GuessWhat?! game

The first and main task of this evaluation suite is a guessing game called GuessWhat?!. The game has been first proposed by (DeVries et.al, 2017). It consists of two agents: 1. an Oracle selects a target in a scene and has to answer questions about it; 2. a Questioner generates visually grounded questions aimed at understanding what is the target object in the scene. Due to the recent unavailability of the original dataset, we made it available at the following link: Dropbox link.

CompGuessWhat?! extends the GuessWhat?! dataset to promote the study of attribute-grounded language representations. The original GuessWhat?! dataset is extended with a semantic layer on top of the perceptual layer (i.e., images). This layer consists of a collection of intentional and extensional attributes of the objects in the reference image. Because to associated a scene graph for every image used in GuessWhat?!, we filtered the original dataset by retaining only those images that have a corresponding scene graph. Our subset can be download from the following Dropbox link.

Previous Post Next Post