From Pixels to Tokens: Revisiting Object Hallucinations in Large Vision-Language Models