Contrastive Learning for Label-Efficient Semantic Segmentation