Researchers have reportedly developed a Generation Query Network that can create 3D models of scenes from 2D photos.
Scientists from Google DeepMind have developed an artificial intelligence network they dubbed as a Generation Query Network (GQN) that allows computers to make 3D renders of scenes taken from 2D photographs. This enables users to view 2D images from different angles.
The GQN’s algorithm allows it to imagine and recreate photo scenes from any angle without the need for human intervention or supervision. Along with this, the artificial intelligence network does not require any training to create the scenes.
According to the research team, their visual recognition system is so good at creating 3D scenes that it can render the opposite sides of 2D images without even seeing it. It can also do this for multiple vantage points and can even add shadows.
In their paper, published in the journal Science, the Google DeepMind researchers said that their primary goal in developing GQN was to replicate the way a human brain learns about its environment and the interactions between the objects.
GQN has two parts. The first is the representation network where representations of the scenes from the images are developed. The other one is the generation network where the potential 3D images of the scenes from new angles are produced.
While the AI innovation might not appear to be groundbreaking, Matthias Zwicker, a member of the research team, claims that its power lies with what the Generation Query Network could do to come up with new angles not shown in 2D photos. The system does not have radar or depth finder to come up with multiple views of a scene. Instead, it works just by using the few photographs that it captures.
However, the artificial intelligence system still requires additional work to improve its capabilities. The researchers are currently working on determining if they could use GQN on more complex objects.