Deep Learning for Visual Question Answering

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

How do these results compare to FB AI research results that we re recently announced on the same topic?