Came across this article, seems like a pretty cool way of creating environments for games...it uses UNREAL Engine 4, so as it is now, wont seem possible within the extremely modified Torque engine that BeamNG uses, but I figured I'd start a discussion thread about it. At least it seems like a pretty awesome bit of technology. CLICK HERE for the link to the article.
It’s not that amazing, IMO. It looks like they just textured some environments and called it a day.....
Something to point out from the article... apparently this takes a massive amount of computing power as they were running it on the new RTX 2080 and they could only just run it at 25FPS. Not fast, but impressive that they can do it at all really.
well I was thinking this type of system could be used to actually just generate the terrain/static objects, and then be used as a map in a game.
Well, I guess the real question is what exactly is the AI doing? Is it generating a 3D map that you can move around? From what it looks like to me is it looks like it is generating images that are smoothly interpolated between screens to make it appear as though you are viewing a 3d environment. Basically it is creating images on the fly and just generating the next one based off of the previous one. Notice the training data that the fed it is only images of them driving around... nothing more. That and they never once mentioned photogrammetry. To me, what this is doing is just trying to hold continuity from image to image while generating its city like setting. It doesn't really seem to "create" an environment so much as just show you one. I am not sure if that made any sense, but to back up what I am saying, notice they mentioned that they had to give it a memory so that the AI would keep an object it created in the image the same color in every frame that object shows up in. They said that the computer kept trying to change the color in every frame before they did that as that's not weird to a computer. While that is true, if it were procedurally generating a 3D environment, then it wouldn't really have that as an issue. Once it created the object, it could just use it without adjusting its color every frame. I am not sure if what I am saying is making any sense... it made sense in my head.
it makes sense to me, I understand what you are saying...the AI is only generating images, not 3D models. But I am sure theres a way that this could be used to create 3D models, give it a few more years.
It dynamically create a coherent environment based on what these networks learned through various videos they were fed. Of course it kind of look like 3D in Google Earth, but this is still an impressive demo. Creating the environment every new frame is sure not that efficient. But using an AI to generate then save a terrain could vastly improve games development time. Of course details would still have to be adjusted by hand but it would still be amazing ! Various CAD programs have plugins where you can generate terrain, mostly urban ones. Pretty good results whitout having AI. Here is an online medieval city generator for example, and here interesting comments on previous link. The generated co-author dance at the end remind me that AI made super-fake of Obama speaking. It could become a real challenge to justice when the only evidences are video based. Also concercing in this age of "fake news", where some head of state want their people to see journalists as public enemies.
