Hierarchical Edge Refinement Network for Guided Depth Map Super-Resolution

Abstract

The Guided Depth map Super-Resolution (GDSR) task aims to solve the low-quality problem of depth maps obtained from different devices. The most difficult challenge for the GDSR task is obtaining consistent depth discontinuity features from the depth map and the corresponding color image. This paper proposes a novel learning-based hierarchical network for GDSR task with the Edge-guided Bidirectional Interaction Module (EBIM). First, the EBIM refines the edge information by recalibrating the color details and selecting the consistent edge feature from information flow of the depth feature to color features. Second, EBIM enhances the depth discontinuity information details in the depth branch by transferring the refined edge information from the color branch. In addition, we use hierarchical edge constraints to intensifier the refined edge information through the EBIM module. Experimental results on existing public datasets show that our method outperforms other state-of-the-art methods in different measures.