Open-Vocabulary Semantic Segmentation with Mask-adapted CLIP