Nvidia researchers present a rendering framework that can produce 3D objects from 2D images

Nvidia researchers have created a rendering framework that uses AI to take 2D information and transform it into a 3D object accurately. The system is called DIB-R, short for differentiable interpolation-based renderer and is built on PyTorch, a machine learning framework. The team will be presenting its findings at the annual Conference on Neural Information Processing Systems in Vancouver, Washington, this week.

Essentially the framework does the opposite of what GPUs usually do. It analyzes an image and then forms a high-fidelity 3D object, including shape, texture, color, and lighting. The encoder-decoder architecture starts with a polygonal sphere and morphs it using the given information in the 2D image.

The process only takes one-tenth of a second, but the neural network takes two days to train using a single Nvidia V100 GPU. Nvidia claims that training using other GPUs would take several weeks. After feeding it several datasets containing pictures of birds, DIB-R was able to accurately create 3D models when given a single image.

While birds were the focus of the experiment, co-author of the research paper on the subject, Jun Gao said that the system can render any 2D image into a 3D model.

“This is essentially the first time ever that you can take just about any 2D image and predict relevant 3D properties,” said Gao.

The researchers believe that the system can be used to give depth perception to autonomous robots improving safety and accuracy when working in an environment. Being able to understand the three-dimensionality of things around it will allow robots to better navigate as well as manipulate the objects they are tasked with handling.

Nvidia has added DIB-R to Kaolin, its 3D deep learning PyTorch GitHub library. Kaolin helps researchers accelerate 3D deep learning experiments.

Permalink to story.

 

ferrellsl

TS Booster
How is this any different from Meshroom? Meshroom has been around quite a bit longer and is certainly more mature that what is being shown here.
 

Evernessince

地獄らしい人間動物園
How is this any different from Meshroom? Meshroom has been around quite a bit longer and is certainly more mature that what is being shown here.
From what I can gather from the framework Meshroom is based off, Meshroom is algorithmically based while this Nvdia solution is AI based.

Meshroom reproduces 3D spaces and objects from 2D images. This Nvidia solution "guesses" based off the input images. Each approach has it's own advantages and disadvantages. The most obvious with the AI approach is that it's going to be less accurate.
 
  • Like
Reactions: onestepforward

Porkous

TS Member
How is this any different from Meshroom? Meshroom has been around quite a bit longer and is certainly more mature that what is being shown here.
I think they are interested in the AI development part, and in part of developer's work around this sphere.
Most interesting about depth perception is that it would help with the robotos.
For example, if you look at the video below, there are many robots in one space, each does its stuff. But with more advancing technology, I assume there would be more freestyle to what robots are adding to the job and to themselves. Maybe in the future, robots would work at nanoscale, as tiny workers. Something like that