DeepMind Self-training Computer Creates 3D Model From 2D Snapshots (ft.com) 27
DeepMind, Google's artificial intelligence subsidiary in London, has developed a self-training vision computer that generates 'a full 3D model of a scene from just a handful of 2D snapshots," according to its chief executive. From a report: The system, called the Generative Query Network, can then imagine and render the scene from any angle [Editor's note: the link maybe paywalled; alternative source], said Demis Hassabis. GQN is a general-purpose system with a vast range of potential applications, from robotic vision to virtual reality simulation. Details were published on Thursday in the journal Science. "Remarkably, [the DeepMind scientists] developed a system that relies only on inputs from its own image sensors -- and that learns autonomously and without human supervision," said Matthias Zwicker, a computer scientist at the University of Maryland who was not involved in the research. This is the latest in a series of high-profile DeepMind projects, which are demonstrating a previously unanticipated ability by AI systems to learn by themselves, once their human programmers have set the basic parameters.
There's already a program for that... (Score:2, Informative)
Re: (Score:1)
Adobe has been doing it for years as well.
Re: (Score:3)
I'm not sure why this is a big deal. MS had the tech for this about 5 years (Send us a buncha 2d pictures and we'll turn it into a VR set) extrapolating to models isn't that far of a reach.
It sounds like the key difference here is they're predicting parts of the scene they haven't seen, such as what the other side of an object they haven't seen looks like.
I don't know if they do that just based on clues like shadows, or the system says "that looks like the front of a sphere, therefore I can assume it's round on the other side as well."
Re: (Score:2)
It sounds like the key difference here is they're predicting parts of the scene they haven't seen, such as what the other side of an object they haven't seen looks like.
Yeah, the key feature here seems to be extrapolation from a very small number of observations. Say you're in a park and you're snapping a few photos and from that you're trying to reconstruct a 3D model of the park. You don't have nearly enough data to do that properly, but you can make an educated guess. At least that seems to be the focus to me, how well can it fill in the blanks and make some copy-paste assumptions.
Re: (Score:2)
I was surprised, because I'd read that it was difficult. Specifically it was about creating a bump map from a photo of things like those decorative carved panels you sometimes see on buildings.
It's something humans can do quite easily.
Re: (Score:3)
This is something completely different than Photosynth. Completely.
At least just glance at TFA.
Re: (Score:2)
At least just glance at TFA.
I can't. It's paywalled.
Re: There's already a program for that... (Score:2)
Man, are you in Shenzhen? Any much near Nanshan area? Want to invite you for a tea... joke, beer will be better.
Re: (Score:2)
Pics or it didn't happen.
It's for.... eh... the archives.
not unanticipated (Score:5, Insightful)
Re: (Score:2)
The unsupervised or weakly supervised achivements are new.
Re: (Score:2)
The unsupervised or weakly supervised achivements are new.
Not really..... [nott.ac.uk]
Anime? (Score:2)
How well does it work on 2D anime drawings?
Link is DEFINITELY paywalled (Score:2)
The system, called the Generative Query Network, can then imagine and render the scene from any angle [Editor's note: the link maybe paywalled; alternative source],
Why didn't you just use the definitely non-paywalled source?