Super-resolution domain adaptation networks for semantic segmentation via pixel and output level aligning