Quantity Matters: Towards Assessing and Mitigating Number Hallucination in Large Vision-Language Models