February 9, 2018

51VR crafts photorealistic city streets in Unreal Engine for AI training

By Ken Pimentel

Tapping into the potential of using digital worlds for AI training, predictive analysis, and navigation, 51VR is on a mission to build a virtual domain for a brighter future. The task of meticulously re-creating interactive real-world locations in VR is no simple feat, but using Unreal Engine and photogrammetry technology, the company’s Chengdu Studio is blurring the line between the real world and the digital one with unparalleled fidelity.

With R&D teams located around the world, 51VR is China’s largest and fastest-growing virtual reality focused technology innovators. The company pairs Unreal Engine with VR and other technologies across disciplines, from real estate to education and beyond. More recently, the team’s impressive work re-creating a photorealistic interactive 440 meter stretch of city street in Unreal Engine showcased its dedication to detail in bringing the real world into VR.

Despite looking like a photograph, this is really a virtual Chengdu Street running at 40+ FPS in Unreal Engine

“The so called ‘Chengdu Street’ is merely just the art part of a whole project,” explains 51 Chengdu Studio’s Kevin Guo. “We used photo scanning technology to recreate the majority of the material and small items of this 440-meter-long street. Since the whole project is designed to run trainings for artificial intelligence used for autonomous driving, it has a high requirement for the fidelity of street recreation in order to match the situation in real life and reduce the uncertainty in VR simulation.”

To fully capture and rebuild the level of detail required to suit the specialized project, the team scouted the location, taking over 3,000 3D photos as a reference for modeling and material making. From street materials and prop objects right down to plants and trees, photogrammetry was used to scan and build an in-depth reference base to help make the digital version of the the street as photorealistic as possible.

The process was painstaking but the high fidelity of the final result holds up amazingly well next to the actual reference photography. “In order to recreate the street, we went through a massive modeling and material making process...it took the five of us a month and a half to finish,” says Guo.

The smallest details of Chengdu Street were captured in the Unreal Engine experience created by 51VR

“For a better performance in Unreal Engine 4, we made LOD for the models as well as other optimizations,” he adds. “Due to the final usage of this project, a very high video game FPS target is not what we aimed for. Instead, we focused on how to recreate and present all the information contained in this street on a GTX1080 setup while keeping the frame rates at 40 to 60 frames per second.”

All told, the team used hundreds of 4096 x 4096 textures and shaders to recreate the Chengdu street in Unreal Engine, and the stunning results speak for themselves.

Importing large volumes of design data into Unreal Engine for major projects doesn’t have to be a time-consuming ordeal. Looking for a tool that streamlines and speeds up that process? Join the Unreal Studio beta today to shave hours, if not days off of your real-time design workflows in Unreal Engine!