One of the most interesting things about Grand Theft Auto V is how skillfully the makers have managed to make San Andreas look very similar to Los Angeles and Southern California.
Now, Rockstar Games is working on an innovative machine learning project that might just drive the game towards photorealism.
A group of researchers including Hassan Abu Alhaija, Stephan R. Richter and Vladlen Kolten have developed a bunch of processes. When you use these in the game, it creates a very interesting result. A visual look is produced that bears a strong resemblance to the type of pictures one would take from behind the blurred front window of your vehicle.
When you see it in motion, you will realize how effective it really is. However, the combination of a few things like bleached lighting, polished pavements and cars that are highly reflective make you feel that you are looking right at a street from behind a dashboard. Even though all of it is virtual, it seems real.
According to the Intel researchers, this photorealism is a result of the datasets that the researchers stored in their neural network. The group of researchers has explained this theory elaborately in their paper (PDF). The Cityscapes Dataset, which was built by putting together pictures of different streets in Germany and used in the process, also throws some light on this subject.
The researchers have stated that the kind of improvisations they have made offers much more than what regular photorealistic conversion processes do. Geometric information from GTA V has been integrated into these processes to produce a more effective result. The researchers have named them ‘G-buffers’ and claim that they can accommodate data or information like how good the textures are, the quality of the polish of the cars and the distance between the different objects in GTA V and the camera.
There is a possibility that you won’t see a ‘photorealism update’ being made to GTA V immediately but there is a high chance of you have already seen a video or played a game that has AI upscaling, a different kind of machine learning, being integrated to it. AI upscaling has earlier been used in the Shield TV manufactured by Nvidia and several other projects wherein the focus has been towards upgrading older games.