Image-to-Image Matching via Foundation Models: A New Perspective for Open-Vocabulary Semantic Segmentation