Distilling base-and-meta network with contrastive learning for few-shot semantic segmentation.