(Title viaTechSpot)

The full name of DIB-R is "renderer based on differential interpolation calculation". The bottom layer is built by pytorch machine learning framework. At this week's annual conference on neural information processing systems in Vancouver, NVIDIA's research team presented their latest progress.

The working principle of the framework is almost the reverse of GPU's daily work. It needs to analyze 2D images and then form a high fidelity 3D object, including shape, texture, color, and lighting.

The architecture of codec starts from the changeable sphere and uses the information given in 2D image to transform it. It is worth mentioning that the process only takes 1 / 10 of a second.

If a single NVIDIA V100 GPU is used for training, the neural network needs to be trained for 2 days. If other GPUs are used for training, it will take more weeks.

DIB-R can accurately create 3D models when it gives a single image after feeding it with multiple data sets containing bird images.

But co-author Jun Gao said the system can also render any 2D image into a 3D model: "in fact, for the first time in history, you can take almost any 2D image and predict the relevant 3D attributes.".

The researchers believe that the system can be used in the depth sensing application of autonomous robot, so as to enhance its safety and accuracy when working in real environment. Through such three-dimensional processing, the robot can better navigate and manipulate the objects it needs to process.

It is reported that NVIDIA has added DIB-R to its pytorch GitHub library for 3D deep learning to help researchers speed up 3D deep learning experiments.