Human beings can easily recognise objects, even if they’re not totally visible. For example, when entering a room and seeing the three legs of a table, it makes you realise the fourth is hidden somewhere.
For machines, it is much more complicated than that. In fact, systems of visual recognition need large datasets of annotated images, painstakingly labeled by humans, in order to be able to recognise elements of a scene.
Obtaining this data requires individuals to label every aspect of every object, in each scene in the dataset, which makes it a costly and time-consuming process. As a result, artificial vision systems become limited by the fact that they can’t capture a scene’s overall content.
That’s why DeepMind, a Google subsidiary that specialises in Artificial Intelligence (AI), came up with a computer vision program called “Generative Query Network” (GQN), which doesn’t require such labeled data.
The new algorithm is capable of generating 3D models of objects from 2D images, importantly without any prior human supervision. In fact, this program works only with the data obtained by itself as the machine moves around scenes.
Compared to the usual systems of visual recognition, DeepMind’s algorithm is capable of creating a 3D representation based on images treated with two neural networks: the representation network and the generation network.
The main role of the representation network is to create an abstract description of the scene, using the images taken from different viewpoints. Based on this representation, the scene is predicted from a new, arbitrary viewpoint thanks to the generation network.
For now, GQN only works on simple scenes containing few objects. Yet, the hope of DeepMind is to be able to make machines capable of understanding autonomously their physical environment (which is more complicated), without needing a large amount of pre-set data. Even though the team working on this project is very far from achieving this goal, GQN technology seems very promising.
Mariem, Consultant, Leyton France
Read more on the same subject: