In this contest, competitors will design systems to learn about two modalities of data: images and text. The provided training data is Louis von Ahn's Small ESP Game Dataset, containing images and word tags for these images. Competitors should train their system to associate images to sets of word tags. At test time, the system is presented with two possible sets of word tags for an image, and must determine which is the correct set of word tags.
Because this task is very easy for humans to do, we will not provide the final test inputs until one week before the contest closes. Preliminary winners will need to release their winning code and demonstrate that they did not manually label the test set. We reserve the right to disqualify entries that may involve any manual labeling of the test set.