampling algorithm. This algorithm adopts a new Bi-perspective discriminative self-learning approach which turns the HR depth recovery process into a multi-stage classification-based problem. It employs Fisher discriminant criterion over different splitted subspaces and sub-subspaces through a new trilateral decomposition process. This new trilateral decomposition approach utilizes joint weighted low-rank and sparse priors which ensure global and local consistency, respectively. In addition, with the proposed Bi-perspective and multi-stage discriminant classification, the HR construction process is converted from single-map into a 3D collaborative multi-map reconstruction process. Moreover, for more accurate discriminative classification-based behavior, the shared common bases or features between subspaces, that don't contribute basically in the classification, are addressed by a specific shared property learning to keep suitable overlapping consistency between different subspaces. Accordingly, the proposed depth upsampling algorithm shows superior accuracy among most of the state-of-the-art algorithms. The performance is tested by a set of depth maps from different depth sensing systems and with different degradation styles. The proposed algorithm achieves the first rank with a robust performance against TOF degradations. (C) 2019 Elsevier B.V. All rights reserved.