Can Large Vision-Language Models Correct Grounding Errors and Reason By Themselves?