D3 sensor efficiency

[ejmartin]

Interesting. What methodology did you use to measure the read noise? Did you correct for the clipping of blacks in Nikon's treatment of the raw data?

I have calculated preliminary gain values for the Nikon D40X:

Full well = 55700e
ISO100 IG = 13.6
IS0800 IG = 1.63
ISO 100 read noise = 18e
ISO800 read noise = 9e

I have to recheck my figures and do some plots for ISO200, 400 and
1600, but this seems like a very efficient sensor.

The ISO400 Gain values are probably on the order of 3.3 which
actually exceeds the D300 or 40D but these values verify the DR data
from the 40D Images Resources Imatest results:

Model 1.0(Low) 0.5(Medium) 0.25(Med-High) 0.1(High)
Fujifilm S3 Pro
(Adobe Camera Raw 2) 12.1 11.7 10.7 9.0
Nikon D40x
(Adobe Camera Raw 4.1) 12.0 10.9 10.3 8.9
Nikon D40
(Adobe Camera Raw 4.1) 11.9 10.9 9.89 8.3
Pentax K-100D
(Adobe Camera Raw 3.6) 11.3 10.3 9.51 8.23
Pentax K10D
(Adobe Camera Raw 3.7) 10.6 10.0 9.29 8.19
Canon EOS-1Ds Mark II
(Adobe Camera Raw 3) 11.2 10.3 9.4 8.14
Nikon D40x 10.8 10.0 9.42 8.04
Fujifilm S3 Pro -- 9.9 9.4 7.94
Canon Digital Rebel XTi
(Adobe Camera Raw 3.6) 10.8 9.88 9.18 7.84
Canon EOS-5D
(Adobe Camera Raw 3) 11.0 10.4 9.21 7.83
Canon EOS-40D
(Adobe Camera Raw 4.2) 11.2 10.1 9.26 7.72
Canon EOS-5D
(Camera JPEG) 10.2 9.68 8.82 7.65
Nikon D200
(Adobe Camera Raw 3) 10.6 9.65 8.96 7.61
Nikon D80
(Adobe Camera Raw 3.6) 11.1 10.4 9.42 7.51

( http://www.imaging-resource.com/PRODS/E40D/E40DIMATEST.HTM )
I can confirm that the D40X does, indeed, have 12 stops of DR, in the
Low (engineering) category as measured by IR. This makes the D40X
sensor one of the most efficient sensors on the market, and at low
ISOs it will exceed the performance of the D300 and 40D.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[DuncanM1]

Interesting. What methodology did you use to measure the read noise?
Did you correct for the clipping of blacks in Nikon's treatment of
the raw data?

I calculated the gain values, and then measured the readnoise from a bias frame. I measured gain by subtracting pairs of images from each other and then using the resulting Sigma values, as per Buil.

Duncan

 

[ejmartin]

Interesting. What methodology did you use to measure the read noise?
Did you correct for the clipping of blacks in Nikon's treatment of
the raw data?

I calculated the gain values, and then measured the readnoise from a
bias frame. I measured gain by subtracting pairs of images from each
other and then using the resulting Sigma values, as per Buil.

Duncan

OK, assuming that the D40x clips the blacks like other Nikons, this underestimates the read noise substantially. I have looked at the D300 read noise histogram and it is actually clipped above black by a substantial amount; if one used the above method to measure the D300 read noise one would get essentially zero (for instance in a D300 black frame at ISO 200, over 90% of the pixels have raw value zero). See for instance

http://theory.uchicago.edu/~ejm/pix/20d/posts/tests/D300_40D_tests/

I have some blackframes from a D200 but haven't had time to analyze them; eyeballing them it looks more like they are clipped at zero rather than above zero. Approximating the noise as gaussian (standard bell curve), clipping at the middle of the histogram pushes all the values below zero to zero and effectively narrows the noise fluctuations by hand and giving a falsely low measure of the read noise. If you chop the gaussian exactly in half, the spread is narrowed by a factor 1.66, the read noise is underestimated by about .7 stop and the DR by the same amount.

(edit: if you take a gaussian and replace all the negative values by zero, which is what clipping does, the spread is narrowed by 2.93, or slightly over 1.5 stops)
--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[ejmartin]

(edit: if you take a gaussian and replace all the negative values by
zero, which is what clipping does, the spread is narrowed by 2.93, or
slightly over 1.5 stops)
--

Oops, I should know better than to calculate under editing time constraints. Forgot to take a square root. The clipped gaussian is narrowed by 1.71 (the square root of 2.93), or .77 stops.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[Zane Paxton]

.... A lot of high ISO samples from the D3
are jpegs exhibiting strong noise reduction (especially chroma noise);

That matches what is posted here:
http://www.astrosurf.com/buil/nikon_test/test.htm

By "moot" I presume that translates to acknowledging that the NR is being done internally beyond our control, but it doesn't matter if it produces good results that we find useful.

--
Zane
http://www.pbase.com/devonshire
Nikon D2x
NAPP Member

'Life is not measured by the number of breaths we take, but by the number of moments it takes our breath away.” ~ Anonymous

 

[aclo]

.... A lot of high ISO samples from the D3
are jpegs exhibiting strong noise reduction (especially chroma noise);

That matches what is posted here:
http://www.astrosurf.com/buil/nikon_test/test.htm

No, Buil is talking about some sort of filtering of the raw data for exposures> 1s (this also occurs on the d200, which I have, and I suppose also other nikons). This is not something particularly noticeable from "normal" images, unless you specifically test for it like Buil does, and it's not terribly relevant for normal photography. Martinec's talking about serious NR done on the jpegs (of the sort done by nikon capture, neat image etc).

 

[Dejan Malikovski]

So the Canon lover John was talking absolute gibberish! Nice to know.
--



2007 Digital Camera Satisfaction Study:
http://www.jdpower.com/electronics/ratings/digital_camera/dslr
How many megapixels you need from a camera:
http://www.bhphotovideo.com/FrameWork/charts/resolutionChartPopup.html

 

[Bill Janes]

I have calculated preliminary gain values for the Nikon D40X:

Full well = 55700e
ISO100 IG = 13.6
IS0800 IG = 1.63
ISO 100 read noise = 18e
ISO800 read noise = 9e

I have to recheck my figures and do some plots for ISO200, 400 and
1600, but this seems like a very efficient sensor.

EJ Martin has raised questions about your calculation of the read noise, but I find that your determination of full well is questionable, because your value is greater than Emil found for the D3 whose pixel area is twice that of the D40x.

It is easy to overestimate the full well if over exposure results in clipping of the noise. If you fully saturate the sensor the standard deviation drops towards zero.
--
Bill Janes

 

[John200973]

Those numbers are exactly why you won't have Canon pros turning in their Mark IIIs in droves and switching to the D3. The D3 does give Nikon and Nikon pros a very slightly "better" camera for however long it is until Canon replaces the Mark III. That might be as soon as the Mark IIIn which Canon might do sooner than later just to shake the stigma of the Mark III and it's focus problems. Personally having all of them I still very slightly prefer the 1.3 crop of the Mark III (if mine were useable). Everything of a wildlife and flower nature I have taken so far with my D3 I have cropped at least that much initially (resolution is slightly better with the Mark III, but the colors are better with the D3). Probably why I will continue to use my D2Xs under 800 ISO for those subjects when using Nikon glass. I would very slightly prefer the Mark III also if I could get one that was fixed properly.

 

[DuncanM1]

You forget that the D3 has a base ISO of 200, versus 100 for the D40X. At ISO200, the D40X Well has about 27000, electrons or less than 1/2 of the D3. I measured multiple data points that were well away from saturation. The D40X full well size is proportionately the same as a 40D. I have confidence in my results since they are validated by the IR DR data, that I listed above, but I will attempt to refine them in due course.

Duncan

EJ Martin has raised questions about your calculation of the read
noise, but I find that your determination of full well is
questionable, because your value is greater than Emil found for the
D3 whose pixel area is twice that of the D40x.

It is easy to overestimate the full well if over exposure results in
clipping of the noise. If you fully saturate the sensor the standard
deviation drops towards zero.
--
Bill Janes

 

[DuncanM1]

I used IRIS to give me stat summary of a bias frame. These values must be correct or the DR would be substantially less than that reported by the IR 40D DR summary. I have used the D50 extensively for astro work. Nikon does not clip data, but they simply use a 0 offset. Any signal in a bias frame is preserved.

Duncan

Interesting. What methodology did you use to measure the read noise?
Did you correct for the clipping of blacks in Nikon's treatment of
the raw data?

I calculated the gain values, and then measured the readnoise from a
bias frame. I measured gain by subtracting pairs of images from each
other and then using the resulting Sigma values, as per Buil.

Duncan

OK, assuming that the D40x clips the blacks like other Nikons, this
underestimates the read noise substantially. I have looked at the
D300 read noise histogram and it is actually clipped above black by
a substantial amount; if one used the above method to measure the
D300 read noise one would get essentially zero (for instance in a
D300 black frame at ISO 200, over 90% of the pixels have raw value
zero). See for instance

http://theory.uchicago.edu/~ejm/pix/20d/posts/tests/D300_40D_tests/

I have some blackframes from a D200 but haven't had time to analyze
them; eyeballing them it looks more like they are clipped at zero
rather than above zero. Approximating the noise as gaussian
(standard bell curve), clipping at the middle of the histogram pushes
all the values below zero to zero and effectively narrows the noise
fluctuations by hand and giving a falsely low measure of the read
noise. If you chop the gaussian exactly in half, the spread is
narrowed by a factor 1.66, the read noise is underestimated by about
.7 stop and the DR by the same amount.

(edit: if you take a gaussian and replace all the negative values by
zero, which is what clipping does, the spread is narrowed by 2.93, or
slightly over 1.5 stops)
--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[nickram]

Um, simple question here, looking for simple answer. Does all this mean the D3 can make more beautiful and more realistic pictures than any other slr 35mm digital type camera? More, as in "noticeable", for example, compared to optimal D2X?

Definitions: "more beautiful": richer, sharper, bigger blow ups, greater dynamic range, better image quality, richer color, better saturation, more pop, more film like, better grain, less grain, cleaner at long time exposures, obviously better in all categories at optimal ISO (high ISO is not optimal image so I don't care about that at all), much closer to MF than the D2X etc, etc.

Definition: "more realistic": as in, real life, that is, what it looked like to you in reality when you took the photograph (assuming of course you want to record reality and not distort it). In a sense, also known as "film like"or "human eye like".

Of course, all these questions assume you have an educated, sensitive, experienced eye, and a good dose of visual sophistication. I know this is a subjective question but it's kind of like this: if you know what I am really asking then you will know how to answer. Thanks!

 

[DuncanM1]

Yes, it does mean that. Image quality is determined by two primary factors:

Signal - photons converted to electrons, that you want

and

Noise - electrons that you don't want.

and the ratio between these two sources of electrons determines the image quality, the "signal to noise ratio".

The D3 collects many more photons and turns them into electrons very efficiently, more efficiently, apparently, than any other DSLR, to date. It will produce images with deeper range and colour fidelity than any other DSLR, (and published samples confirm this, it seems) because it collects more photons with which to 'sample' the image and reproduce it accurately. Hopefully the D3 will reverse the trend towards smaller and smaller pixels, and lead a renaissance towards better IQ over "megapixels".

Um, simple question here, looking for simple answer. Does all this
mean the D3 can make more beautiful and more realistic pictures than
any other slr 35mm digital type camera? More, as in "noticeable",
for example, compared to optimal D2X?

 

[ejmartin]

I used IRIS to give me stat summary of a bias frame. These values
must be correct or the DR would be substantially less than that
reported by the IR 40D DR summary. I have used the D50 extensively
for astro work. Nikon does not clip data, but they simply use a 0
offset. Any signal in a bias frame is preserved.

Duncan

Here's what I'm talking about: a blackframe of the 40D looks like this:



an offset is added to Canon raw data before quantization, and so the full gaussian of blackframe noise is evident and the std dev as measured in IRIS is indeed the true blackframe noise. On the other hand, a D200 blackframe looks like this:



The voltage fluctuations due to noise can be either positive or negative. Because there is no offset, negative values are set to zero upon quantization. If you go tthrough the math, this artificial clipping of the noise spectrum narrows the std dev of the noise histogram by about a factor 1.7, so if you simply ask IRIS what the std dev of the blackframe is you get a value too low by this factor, which translates into about 3/4 stop exaggerated DR. You don't get extra DR by artificially setting the black point too high; quite the opposite, in fact.

If you perform this exercise for the D300, the blackframe noise histogram is clipped well above the median; the left 90% or more of pixels are artificially shifted upwards and the std dev of the clipped histogram is about ten times smaller than the actual read noise.

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[Bill Janes]

You forget that the D3 has a base ISO of 200, versus 100 for the
D40X. At ISO200, the D40X Well has about 27000, electrons or less
than 1/2 of the D3. I measured multiple data points that were well
away from saturation. The D40X full well size is proportionately the
same as a 40D. I have confidence in my results since they are
validated by the IR DR data, that I listed above, but I will attempt
to refine them in due course.

The ISO at which the full well capacity is reached is a measure of the sensitivity of the sensor to light, but the full well is the full well, regardless of ISO. Pending verification of your data, I am skeptical of your measurement. If anyone else has data on the D40x, please post it.

Bill Janes

 

[kevm14]

You forget that the D3 has a base ISO of 200, versus 100 for the
D40X. At ISO200, the D40X Well has about 27000, electrons or less
than 1/2 of the D3. I measured multiple data points that were well
away from saturation. The D40X full well size is proportionately the
same as a 40D. I have confidence in my results since they are
validated by the IR DR data, that I listed above, but I will attempt
to refine them in due course.

The ISO at which the full well capacity is reached is a measure of
the sensitivity of the sensor to light, but the full well is the full
well, regardless of ISO. Pending verification of your data, I am
skeptical of your measurement. If anyone else has data on the D40x,
please post it.

I don't think we know full well for the D3 yet. Emil has calculated the capacity that corresponds to the max ADU value at base ISO, but I think full well is higher, reachable at an ISO slightly lower than 200.

 

[DuncanM1]

Here's what I'm talking about: a blackframe of the 40D looks like this:

I understand what you are saying, but the effect of that would be be reduce the D40X DR to somewhat less than the 350xt, yet as we can see from the IR DR data, this isn't the case, in fact this would cripple Nikon DR, in general and have a noticable impact on IQ, yet, again I don't observe this to be the case. If we increase readnoise to 30 electrons at ISO100, the DR drops to absurdly low levels, well below the D200, as per Clark, in fact.

At ISO1600 I get read noise, based upon the bias frame of 9e, the 40D has about 5.6e, and this is a good fit with observed IQ., but 15e versus 5.6e gives the 40D nearly 3 times the DR at ISO1600 and this does not match observed results. At ISO400 I get readnoise of about 9e, again, and this compares well to the 7.4e for the D200, as per Clark and the 7e for the 400D and is a good fit with observed IQ, but at 15e it doesn't make sense, it would give the 400D a huge advantage in IQ, which we simply don't see. My noise measurements are also a good fit to the Imatest noise results, so I have confidence that they are correct.

cheers

Duncan

 

[DuncanM1]

The ISO at which the full well capacity is reached is a measure of
the sensitivity of the sensor to light, but the full well is the full
well, regardless of ISO. Pending verification of your data, I am
skeptical of your measurement. If anyone else has data on the D40x,
please post it.

I have no problem with you being skeptical, but the 20D, with 6.4 micron pixels for example, has a measured full well capacity of 51,000e at ISO100, are you skeptical of that? The 40D is about the same as the 20D, despite its 5.7 micron pixels, while the D40X falls in between at 6.1 microns.

 

[ejmartin]

Here's what I'm talking about: a blackframe of the 40D looks like this:

I understand what you are saying, but the effect of that would be be
reduce the D40X DR to somewhat less than the 350xt, yet as we can see
from the IR DR data, this isn't the case, in fact this would cripple
Nikon DR, in general and have a noticable impact on IQ, yet, again I
don't observe this to be the case. If we increase readnoise to 30
electrons at ISO100, the DR drops to absurdly low levels, well below
the D200, as per Clark, in fact.

Hmmm. I suppose I'll have to make time to analyze the D200 test images I have. I suspect the same issue I am harping on may afflict Clark's measurements as well, though I don't know for sure. He simply states that the analysis he uses is the same as he used for the Canon 1D2, ie subtracting two bias frames and measuring the standard deviation. That will under-report the read noise for the reasons stated in my previous post. Clark makes no mention of any compensation for clipping of the blackframe noise.

I don't think the compensation for this effect drops the D40x DR to "absurdly low levels". It comes down from 12 stops to about 11.3, much more in line with its contemporaries, for instance about the same as the D300 and 40D.

edit: BTW, from your figures above, isn't 55700/18 ~ 11.6 stops of DR, not 12? If so, then applying the correction would drop this to 10.9 stops of DR, a bit less than the D300 and 40D (but still very good).

--
emil
--



http://theory.uchicago.edu/~ejm/pix/20d/

 

[DuncanM1]

There is an offset. It is 1 ADU, which is sufficient to allow measurement of the true SD and noise.

this is from a D40X ISO100 Bias frame:

Mean = 0.97 - Median = 1
Noise (deviation) = 1.25
Maxi = 10 - Mini = 0

x=737 y=314 i=0
RMS noise: 1.185

cheers

Duncan