Iliah Borg
Forum Pro
Interesting what took Adobe so long.
Take same raw data, pack it to cr2 and arw, and you will have plenty of evidence.
--
http://www.libraw.org/
-
Welcome to the new forums! Please read this first. For known issues we are working to resolve, click here.
EP1 Dynamic range - dxo says 10eV
- Thread starter Eric_1
- Start date
--
http://www.libraw.org/
For "JPEG DR", if one wants to go that way,
Then I'd throw in a handful of differently-coloured sewing threads (not on their spools) in a dark place so the camera's JPEG engine has some fine colour detail to deal with.
Then darken the bright gradient and brighten the darkest parts of the scene to some predefined average levels using some standardised method (e.g. gamma) where visual inspection can be done. Compare to a well-exposed reference. Everyone can then make up their own mind.
It would still be affected by light spectrum and things like that (which the test should point out), but at least the effect of NR, sharpening, JPEG compression and pattern noise would be factored in.
Hopefully, at some point, there will be software capable of, in the vast majority of cases, reducing line noise at the source without damaging the real image noticeably.
The GH1 has the unique problem of clipping 99.4% of its raw data at black at ISO 100, which, I'd expect, would limit the amount of subject details you could salvage from the deepest shadows, as well as fool the wedge/Imatest method.
Yes like a cavity acting as a light trap. DPR's new test scene has a bit of that which is good. Lots of varying low-contrast fine textures to trick the NR algorithm.
Then I'd throw in a handful of differently-coloured sewing threads (not on their spools) in a dark place so the camera's JPEG engine has some fine colour detail to deal with.
A bright, slow gradient to match the clipping against. As small as possible not to reduce overall scene contrast by lens flare needlessly.
I'd use a sequence 1/3 stops apart and pick the one where I get 255,255,255 closest to the middle of the clipping gradient reference (or better also pick a few around that). I'd also put 1/3 stop reference marks on the gradient.
Use the lowest ISO (or the ISO performing the best if there's a compromised lowest ISO).
Then darken the bright gradient and brighten the darkest parts of the scene to some predefined average levels using some standardised method (e.g. gamma) where visual inspection can be done. Compare to a well-exposed reference. Everyone can then make up their own mind.
It would still be affected by light spectrum and things like that (which the test should point out), but at least the effect of NR, sharpening, JPEG compression and pattern noise would be factored in.
Problem with the wedge and Imatest, or DPR, or even DxO, is that none of them registers pattern noise like banding. That can often be the factor that limits the practically useful DR. It also doesn't go away well when downsizing, unlike random noise.
Hopefully, at some point, there will be software capable of, in the vast majority of cases, reducing line noise at the source without damaging the real image noticeably.
The GH1 has the unique problem of clipping 99.4% of its raw data at black at ISO 100, which, I'd expect, would limit the amount of subject details you could salvage from the deepest shadows, as well as fool the wedge/Imatest method.
Iliah Borg
Forum Pro
ejmartin
Veteran Member
Interesting. Do you generate and use a four-channel color profile?
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
Iliah Borg
Forum Pro
Profiles can be applied to demosaiced data, right? Demosaicing "4 to 3" does not need 4-channel profiles. But pre-demosaicing linearization needs to be in 4 channels, independent.
--
http://www.libraw.org/
ejmartin
Veteran Member
Yes. But one also has the option of interpolating an additional bit of chrominance information, so I was asking whether that was used or not.
What do you mean by linearization? More than just normalization of color channels? Again, what is puzzling me is what use could be made of the extra channel data that one doesn't have with the two greens being spectrally identical.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
Iliah Borg
Forum Pro
It is used.
Flare and vignetting correction at bare minimum.
You pointed at metamerism, right? A subtle difference in spectral responses of green channels allows to significantly improve differentiation of colours.
If you look at EXIF of say E-330 you will see that the sensitivity coefficients for both green channels are stated equal. But in fact if you use 4-channel WB the difference can be as large as 0.05 EV, depending on the light spectrum.
--
http://www.libraw.org/
Jay Turberville
Forum Pro
Fool it how? That should simply create a final step where you can no longer distinguish the next step. You may get an odd graph where the last step is less noisy than the next to last step, but I don't think that would affect the DR range at specific noise thresholds report.
--
Jay Turberville
http://www.jayandwanda.com
Jay Turberville
Forum Pro
I was offering that as a possibility that would make it difficult for it to show up. But if there is no good design reason for such an approach, then I agree it is unlikely.
--
Jay Turberville
http://www.jayandwanda.com
It would reduce the noise amplitude (std.dev.) so the noise curves Imatest put out would not be correct towards black. But maybe the wedge test never comes close to real black anyway?
Yes, I suppose if the thresholds are conservative enough the effect will be outside their realm.
One thing I've been thinking about is that when we view the image at a smaller scale, tonal levels that were too noisy at 100% become acceptable. But subject detail that was clipped, how much of that can we retrieve? The noise will help push some of it above the clipping threshold but how much does that help?
Keyboard shortcuts
- f
- Forum