4 ms·
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
by murbard2 6y ago
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.
- _visgean 6y agosauce?
- morsch 6y agoLiterally the first Google result for super-resolution from video: https://paperswithcode.com/task/video-super-resolution https://paperswithcode.com/task/video-super-resolution
- murbard2 6y agoLots 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 https://www.di.ens.fr/~mallat/papiers/whitepaper.pdf
- codetrotter 6y agoIMO 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.
- _visgean 6y agoYeah 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.
- murbard2 6y agoMany 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.
- gchamonlive 6y ago"sauce"? Like in tomato sauce? What's going on here?
- Kliment 6y ago"sauce" is common internet slang for "source"
- michaelmior 6y agoOh, I assumed it was referring to what secret sauce was used.
- ceejayoz 6y agoInteresting. 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.
- wtallis 6y agoSimply 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 https://en.wikipedia.org/wiki/Lucky_imaging
- coronadisaster 6y agoIf 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).