TRL: Transformer based refinement learning for hybrid-supervised semantic segmentation