Interesting: apparently there is now an open visual Q&A training dataset: 254,721 images, 764,163 questions, 9,934,119 answers!