This version of our system only gets about 50% of typical questions asked by people correct. Our near state-of-the-art system is around 60% accuracy, but it requires significantly more processing time per image.
There are a lot of biases in the way people tend to ask questions that it has learned to exploit, so don't expect it to handle unusual questions well.
There are a lot of biases in the way people tend to ask questions that it has learned to exploit, so don't expect it to handle unusual questions well.