Seol mar théacs é seo: Looking beyond single images for contrastive semantic segmentation learning