Comprehending an image involves more than recognising what objects or entities are within it, but recognising events, relationships, and context from the image. This problem requires both sophisticated image recognition, language, world-modelling, and "image comprehension". There are several datasets in use. The VQA datasets, were generated by asking Amazon Mechanical Turk workers to propose questions about photos from Microsoft's COCO image collection. (Source)
Concretely, this involves questions like looking at an image of a pizza and identifying if it is vegetarian, or how many slices it has been cut into.
Human performance on this dataset is currently ~83%, so it is quite challenging!
The question resolution date will be set to the earliest of the following dates:
*Publication date of a credible paper, blog-post, video or similar demonstrating an AI with performance >=83.00%
*A credible paper, blog-post, video or similar, referencing a date earlier than its publication date, by which the feat had been achieved (similar to how DeepMind kept AlphaGo's victory over European champion Fan Hui secret from October 2015 to January 2016, in order to coincide with the publication of the corresponding Nature paper).
Data
Data on previous performance on this benchmark can be found here: https://www.eff.org/ai/metrics#Visual-Questio… (data in table "COCO Visual Question Answering (VQA) real images 2.0 open ended") as well as on the VQA challenge site leaderboard (https://evalai.cloudcv.org/web/challenges/cha…).
You might improve your forecasts by also gathering data from other benchmarks for visual question answering. There is more data on performance on the preceding VQA1.0 dataset. This was eventually superseded due to language biases inflating performance metrics. VQA2.0 is harder as for most questions (e.g. "Is the child riding a bike?") it adds another similar image that has a different answer (e.g. a child sitting next to a bike).
In addition, the Visual7W dataset, which can also be found in the EFF dataset, is based on the same underlying dataset of images (Microsoft COCO) but provides richer questions and longer answers than VQA (see https://github.com/yukezhu/visual7w-toolkit).