Hacker News Reader

Comment by virgilp | Hacker News Reader

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.

View on HN

Well, this does explain the disproportionate amount of actors that are arrested in CSI after reviewing the video evidence.

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.

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).

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.

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!

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.

Isn't that more of a case of "a broken clock is right twice a day"?

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

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.

Reply on news.ycombinator.com