Scientists at MIT have created a neural network that can accelerate the creation of 3D objects.
Scientists at the Massachusetts Institute of Technology (MIT) have created a method that can accelerate the creation of 3D models by thousands of times. This is reported on the website of the organization.
The experts noted that many areas of life, such as precision surgery or agriculture, require three-dimensional visualization of objects. Typically, neural networks take a two-dimensional image and use it to create a 3D object. MIT scientists said that their new method can accelerate the process by about 15 thousand times compared with existing models.
The authors have created a light field network (LFN), on the basis of which the artificial intelligence learned to reproduce three-dimensional objects after a single observation and frame rate in real time. This method represents a scene as a 360 degree light field and a function describing all light rays in 3D space, passing through each point and in all directions. The light field is encoded into a neural network, which speeds up rendering of the 3D scene.
The experts tested the model on several scenes. They found that by using LFN, the neural network was able to generate three-dimensional objects at a rate of more than 500 frames per second, which is about three orders of magnitude faster than other methods. The scientists also specified that the new light-field network made more efficient use of resources, requiring about 1.6 megabytes of memory.
"Neural rendering has made photorealistic rendering and image editing possible from just a small set of input data," said study participant Associate Professor Gordon Wetzstein. He said the new technology will find applications in computer graphics and other fields.
At the end of November it became known about the application Ricult, which helps farmers to increase yields. The service analyses weather changes to help farmers in Thailand and Pakistan choose the best time to sow crops.
