Google DeepMind unveils an AI (artificial intelligence) capable of transforming a 2d flat image into 3d. While for a human brain it can be easy to imagine an object from different angles, achieving it for a machine is another pair of sleeves. The IA Branch team, DeepMind, is working on a new algorithm capable of meeting this challenge.
DeepMind, the UK-based Google subsidiary specialized in AI, has just presented new work. The AI is now able to generate a 3d rendering based on a flat image like a photo for example or any other image. The human mind is able to imagine what is behind an object, the researchers have managed to endow an AI with the same imagination to some extent. This AI has been trained to achieve this result.
Google unveils an AI that knows how to create 3d renderings based on simple 2d images
How does this AI work? DeepMind researchers have developed a GQN (generative query network). This device allows the AI to be taught how to imagine a scene from several different angles starting from a flat image. At no time does the man intervene to give him instructions to achieve this, which is a real feat. In just 3 images, AI is able to launch a small video in the form of 3d rendering.
What researchers want to develop is an AI capable of "understanding the unsupervised scene." The photos on which AI trains are not images representing the real world, but only relatively basic forms. The next step is to achieve a 3d rendering of a real scene. It is possible one day that this AI can end up in a Google service, why not Photos to create videos from a few images although this is not suggested by DeepMind.
The video available at the bottom of the article illustrates perfectly how this artificial intelligence works. Google develops its services enormously thanks to AI, the scariest example is probably Duplex able to make a call for you, to make an appointment at the hairdresser or elsewhere.