Contrastive Learning for Label Efficient Semantic Segmentation