80 comments

[ 3.8 ms ] story [ 157 ms ] thread
Hilarious; now that I look at the original photo, I could see myself making the same mistake, perhaps not in such detail!
It's only now that we fully realize how bad the depicted policing was in all those TV shows where they zoomed & "enhanced" the photo evidence.
Well, this does explain the disproportionate amount of actors that are arrested in CSI after reviewing the video evidence.
Think that's bad?

Somewhere there are cops using this thing for real

Indeed, that may very well be the case. I know we have all seen mugshots of persons of interest that are pulled from CCTV and have less pixels than minecraft avatar faces, along with the colour pallet of a ZX81/C64.

How AI and visual processing adapts too face-coverings is sure going to be interesting in many avenues, let alone humans.

Often it doesn't take much for someone who knows a suspect to recognize them from a few blurry frames of video or a drawing.

I am constantly surprised by how extremely bad surveillance cameras are and also how they still sometimes make it obvious who was or wasn't at the scene because of minor (and slmetimes not so minor) details in clothing, shoes, face or gait.

See also https://www.washingtonpost.com/news/true-crime/wp/2018/02/08...

That’s the 5G conspiracy theory “they” don’t want you to think about: 5G will allow pervasive surveillance with quality much better than today’s potatocams
I suspect Ryan Gosling might be the perpetrator in many future crimes!

Almost as scary as the prospect of an AI that can "predict future criminals" based on their facial features!

Not quite. Super-resolution from face prior is probably a bad idea if you're looking to identify someone, but super-resolution from video works. It doesn't make up information based on what it expects the face to be, but collects and aggregates information from multiple frames.
sauce?
Literally the first Google result for super-resolution from video: https://paperswithcode.com/task/video-super-resolution
Lots of neural-network approaches for this which work quite well but are also likely to learn priors from the data which can introduce tricky issues for police work.

There are however approaches which aren't based on machine learning at all, for example https://www.di.ens.fr/~mallat/papiers/whitepaper.pdf

IMO it’s still nice to provide links. For one, readers won’t know ahead of time if it’ll be the first result or not. Aside from that, search results can vary from person to person what with Google filter bubble. And on top of that, if I find a source I still don’t know if it’s the same thing that was being talked about. Worst case I might find a bad source instead and then disregard the information because the source I found for it was bad while there existed good sources for the information unbeknownst to me.
Yeah thats not a source at all. Its a collection of paper on the relevant topic.

The claim of the parent comment was

> but super-resolution from video works. It doesn't make up information based on what it expects the face to be, but collects and aggregates information from multiple frames.

So I was simply interested in hearing where this claim comes from. Has any of these papers mentioned that this is the case? I briefly went through some of them but did not see any such guarantee.

Many of the modern ones are based on deep learning and might indeed capture priors for faces. However, other approaches do not depend on machine-learning. The priors that they implicitly use are weak or rather general (e.g. natural images have mostly sparse wavelet representations, videos represent things that move mostly continuously, etc). They essentially compensate for the motion over several frames and use the redundant information to denoise the result.
"sauce"? Like in tomato sauce? What's going on here?
"sauce" is common internet slang for "source"
Oh, I assumed it was referring to what secret sauce was used.
Interesting. Sounds a lot like stacking in astrophotgraphy, where you take an hours long video of Jupiter or something similar and software identifies the couple hundred frames with the least atmospheric distortion and uses them to bring out detail.
Simply stacking multiple images helps, by averaging out errors (in more or less the same manner as doing a long exposure with perfect tracking). But what you're referring to is a technique for doing even better: https://en.wikipedia.org/wiki/Lucky_imaging
If you take 10 pictures of a fixed object from the same position with the same camera & settings, you also can get a better quality image if you combine them (using 10% opacity of each).
Actually I felt a lot like I was using just that Software when I enhanced a low res picture with this software. Give it a try, it works absolute magic.
It's like robots are now like human, seeing faces everywhere ! https://www.reddit.com/r/facesinthings/
I wonder what AI would make of some Martian landscapes - certainly been the case of humans seeing faces in it before, AI would probably take that up a whole new level.
My thoughts exactly. This is just a case of an ML algorithm describing the face it sees on the Martian surface[0], kind of like a "machine vision Turing test." It's not that it isn't interesting; it's just that it's kind of the expected result when you train a neural net on zillions of images, some of which include faces.

---

[0]: https://www.space.com/11947-photos-mars-illusions-martian-fa...

That just shows how far away from the expectations is the quality assurance for AI-supported projects. Black-box methods like NN will at some point produce weird results. Now we need an AI to check AI if a random face was added to the image in the process.
That does not look like Gosling to me. Why does the photographer think it is him?
I just tried it with his images and did not come close to replicating his results. I am using the current version of Gigapixel AI on Windows.

My first thought when I saw the image was that he photoshopped it. Looks like I am not wrong. Why? Who knows. Bored, I suppose.

I reckon it's like that one paper that shows a neural network swinging from 1.0 confidence that a picture of a lion is a picture of a lion to 1.0 confidence that it's actually a baseball bat once a single pixel had been changed.

So to repro, you would need the original image, not this zoomed in crop nor his final retouch.

Wait. Because you couldn’t replicate the results, the author is a liar. And then you go on to speculate why they lied?
The main problem I see here is that the inserted image is much sharper than the original. I'd guess that the result would be more acceptable and less of a problem if the "blurriness" was preserved. Perhaps they should train on this aspect; it shouldn't be too difficult.
I have a feeling the "detect faces" feature is a hand-engineered feature. It isn't end-to-end ML like the rest of the software.

I guess they run a face detector over the whole image, then cut out faces, pass them through a system to generate a similar but sharp and clear face, and then blend that back into the original image.

If Gigapixel AI works like other state of the art upscaling networks it upscales patches at the time, not the whole image at once. If the only context you have is small enough, scaling it up as Ryan Gosling's face seems completely reasonable. After all I see a face there too. When looking at the larger context(some kind of building) it becomes clear that there isn't supposed to a face, but the AI likely didn't have that much context.
It's like the AI is replicating the same mistake a human brain on LSD makes.
Before casting stones, have they checked with Ryan Gosling to make sure that wasn't him behind a giant magnifier at that window?
Exactly. Perhaps the AI is just smarter than us. After all, there’s really no correct way to upscale — it’s all just making up data that isn’t there based on presumptions.
Pretty sure that's not Ryan Gosling. Just a random face, but adding a celebrity to the headline probably generates more clicks.
Yeah it's just some dude's face. Excellent headline writing to take advantage though.
Turns out the article was written by a GPT-3 bot optimized for clicks. Welcome to the future.
It's like a kind of virtual grey goo EOTW scenario where all the world's resources are consumed writing increasingly effective clickbait.

The Utopian version of the story has the clickbait getting so good that it actually evolves into a high form of art and expression.

CelebA is a relatively standard dataset that is used to train DL-based CV models, it's around 200k of celebrity faces. While the article title is no doubt clickbait, it wouldn't surprise me if Gosling is indeed lurking in those model weights somewhere.
The irony of a human seeing Ryan Gosling in a random face is pretty good.
Yes, just like any camera since cameras have a "detect and focus face" feature like to see faces in things like curtains or other seemingly random structures, just without trying to match a specific known face to it and replacing it. So how's this news? Because of a celebrity name? Surely it's not because it's screaming "ad for Gigapixel AI" in my face. My Sony Ericsson C905 did this over 10 years ago. But looking at the other comments its at best entertaining a little because it's just silly.
Yes, just like any camera since cameras have a "detect and focus face" feature like to see faces in things like curtains or other seemingly random structures, just without trying to match a specific known face to it and replacing it. So how's this news?

It's news because it's trying to match a specific known face to it and replacing it.

It doesn't seem like it's matching a "specific known face". The face only marginally looks like Ryan Gosling; more realistically it just inserted a face that happens to look like a random celebrity and the site ran away with it for clicks.
Enhance! Enhance!

I bet the Pentagon pays good money for such unblur features.

Now you a have a fractal of faces as you go deeper!
Besides the added face, rest of the upscaling done seems to be terrible as well. I wonder why the photographer is using the software
Upscaling is always a problem, no matter how you approach it - weighted neighbours of various sorts, fractals, vector transformations... they all do some things reasonably well and look like poop at least half of the time. When Gigapixel AI works well, it works really well. When it's the wrong program to use, it leaves you no room for doubt about that.
The image is likely zoomed in 800% or more. It's normal that it looks terrible from that close. When you see the images at a normal scale using Gigapixel AI they often look much more detailed (but not always).

Your phone also does this, but even more poorly.

It's not just Ryan Gosling. It's Ryan Gosling with "Charlie Chaplin mustache".
Ryan Gosling is the new Rick Astley. You can now get Ryan Rolled!
I'm sorry, but that does not look like Ryan Gosling.
There are lots of crossing lines, if the AI took those geometrical shapes into account it should have placed lines in front of his eyes and mouth. That would have been more accourate, incidentally avoiding the match with Goslin. Is that possible to implement in Gigapixel? Asking as a CV enthusiast that tailored a volume recognition solution in OpenCV years ago.