The neural network self-created world in 3D

New technology memorable fact that for training she did not need additional data. It very works well at any location of the object.

Programmers have created a neural network, independently investigating the scene and calculate how located in it the items will look with strange view without 3D. In the basis of mechanisms for machine vision are vast neural networks, which need a large set of images for training. Their development requires a lot of time in this case, programmers can quickly and efficiently create a good product, so similar algorithms work with disabilities.

The group of experts created a neural network is a representative network that doesn’t know what angles generative component need to “picture the scene”, she sverhdorogie describes the location of objects in the room and their color. During the testing network understands how objects look in the scene as they are to each other, and what properties has the room.

A realistic neural network reproduced the scene, creating quality images, not understanding the rules of perspective, occlusion and light. Also, the algorithm is able to count the number of objects and classify them, even if they are not visible.

Read More