Depth-aware Multimodal Question Answering