Depth Helps: Improving Pre-trained RGB-based Policy with Depth Information Injection