Evaluation and Enhancement of Semantic Grounding in Large Vision-Language Models