Hacker News Reader

Comment by crazygringo | Hacker News Reader

> Photo response non-uniformity, however, is far more difficult to extricate.

I don't see why? Seems like all you need to do is take a photo of an evenly lit gray card, determine the variation of each pixel, and then apply the reverse of that to every photo you take.

And a simple filter can even get around the "evenly lit" part trivially, as long as it's smoothly lit. (I.e. handle gradients.)

The article doesn't mention this at all -- yet isn't this the kind of thing astronomy photographers do when stacking photos, to prevent the "fingerprint" from accumulating?

Of course, it's not even clear this "fingerprint" is even remotely readable from average photos -- it seems like it would be completely indistinguishable in any scene with detailed textures. So it's really only in areas of blank walls, sky, etc. where it poses a risk.

View on HN

It's difficult to extricate for the reasons you mentioned - various textures interfere with extracting it. But it's also surprisingly difficult to purge once you extract that fingerprint.

Astronomers (I presume) do much more in the way of non-uniformity correction (NUC). Often one of the biggest differences between consumer cameras and industrial/scientific ones is the ability to continuously recalibrate them, either in firmware or in post. You could do the same with off-the-shelf cameras and some software, but there's little need most of the time.

> But it's also surprisingly difficult to purge once you extract that fingerprint.

Can you explain why? Is it not just a simple subtraction?

E.g. is it not a pattern of predictably varying intensity per-pixel, but a spatial pattern of nondeterministic noise? So each individual pixel has a likelihood of contributing to the "fingerprint" in any given photo, but you don't know if it contributed in this one?

Or does it totally predictable on the sensor, but the on-camera RAW processing followed by JPEG compression transform it in non-predictable ways, so that you'd need to extricate it on RAW data directly before any processing?

Or is it something extremely non-linear that changes depending on pixel brightness per color channel, or is drastically different depending on exposure time?

I'm super curious if you or anyone has actual knowledge here.

The theory sounds simple, but my (barely) educated guess is that an "evenly lit grey card" is an understimated endeavor.

There is a reason why 3D rendering is difficult; light bounces an awful lot. Objects slightly tint the whole room even if unperceptibly. Lamps vary in intensity, and their shape affect the outcome. So how do you even set the scene up? With a small card the phone being so close can affect lighting. A distant large surface requires more distant lights for uniformity, o simply more lights. Getting the camera out of focus might be useful, but can't correct large scale phenomena properly. Getting the watermark is the difficult part imo.

After all that, once you substract it from every picture you take, every subtle mistake you made in the watermark generation phase will become your new watermark. You bet your privacy on your error being very close to the zero function across the whole grey card. As mentioned in every other comment, astronomers do these calibrations often... for a living. I'm simply not that invested beyond removing EXIF data.

Reply on news.ycombinator.com