Stereo Domain Translation for Denoising and Super-Resolution Using Correlation Loss

This paper proposes a GAN-based denoising and super-resolution network for stereo images. The proposed network solves the two problems separately in an end-to-end training fashion. A matchability attention module are introduced to compute matching cost spaces and provide the stereo information between generated stereo images. In addition, the correlation loss is proposed to preserve the correspondence between a stereo pair. We evaluate the proposed network using the KITTI 2012 and KITTI 2015 datasets. In addition, we compare with state-of-the-art denoising and super-resolution methods. Experimental results show that the proposed method significantly outperforms the existing method both in terms of qualitative and quantitative analysis.