Coarse-to-fine Face Depth Super-Resolution with Attentive Feature Selection

The application of face depth maps is promising but limited due to the multiple degradations introduced by depth sensors, such as low spatial resolution, noise and blurry edges. In this paper, we propose a novel coarse-to-fine framework that progressively denoises and super-resolves face depth maps with two stages. The coarse stage consists of a denoising sub-network and an edge inference sub-network, and it recovers a denoised coarse depth map and an edge prior to assist the following depth refinement. The refinement stage consists of multiple attentive feature selection and fusion (AFSF) blocks that can enrich the feature diversity and aggregate important features selectively. Moreover, a residual learning scheme is used in the refinement stage to enhance the details. Extensive experiments on both synthetic and real-world face depth datasets demonstrated the superiority of our method over several state-of-the-art methods.