SegCLIP: Patch Aggregation with Learnable Centers for Open-Vocabulary Semantic Segmentation