Which is why Nikon guys were so mad at Canon guys (for a long time, actually) who insisted on using their standard raw converter (quite often ACR) when comparing to Nikon, then concluding things based on that.
-
Welcome to the new forums! Please read this first. For known issues we are working to resolve, click here.
DxO Mark and MF x Nikons
- Thread starter rhlpetrus
- Start date
ejmartin
Veteran Member
Roughly correct. The tonal range number is the number of distinguishable gray levels in the range of recorded RAW levels, where the criterion for distinguishability is the std dev of the noise. Mathematically, the integral over raw levels S from zero to Smax (=2^bitdepth-1) of dS/N(S) where N(S) is the noise at a particular signal level S. Now, over the part of the range where the dominant noise is photon shot noise, a gamma transformation of 2 linearizes N(S) and so the post-gamma bit depth needed is close to DxO's tonal range (though it will differ in shadows where N(S) is read-noise dominated).
Here it's better to look at the SNR curves and see the range that exceeds one's minimum acceptable SNR. Universally subtracting off a couple of EV shortchanges cameras like the D3x which has a shot-noise dominated SNR graph over a much wider range than its competitors (including the D3/D700).
Color sensitivity is essentially a 3d version of the tonal range, the number of 3d "cells" of distinguishable RGB tonalities that the camera can record, where the size of the cell is governed by the noise std devs in R,G,B at a given RGB value.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
strawman_John
Leading Member
At the print size you mention, do you need more than say 12mp?? And if the scene is one without a lot of contrast and shadow area in good light, then the compact is good enough.
Its like how many of us actually need more than 12mp? How many print bigger than say 12"x18". I know I do not.
Its like how many of us actually need more than 12mp? How many print bigger than say 12"x18". I know I do not.
DSPographer
Senior Member
Emil,
Thanks for taking the time to write up a much better description of these benchmark values. My reply was just a quick and sloppy approximation of their meaning.
Thanks for taking the time to write up a much better description of these benchmark values. My reply was just a quick and sloppy approximation of their meaning.
Iliah Borg
Forum Pro
Apart from "The dynamic range of 12.84 EV represents the highlight to zero SNR dynamic range. This is useful for comparing cameras but most photographers would use a higher than zero SNR for the minimum acceptable shadow level. I just subtract a couple stops from the given DR andfigure that gives an idea of the EV range the camera can capture" I was not able to find interpretation. Explanation is not the same as interpretation, right?
Now, here is something that puzzles me. On http://www.dxomark.com/index.php/eng/Image-Quality-Database/Nikon/D3X , "Overview", dynamic range is stated as 13.7. On the same page, "Dynamic range" tab, the number for the highest dynamic range is 12.84.
--
http://www.libraw.org/
ejmartin
Veteran Member
You need to click on "print" instead of the default "screen" display on the DR tab to get the value 13.7. "Print" display values are mathematically derived scalings to fixed output dimensions; "Screen" values depend on pixel pitch and represent what is seen in measurements of pixel properties.
The "Print" version is normalized to account for the number of pixels/image height, and therefore gives a guide to SNR, DR, etc at a fixed scale in the image rather than the unnormalized scale of the pixel. The difference is quite apparent if one compares the D3 and the D3x -- "Print" values are quite close whereas "Screen" values are rather different due to the factor of sqrt[2] in pixel dimensions.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
rhlpetrus
Forum Pro
--You need to click on "print" instead of the default "screen" display
on the DR tab to get the value 13.7. "Print" display values are
mathematically derived scalings to fixed output dimensions; "Screen"
values depend on pixel pitch and represent what is seen in
measurements of pixel properties.
The "Print" version is normalized to account for the number of
pixels/image height, and therefore gives a guide to SNR, DR, etc at a
fixed scale in the image rather than the unnormalized scale of the
pixel. The difference is quite apparent if one compares the D3 and
the D3x -- "Print" values are quite close whereas "Screen" values are
rather different due to the factor of sqrt[2] in pixel dimensions.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/
Good shooting and good luck
(after Ed Murrow)
Iliah Borg
Forum Pro
rhlpetrus
Forum Pro
But can you actually compute pattern noise as a separated entity? Doesn't seem so easy.
And I may be missing something here, but the PRINT option just recomputes mathematically the graphs, no extra measure is done, using the downres'ing factor from actual res to 8MP.
Exactly how pattern noise would become relevant for that option only?
--
Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/
Good shooting and good luck
(after Ed Murrow)
Iliah Borg
Forum Pro
There are several ways one can use, to get an idea you can push D3X file shot at ISO 100 10eV in a converter.
And that is pretty meaningless for the resulting actual print.
--
http://www.libraw.org/
antoineb
Veteran Member
Also, why should the DxO mark scores represent THE truth?
The fact that they're the only ones to have come up w sensor scoring, does not mean that their criteria are good, that their methodology is good - no one knows!
And clearly, the fact that their "scoring" does not factor anything for resolution, is clearly a big hole in the methodology.
The fact that they're the only ones to have come up w sensor scoring, does not mean that their criteria are good, that their methodology is good - no one knows!
And clearly, the fact that their "scoring" does not factor anything for resolution, is clearly a big hole in the methodology.
bobn2
Forum Pro
To return to the discussion, the key think about quantitative measurements is to measure things that are subjectively important. If pattern noise is subjectively important, ways can be found to measure it. Returning to the MR's audio analogy, the problem with audio testing wasn't that it was quatitative, but that it was testing things that weren't subjectively important. It turned out that TID was much more subjectively important that THD, and that it also helped to know the distribution of harmonics in the distortion.
The key to measuring 'pattern noise' is in the word 'pattern'. Random noise has a known (spatial) frequency distribution. Noise that has other distributions is pattern noise, so a noise spectrum would give a fair indication of the relative strength of random and pattern noise, and some clues about the nature of the pattern noise. With some research, it might also be possible to quantify the objectionableness of different frequency distributions.
May or may not be the case. A problem with all these tests is that no attempt is made to justify the effects of what's tested subjectively. This isn't too hard to do. With a sufficiently large panel of observers in a blind test, one can rank the subjective effects of various phenomena. If a quantitative physical test correlates well with that ranking, it's probably a useful test.
--
Bob
rhlpetrus
Forum Pro
1) DxO Mark, properly understood, is a measure of Sensor+ADC specifications. If you think it's flawed in any sense, please call our attention to that. It's not a measure of final IQ, since that involves, demosaic'ing, lenses, AA filter system, etc.
2) The final score actually takes into account resolution, since it's computed from the full resolution data to simulate a common resolution for all sensors (8MP, actually). That's what the "PRINT" option does, for the graphs. For example, at full resolution, DR for D3x is 12.8, 13.7 comes from reducing resolution to 8MP.
3) DxO Mark is a decent digital imaging company, they have worked hard on this methodology, I think you would enjoy reading the various technical explanations they have posted about their methodology.
4) They warn against concluding that MF are worse than dslrs because of final scores, since that takes into account low-light ability of sensor, which is not a concern with MF, actually some of them (the Hassy, for example, ISO50) only have one ISO setting, other values are obtained by digital amplification. That distorts the final scoring, and they say that explicitly.
--
Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/
Good shooting and good luck
(after Ed Murrow)
ejmartin
Veteran Member
Pattern noise is due to the camera electronics, in particular the ISO amplification and ADC. It can be measured from black frames in Canons, which don't clip blacks, and from suitable near-black exposures on cameras such as Nikons that do clip blacks. Take the black frame and do a Fourier transform of the image. Here is the result for a Canon 1D3:
and here is the Canon 40D:
The lines are the banding component of the pattern noise -- vertical lines represent horizontal banding, horizontal lines represent vertical banding. There are four vertical lines on the 40D I suspect due to its four-channel readout, and eight for the 1D3 due to its 8-channel readout (so the D3 should have 12 vertical lines). The 40D also has some isolated off-axis bright spots, which represent noise that is periodic in both horizontal and vertical directions. Though I haven't done the analysis, one could make a quantitative measure of banding noise by subtracting off the diffuse background (which is the spatially random component of read noise) and/or averaging along the bright lines to get a quantitative measure of the amplitude of the pattern.
I believe Iliah's point is that pattern noise doesn't scale the same way that random noise does; in fact it reduces very little under moderate amounts of downsampling. Pattern noise is coherent across spatial scales.
Because it doesn't reduce as much as random noise under downsampling.
However, just because DxO have not included it in their analysis does not invalidate their graphs for "Print" resolution. One just has to remember that what is being displayed for SNR etc relates to the random component of noise relative to signal.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
rhlpetrus
Forum Pro
Thanks, makes sense.
They try to show with images the effect of measurement scale (click on right bar of graphs). Not that useful, except in extreme situations.
Re a panel of expert: I still think these measurements are what they are, sensor specs, essentially, with some bearing on final IQ, more or less relevant in each case. For example, detail outweighs noise, for subjective appraisal, of final images, but it's not something that can actually be measured, beyond the traditional resolution chart shots. DR is actually much hard to evaluate subjectively, since some take contrast for detail, etc.
One point: I think that the way to actually evaluate IQ would be as follow: print both files at full resolution. View them at different distances, compensating so that angle of view is same, thus resolution difference is cancelled.
Why that: because any type of up or down res'ing has IQ effects. I realized that when I saw one of images printed at full resolution with a good lightjet printer. It's much better than anything I can get from my monitor, either at 100%, downres'ed, sharpened, etc. In fact, it's much better than anything I've done in film, I'm still amazed, regarding detail, tonal range, noise (grain), etc.
--
Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/
Good shooting and good luck
(after Ed Murrow)
rhlpetrus
Forum Pro
Very good examples, thanks, learning a lot from this thread.
--
Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/
Good shooting and good luck
(after Ed Murrow)
--
Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/
Good shooting and good luck
(after Ed Murrow)
sidelight
Senior Member
I've also read the replies to Iliah's post on the pattern noise topic (Edit: just now saw Emil's). I think this is something that is important to IQ and it can be measured as shown in the Imatest (Norman Koren's program) plots generated by Imaging Resource. Perhaps you have seen these. They emphasize looking at the Noise Frequency plots (lower right in the grouped area) as being most diagnostic. In these, they plot Noise Power vs Frequency and you can easily see the distribution for the color channels and luminance. Also, the area under the curve (AUC) gives some idea of the magnitude of the noise over the frequency range. They show plots at different ISOs.
I've provided links below for the D3 and D90 as clean cameras (but look at the differences) and for my D200, which is acknowledged as a noisier camera. It would be easier to compare if these plots were all side by side but I didn't see how to do this. Still, I think the plots show differences and these get to Iliah and Bob's points as at least one way to correlate measured values with subjective perceived IQ.
http://www.imaging-resource.com/PRODS/D3/D3IMATEST.HTM
http://www.imaging-resource.com/PRODS/D200/D200IMATEST.HTM
http://www.imaging-resource.com/PRODS/D90/D90IMATEST.HTM
--
David
Iliah Borg
Forum Pro
Aside from repeating pattern noise, there is also sensor pattern noise (for the lack of a better short term). You can easily see that sensor pattern pushing a shot with a body cap on and viewfinder closed, base ISO, short exposure, camera covered from light, if you push the resulting raw about 10 stops. Worse, when the camera gets heated from a battery or from sun/ambient the patterns develop further.
Statistical analysis done over relatively large portions of a sensor can show separation where it does not exist for an eye.
--
http://www.libraw.org/
Statistical analysis done over relatively large portions of a sensor can show separation where it does not exist for an eye.
--
http://www.libraw.org/
ejmartin
Veteran Member
By "sensor pattern noise" do you mean banding noise, or something else? Banding shows up in the noise power spectrum (the Fourier transform of the black frame) as in the examples I posted -- it's the bright lines. The repetition of the bright lines is an artifact of the multichannel readout; that produces a rather less apparent regular oscillation of the brightness of the line in the line noise.
--
emil
--
http://theory.uchicago.edu/~ejm/pix/20d/
with an old P45+ ( 3 years old back )
http://forums.dpreview.com/forums/read.asp?forum=1021&message=30419028
a recent MF back would give better picture.
Dxo use some 4 years old back.
We should compare them to a 4 years old Nikon ...
http://forums.dpreview.com/forums/read.asp?forum=1021&message=30419028
a recent MF back would give better picture.
Dxo use some 4 years old back.
We should compare them to a 4 years old Nikon ...
Keyboard shortcuts
- f
- Forum