Adding object detection skills to visual dialogue agents

Our goal is to equip a dialogue agent that asks questions about a visual scene with object detection skills. We take the first steps in this direction within the GuessWhat?! game. We use Mask R-CNN object features as a replacement for ground-truth annotations in the Guesser module, achieving an accuracy of 57.92%. This proves that our system is a viable alternative to the original Guesser, which achieves an accuracy of 62.77% using ground-truth annotations, and thus should be considered an upper bound for our automated system. Crucially, we show that our system exploits the Mask R-CNN object features, in contrast to the original Guesser augmented with global, VGG features. Furthermore, by automating the object detection in GuessWhat?!, we open up a spectrum of opportunities, such as playing the game with new, non-annotated images and using the more granular visual features to condition the other modules of the game architecture.