Fast Hierarchical Depth Super-Resolution via Guided Attention

Depth maps captured by mainstream depth sensors are still of low resolution compared with color images. The main difficulties in depth super-resolution lie in the recovery of tiny textures from severely undersampled measurements and texture-copy artifacts due to depth-texture inconsistency. To address these problems, we propose a simple and efficient convolutional filtering approach based on guided attention, named HDSRnet-light, for high quality depth super-resolution. In HDSRnet-light, a guided attention scheme is proposed to fuse features of the pyramidal main branch with complementary features from two side-branches associated with the auxiliary high-resolution color image and a bicubic upsampled version of the input depth map. In this way, high-resolution features are progressively recovered from multi-scale information from both the depth map and the color image. Experimental results show that our method achieves state-of-art performance for depth map super-resolution.