50 comments

[ 3.3 ms ] story [ 92.0 ms ] thread
It's almost as if you can zoom in on a picture of dog poo and find flies. WOW that's so amazing! ...
Well, if this was not the Ai age, I would probably agree with you. But now we have super resolution and this papers can serve as a warning that now someone can recover the image, enhance it and use as evidence.
???

If I had a picture of an apartment building that was big enough, I'd be able to see through every window. How is that even remotely "technological"? You shouldn't write a paper saying "a mirror can reflect light" when we already use mirrors every day.

Being able to zoom in on an image and see something isn't a breakthrough, holding the details in the picture is.

Another analogy: Being able to find a grain of rice in a picture of a factory-sized floor isn't amazing, taking a picture of the entire floor whilst maintaining detail is.

You're being incredibly disingenuous if you're suggesting this isn't a non-obvious and surprising finding for most people.

Everybody "knows" that scene from Blade Runner is "just science fiction". Well now it's not.

This is definitely not non-obvious. This is incredibly obvious.

I've never met someone who doesn't know that eyes are reflective.

>"Well now it's not."

It never was. Being able to turn corners in a picture was the science fiction part, not the fact that those corners had people in them.

Jesus how am I the only person here who understands this concept. It really isn't rocket science... the solution is solved by technology, not by progressive "methods" of using that technology.

Sure. Eyes are reflective.

But resolution is limited.

If you are so sure it's "obvious" you can extract recognisable faces from reflections on photographs containing eyes - show me you can do it using any photo containing a headshot you can find on Google Images or Flickr... Or even any staged photo you have or can take - showing enough resolution from a headshot that you can extract recognisable faces from eye reflections. I have a 12 megapixel camera in my phone - and I'm not convinced I could reliably take a photo of someone's face with enough resolution to pick to recognisable detail in a reflected face in an eyeball...

Resolution is limited.... to the camera that takes the photo. Yes. You won't get information from something that doesn't have information. Are you going to write a paper that 100 < 1000 ? I wouldn't, but I guess I'm some genius or something.

You can't, because the information isn't there. The same way you can't read a book that is closed, even with a super computer.

Wait a second, if it all relies on the camera, sure you can't read a book with your natural camera - your eyes - alone. But with xray for lead based ink for example, book scanning would be well awesome.
>I have a 12 megapixel camera in my phone

I am sorry, but megapixels do not a good camera make.

My Fuji X100 (35mm prime) and Nikon D200 (50mm prime), the latter from over a decade ago, both easily produce clear pictures in eye reflections when the light is good.

That said -- have you actually tried?

I just took a selfie with my MotoG5s front camera, and I can see a face from my FB feed reflected in my eye. And that's in crappy office lighting, from a screen with brightness turned down to minimize strain!

>Or even any staged photo you have or can take

To re-iterate: it is ridiculously easy to take such a photo with a real camera with decent lens. Modern cell phone cameras are capable as well if the light is good.

Here's a non-attempt: just a crop from the selfie I took

https://imgur.com/ax9Vz5B

> I've never met someone who doesn't know that eyes are reflective.

"Moon landing? Bleh, who cares. We know the moon is a planetoid and how far it is, why is it interesting when we land on it? I've never met anyone that didn't know it was possible to land on the moon."

Please go and read the paper. You seem to imply that the linked paper is something that it is not.

The contents of the paper is pretty much contained in its title. They took photos and saw faces reflected in eyeballs! That's it!

People have been doing that since, well, cameras were invented. And people have been staring into reflections in eyes since forever.

"My face in thine eye, thine in mine appears,

And true plain hearts do in the faces rest;

Where can we find two better hemispheres,

Without sharp north, without declining west?"

-John Donne (1572-1631)

>Everybody "knows" that scene from Blade Runner is "just science fiction". Well now it's not.

It still is, that's not what the paper is about.

The paper really seems to amount to "gee, eyes are reflective".

I'm with him, though I maybe don't agree with the assumed tone about it.

Precisely what surprising thing went from being assumed non-possible to known possible through these folks' efforts?

This paper is pointing out that if you have sufficient fidelity, you can do this -- which isn't the science fiction part of the film plot, it was fundamental and now widely known (and frequently repeated in popular CSI type television and film, to the degree it has now spawned many parodies).

What was science fiction in the film was that somewhere "beyond the year 2000" you'd have a source image of sufficient fidelity to try this with -- and the paper hasn't changed that either, it is quite unlikely you are going to find matching fidelity images (39 megapixel) out of e.g. security camera or child abuse stills anytime soon.

The only novelty here is that the researchers have access to still camera that can take a photo of fidelity required to make this a real thing in their controlled conditions -- which is a little cool I suppose, but is that surprising? Things reflect light, and that kind of capability from imaging sensors is fairly established (yet still not widely available enough to produce real world applications).

They lean on this abstract with these crime solving use cases -- and if those things were real (even if you cheated and used the very top of the line consumer video recorders or security cameras or phone still cameras) then sure yeah that'd be remarkable and maybe people would be surprised by that (precisely because it is not physically possible).

(comment deleted)
>that kind of capability from imaging sensors is fairly established (yet still not widely available enough to produce real world applications).

It has been used in real-world applications for a while -- just not the dystopian kind we love talking about here.

E.g. check out this photogprapher's work:

https://petapixel.com/2015/11/27/a-wedding-photographed-thro...

I agree with you. But what I'm focusing here is that although the paper uses high resolution cameras to retrieve identifiable images, one can may now use super-resolution to retrieve the same images from normal cameras.
One could always see through windows in a building. Do we write a proof for how to do it?
Superresolution can only process information that's even there in the first place -- it might be trained and aware of faces enough to produce a face from very low resolution, but it won't be an accurate identifiable face, it'll be a machine's hallucination of a plausible face, and to even find yourself in that whole courtroom argument of "is this the accused or does the AI just dream that up" you'll need enough resolution in the first place that your entire reflected face isn't just a contributor of a single captured pixel -- enough only to guess at skin hue.
If a system is able to identify a face with better-than-chance accuracy then it shifts things from "We have absolutely no evidence you were there so we have to believe your alibi." to "We have cause to believe you might have been there so we can doubt your alibi. " That's a long way from proof, but it could be enough to get a judge to issue a search warrant or persuade a jury that other evidence is true.

This is not a good thing. Probability-based evidence such as DNA relies on people understanding probability, and most people really don't. If this system is used in the criminal justice system it will very likely lead to innocent people going to prison.

Super resolution doesn't do what you seem to think it does.

It probabilisticly fills in details to make an image look good. It doesn't make those details correct.

>Well, if this was not the Ai age, I would probably agree with you. But now we have super resolution and this papers can serve as a warning that now someone can recover the image, enhance it and use as evidence.

What does Ai have to do with this? AI can't create lost data. If you zoom in on Facebook pics and don't see anything, there's not much that can be done.

On the other hand, if you ever tried taking a close-up of anyone ever, this wouldn't be a surprise.

As far as "using it as evidence..."

The secret is that your "AI" doesn't actually need to create lost data. All you have to do is qualify yourself as an expert witness in Artificial Intelligence, and then you can say that your AI (which, let's be real, is a huge black box whose internal workings are a mystery to anybody) has identified X suspect in an image of y extracted from the scene with a Z probability score.

Fill in X with the appropriate suspect at your discretion.

(This already happens, athough not with imaging yet: https://www.propublica.org/article/where-traditional-dna-tes...)

One of the Japanese camera companies presented this technique at a Siggraph in about 2004 +/- about 3 years. They also showed they could model the shape of the cornea and use that to flatten the image, find the resolution was highest in the direction the person was looking because the curvature of the cornea is highest in the direction the eye is looking.
Hmm, are you thinking of https://www.cs.drexel.edu/~kon/pubs/KNishino_SG04.pdf or something else? The Tsumura paper they referred to definitely doesn’t seem like it...
The timing is about but right but it was definitely different work focused specifically on extracting the scene the subject of the photo was looking at.

I've hunted repeatedlty for a paper or similar covering the presentation and can only conclude that the presentation (and possibly one paragraph conference talk description that probably exists but that I've not been able to find) is likely the only "publication" of the work which is definitey a bummer.

Was it this one?

https://www.jstage.jst.go.jp/article/ipsjtcva/5/0/5_1/_pdf

I remember seeing it that year...was an interesting read.

This one is even closer, but it looks like these researchers used a custom camera setup and the presentation I saw made a big point of the fact that they were doing the work using their existing consumer grade cameras and lenses (since the presenters worked for one of the big digital camera companies of the day)
“Pan right and pull back. Stop…track 45 left. Stop. Enhance 15 to 23. Give me a hard copy right there.”
I have to join jesterjester here to say that this is about "sky is blue" level obvious.

To punch the nail in that coffin:

"Wedding photographed through the eyes of guests"[1] is exactly what the title implies.

That was from 2015.

Someone should let the wedding photographer know he is a scientist now.

[1]https://petapixel.com/2015/11/27/a-wedding-photographed-thro...

This is my warning to those of you attempting to push the boundaries of AI; You will never achieve the utopia you desire because your creations are only 95% perfect. That kind of "perfection" will be enough to enslave us all. You've been duped into betraying every ethical boundary you've ever considered based upon a promise that your intellect will save us all. Unfortunately you've doomed us all to live under the auspices of a system that only needed the power to punish some of us in order to silence all of us.
This paper is not about AI.
Hyperbole much? The same arguments could be made about any technology. Modern medicine would have been considered whichcraft centuries ago. Rock and Roll music was the “devil’s music.” People were murdered when they dare share scientific discoveries that were deemed blasphemous or heretical.

And, to be fair, the “system” has always had the power to punish and silence. The NKVD didn’t need AI to send people to the Lubyanka basement or to the gulags. This anti-AI pontification is ridiculous. ANY technology has the power to enslave just as it also has the simultaneous power to enlighten and uplift. People were brutally murdered long before there were tools to do so and yet, those very tools have protected people just as much as they have harmed. So it goes..

(comment deleted)
The use case here is for identifying criminals based on noisy data with results that are just slightly better than guessing. Couldn’t they start with cat photos rather than something so dystopian?
i read it as cornmeal and was very excited to see what they had done
Reminds me a lot of the Ghost in the Shell episode where a similar technique is used.
(comment deleted)
Now everyone is a mirror and cameras can see around corners
This is a fundable project to deliver a product. The doj and dod both have funding for business for prodictizing this kind of thing.