Artificial intelligence is learning to see in the dark
13 replies, posted
https://www.youtube.com/watch?v=qWKUFK7MWvg
https://qz.com/1279913/artificial-intelligence-is-learning-to-see-in-the-dark/
Phone camera is gonna be really good with low light in the future
theres a pretty big issue with this in that it has to be specialized iirc; neural nets have to make a lot of assumptions that can very well be wrong
I'm guessing that this is a proof of concept that has been trained on a very narrow set of samples, so it has no much room for errors. Also the last photo at least is RAW so there's enough data to brighten it without a lot of guessing.
Sure it can make wrong guesses but for general purposes I think it's helpful still. Human visual processing uses the exact same sort of noise removal which is what causes us to sometimes see faces and figures and whatnot when looking at vague poorly lit shapes and textures
This page contains some of the sample images they presented. This image is the one used in the article. The black clipped areas are still mostly noise, just noise with the right colour and brightness level so it doesn't look out of place. Also note that they are using raw data from the image sensor, which contains more information than an image file so at least there's that to work with. I had a quick look through their paper and I didn't see any mention of testing the algorithm on images outside of the training dataset so it remains to be seen how applicable this really is.
Great, now they’ll be able to hunt us in the dark.
Wait a second, is AI actually going to make the enhance button a reality?
Most this'll be able to do if data is entirely missing is make the photo look correct rather than being correct.
I'm waiting to see what they do with multiple captures or using a photo + short video and compiling that data into a much higher quality photo
https://i.imgur.com/EROu8tn.png
This part looks like it's just really heavy use of a clone tool to fill in missing information.
You'd be surprised with the amount of accurate guesses neural nets can do. They often guess about certain parts of an image not by what you'd think - looking at the object - but by the surroundings. It could definitely predict a bush being there if it's surrounded by greenery or a tree.
They're incredibly good at picking up patterns that we'd never even realize were there, I have this experience all the time when working with DCNs.
Yup, they usually have a hard time escaping the features present in their test data, too.
They've also got convergence issues, not sure if we've fixed that problem yet or not.
"I see you hiding, human. You can't escape."
http://bloody-disgusting.com/wp-content/uploads/2014/09/AI_AndroidsBD.jpg
There has been some amazing work by a Chinese research team and some data scientists in California dealing with convergence consistency, it's pretty great.
GANs have the potential to become reeeeeally interesting and I can see bucketloads of potential commercial use for them
Not really. The thing can add only detail it has seen before elsewhere (which is a general limitation, not something that can be solved eventually), so it could at most falsify evidence.
That's actually one of the things this technology is really good at, see video face swaps and such.
Sorry, you need to Log In to post a reply to this thread.