Nvidia has posted a fun demo showing how GANverse3D, its new deep learning engine, was used to recreate KITT from Knight Rider – the “first AI car” – in 3D.
The technology, due to be added to Nvidia’s Omniverse collaboration platform, can generate a textured 3D model of any type of real-world object on which it has been trained, from a single 2D image.
Uses StyleGAN as a synthetic data generator to train an inverse graphics framework
Nvidia’s method uses StyleGAN, its open-source image synthesis system, to generate multi-view images from real-world reference data: in this case, photos of cars available publicly online.
The synthetic images, which infer how the real cars would look when seen from a series of standard viewpoints, can then be used to train Nvidia’s DIB-R inverse graphics framework.
In its turn, the framework infers a 3D model from the 2D images.
In the case of the KITT demo, the inverse graphics framework was trained with 55,000 car images generated by the GAN: a process that took 120 hours on four V100 data center GPUs.
The trained framework took just 65ms to generate a 3D model from a source image of KITT on a single V100.
Nvidia then created a driving sequence using the model inside Omniverse, including converting the original predicted textures into higher-quality materials.
Not ready for hero renders just yet, but could be used as background content for arch viz
For production work, the 3D models of KITT and the other cars shown in the video leave a lot to be desired.
They’re not particularly detailed, and they’re often distorted, particularly around the roof – a part of a car not often shown in online photos.
However, they were generated from single source images, without any human input, and provide a starting point that could be refined manually.
Nvidia suggests that they could even be used as is as “ambient and background content” in situations where a higher-quality asset, like a stock 3D model, would be “overkill”: for example, to populate large parking lots seen in the background in architectural visualisations.
Available soon as both source code and an Omniverse app
Nvidia aims to release the source code for GANverse3D publicly this summer. The firm also plans to add the technology to Omniverse as part of an ‘AI playground’ app.
Tags: , , , , , , , , , , , , , , , , , , , , ,