An artificial intelligence called InfiniCity can build virtual cities that extend in all directions seemingly without end. It could lead to virtual reality worlds that millions of people can interact in or be used for training driverless cars how to cope with new surroundings.
Creating detailed three-dimensional environments can be an intensive process. Making ones that represent the real world requires collecting a huge amount of real-world data, for example by Google鈥檚 Street View cars. Producing them without such reference data can take a big team of 3D designers a long time.
at Snap Research, part of the technology company behind Snapchat, and his colleagues have now developed InfiniCity, an AI model that can generate virtual cities that resemble real ones, but extend infinitely in any direction.
Advertisement
Spaces like these will be a requirement for any future virtual reality space that large segments of the population use, says Lee. 鈥淲e cannot hire millions of 3D designers to create these environments, so in the future, we鈥檒l really need these generative models to help us to generate this infinitely large environment so that we can accommodate millions or billions of users.鈥�
The cities are made in three stages. First, Lin and his team generate 2D satellite maps of virtual cities using a neural network that has been trained on real maps of London. Then, another neural network turns those maps into clouds of 3D pixels, called voxels, that represent shapes like buildings or trees. A final layer of AI input then gives these objects texture and transforms them from block-like shapes to make them look like real, photographed objects.
Generating the realistic images of a city takes about a week, but once it has rendered, you can go anywhere nearby in a second or two, says Lee.聽Or you could jump to a point several light years away in a few seconds, he says, and the coarse voxel structure would be generated instantly, but it would take about a week to render better visuals.
Aside from creating VR environments for us to roam in, the maps could help train self-driving cars in all manner of new situations, says Lee. This is useful because driverless vehicles trained to handle the freeways of California are unlikely to be able to deal with the traffic in Delhi or a narrow country lane in Wales without learning about such environments.
For this to happen, the team would need to increase the level of detail of the maps, says Lee.
It is an interesting combination of several technologies, says at Anglia Ruskin University in Chelmsford, UK. It could also be useful for designers and architects when planning new buildings or larger developments, but for that they would probably need access to the full 3D data rather than just the final rendered map, she says.
Reference:
