annotations cvpr cvpr2021 dataset multimodal multimodal-deep-learning paper traffic-events video-qa video-reasoning vqa vqa-dataset