Looking at the photo, anyone will easily be able to imagine what is depicted on it as a three-dimensional image. And even if some of the objects behind the scenes (for example, we see only the front part of the back of the chair, but don’t see the rear), we are quite easy to guess how it will look in the subject on the reverse side. AI is such a complex tricks, but recently, a group of project engineers DeepMind not only taught artificial intelligence to do the above, but also to translate the 2D image into full 3D.
A new system called “Generating network requests” (Generative Query Network — GQN). To teach artificial intelligence to analyze two-dimensional scenes, a team of researchers showed him pictures of the same scene from different viewpoints. This chain of images AI used in order to understand the principle of object changes when you change the angle. Moreover, this approach allowed to learn to predict how will the stage look like from a different angle, taking into account not only the position of objects, but also lighting. As reported by one of the authors Ali Eslami,
“Imagine that you are looking at mount Everest and move to the meter. The mountain will not change its size, which will give you information about the distance to and size. But if you look at the mug and doing the same thing – it will change. Approx and our technology.”
“Moreover, the algorithm learns in the same way as people. Seeing the same object many times, the AI analyzes its characteristics, remembers and then uses repeated interaction. Now, artificial intelligence can recreate a maze, scanning just a few photos taken inside.”
Google DeepMind neural network learned to turn 2D images into three-dimensional objects
Vladimir Kuznetsov