PEVL: Position-enhanced Pre-training and Prompt Tuning for Vision-language Models

29956

Knowledge Triples