Stereo Refinement Dehazing Network

The performance of stereo vision tasks degrades when haze exists in the input image pair. Independently applying single dehazing algorithm on left and right images is not optimal. To overcome problem, we propose an effective framework, called SRDNet, for simultaneously images. main idea SRDNet to make full use information from cross views improving performance. It does explicitly employ disparity estimation correlation matrix. comprises two parts: a weight-sharing coarse network (WSCDN) guided separated refinement (GSRN). WSCDN utilized predict dehazed Then GSRN introduced residues different by extracting fused separating features with channel spatial module. are added pair so as remove remained haze. Experimental results demonstrate that our proposed surpasses previous methods significant margin both quantitatively qualitatively. Moreover, could be preprocessing step stereo-based 3D object detection boosts accuracy hazy scenes.