• Please use real names.

    Greetings to all who have registered to OPF and those guests taking a look around. Please use real names. Registrations with fictitious names will not be processed. REAL NAMES ONLY will be processed

    Firstname Lastname

    Register

    We are a courteous and supportive community. No need to hide behind an alia. If you have a genuine need for privacy/secrecy then let me know!
  • Welcome to the new site. Here's a thread about the update where you can post your feedback, ask questions or spot those nasty bugs!

For owners of 1DIII with no mega-pixel peeping allergy only ;-)

P

Peter Ruevski

New member
Hi all,

There is something I would like to know about the 1DIII and since I do not have one (for now) I have a "strange" request. I would be grateful if someone who has one does the following test:
1. Take the lens off and put a body cap on.
3. Take one photo at all possible ISO settings using 1/100s shutter speed (all pictures will be pitch black).
4. Upload the the raw files somewhere and post the link (or PM me - I can come up with FTP space or a Gmail account)

Regards,
Peter

P.S. Whole ISO settings (100, 200, 400 etc.) only would be gladly accepted too, but I prefer all of them.
P.P.S. Tell me what the temperature was if you know it.
 
J

John Sheehy

New member
Peter Ruevski said:
Hi all,

There is something I would like to know about the 1DIII and since I do not have one (for now) I have a "strange" request. I would be grateful if someone who has one does the following test:
1. Take the lens off and put a body cap on.
3. Take one photo at all possible ISO settings using 1/100s shutter speed (all pictures will be pitch black).
4. Upload the the raw files somewhere and post the link (or PM me - I can come up with FTP space or a Gmail account)
Click to expand...

There really should be a repository somewhere of stuff like this. All the RAW sample sites have sharply focused images without real black; it would be nice to have blackframes, OOF color checkers or step wedges, clipped files, and resolution tests from various cameras at various ISOs. It gets really tiring listening to all the talk about how much noise this and that camera have, how much DR, etc, all based on certain converters.
 
B

Bart_van_der_Wolf

pro member
Peter Ruevski said:
3. Take one photo at all possible ISO settings using 1/100s shutter speed (all pictures will be pitch black).
Click to expand...

Hi Peter,

Why 1/100s shutter speed?

An unexposed shot at 1/8000s would closely approximate a read-noise only shot, while 1/100s would already accumulate 'some' thermal and amp glow noise, making it a bit harder to compare between different cameras.

I've done such a 1/8000s ISO sequence for my 1Ds Mark II, and posted the results here at OPF some time ago:
ReadNoise_1Ds2.png


Which clearly shows the difference between the 'real' analog gain ISO sequence of 100, 200, ... , 1600, and the calculated digital intermediate (and L/H) ISO's (e.g. ISO 200 generates less noise than ISO 125 and even ISO 160).

Bart
 
P

Peter Ruevski

New member
Bart_van_der_Wolf said:
Hi Peter,

Why 1/100s shutter speed?

An unexposed shot at 1/8000s would closely approximate a read-noise only shot, while 1/100s would already accumulate 'some' thermal and amp glow noise, making it a bit harder to compare between different cameras.
Click to expand...

You are absolutely right about 1/8000 vs. 1/100 - I would ideally like to have both, but wanted to see if someone would bother to take the exposures (not really photos ;-) in the first place.

And I am not interested in the read-out noise alone. Starting with the 1/8000 as an approximation of the read-out noise and then having a series of lets say 1/1000, 1/100, 1/10, 1, 10 seconds would be quite interesting (for you, me and John at least ;-), but that, I'm afraid, is asking too much.

For a while I have been considering writing an "engineers" review of the 30D (something like this one of the 10D vs. D70 or this one of 20D vs. 10D - but without the astronomy bias). And wanted at least some information from the MkIII because of the 14 bit ADCs - as a way of finding interesting differences - and therefore areas to concentrate the "digging" in :)
 
P

Peter Ruevski

New member
John Sheehy said:
There really should be a repository somewhere of stuff like this. All the RAW sample sites have sharply focused images without real black; it would be nice to have blackframes, OOF color checkers or step wedges, clipped files, and resolution tests from various cameras at various ISOs. It gets really tiring listening to all the talk about how much noise this and that camera have, how much DR, etc, all based on certain converters.
Click to expand...
Amen!
There is RAWpository but their goal is different - they concentrate solely on file format - not very useful.
 
B

Bart_van_der_Wolf

pro member
John Sheehy said:
I am curious why you have any color information for read noise.
Click to expand...

It's simply the result of the different ADC analog gain settings per channel. The R/G/B filtered sensels have a different gain setting to compensate for the filter densities (in addition to silicon wavelength sensitivities).

Bart
 
J

John Sheehy

New member
Peter Ruevski said:
And wanted at least some information from the MkIII because of the 14 bit ADCs - as a way of finding interesting differences - and therefore areas to concentrate the "digging" in :)
Click to expand...

Well, I have only seen RAWs from the 1Dmk3 at ISOs 100 and 3200 (they are not blackframes, but IRIS doesn't understand the mk3 yet, so it loads the black masked pixels as part of the image). ISO 100 has about the same pixel read noise as the 1Dmk2, when you scale the 14-bit data to 12-bit. The mk3 has a blackpoint of 1024, and a clipping point of 15280, for a usable range of 14,256 RAW levels, and a blackframe noise of 4.88 ADU. 14256/4.88 = 2921. A mk2 has a blackpoint of 128, a clipping point of 3711, for 3583 usable RAW levels, and read noise is 1.26, and 3583/1.28 = 2843, a very small difference.

log(2921/2843)/log(2) = 0.039 stops

These figures are from individual cameras, and individual cameras of the same model vary more than this. It is hard to know exactly what the real, absolute S/N is; the ratio of clipping to read noise floor is much easier to calculate, and is more relevant to DR.

I have taken the RAW data from mk3 ISO 100 images and quantized it to 12 bit, and even in the most extreme push of the shadows, I could not see any difference between the two. The two extra bits are extremely inefficient, IMO. I have quantized RAW data from numerous cameras, and my conclusion is that unless the quantization results in a read noise of less than about 1.3 ADU in the new bit depth, the quantization has no significant disadvantage, as long as you allow full precision in the conversion; IOW, if you integer-divide all RAW values by 4, multiply them by 4 afterward (zeros in the 2 LSBs), otherwise you quantize the conversion at several steps.
 
J

John Sheehy

New member
Bart_van_der_Wolf said:
It's simply the result of the different ADC analog gain settings per channel. The R/G/B filtered sensels have a different gain setting to compensate for the filter densities (in addition to silicon wavelength sensitivities).
Click to expand...

But that depends on the actual WB used; your data seems to have a slightly different WB for different ISOs.

Wouldn't it be simpler just to measure the greyscale RAW data? I don't know what scale your green values have, if any (other than the 16x going from 12-bit to 16-bit).
 
B

Bart_van_der_Wolf

pro member
John Sheehy said:
But that depends on the actual WB used; your data seems to have a slightly different WB for different ISOs.
Click to expand...

I based my analysis on non-colorbalanced non-demosaiced Raw sensel data. In fact I subtracted 2 read-noise frames (and divided the result by sqrt(2)) to avoid non-random/systematic noise (e.g. from per sensel sensitivity/amplification differences). The results were averaged over the same 200x200=40,000 central sensels used in my other incremental exposures used to determine the analog gain per channel.

Wouldn't it be simpler just to measure the greyscale RAW data? I don't know what scale your green values have, if any (other than the 16x going from 12-bit to 16-bit).
Click to expand...

Indeed, that is what I did, including the 12 -> 16 bit scaling to make it comparable to other cameras' 16-bit file read-outs.

Bart
 
Last edited:
B

Bart_van_der_Wolf

pro member
John Sheehy said:
The two extra bits are extremely inefficient, IMO.
Click to expand...

The real advantage, assuming a 14 or 16-bit ADC, is in the more accurate quantization (in particular of the shadows). My 1Ds Mark II uses appox. 18-19 electrons per Digital Number (ADU in IRIS speak) at ISO 100, that's how the amplifiers are set up. I'd expect the 1D Mark III to be able and differentiate between 15-16 electrons per DN, and thus get more accurate DNs for all exposure levels (Read-noise and Black-level in particular).

Whether that happens in practice, is one of the issues to be empirically determined.

Bart
 
S

Steve Saunders

New member
Okay I've just took the images, all ISO's from 50 to 6400 and all RAW files. Total file size is 169MB so it will take a while to upload to RS. I'll upload maybe in two or three loads and get back to you when they are uploaded later.
 
S

Steve Saunders

New member
Uploading now, I've split the load into two zipped files and they will take about 30 minutes each to upload to RS. I renamed each RAW file ISO50, ISO100 etc so they are easily identifiable and the first zipped file will have ISO 50-500, the second will have ISO640-6400.

I have to say that a cursory glance at them on-screen doesn't really reveal any difference, but then again I don't know what I'm looking for.
 
P

Peter Ruevski

New member
Bart_van_der_Wolf said:
I based my analysis on non-colorbalanced non-demosaiced Raw sensel data.
Click to expand...
Bart,

What software did you use? I am asking because as John pointed out there should be no difference between the R, G1, G2 and B pixel's noise in non-colorbalanced, non-demosaiced pure raw data. I have done this test myself and here is what it looks like:
IMG_8822.png


And the original data is in this file:
http://www.pages.drexel.edu/~par24/rawhistogram/IMG_8822.csv

As you can see in this histogram the spread (noise) of all types of pixels (R G1 G2 B) is pretty much the same at around 4-5 counts (this is on a 0-4095 scale). If your data was scaled to 16bits, that would mean multiplying the values by 16. This would give about 64-80 counts of noise for my 30D and you are getting about 50 (on the red) on your 1Ds MarkII - I can believe that.

However the significant difference between the channels should not be there if the data was not white balanced. This is why I am asking what software you used.
 
B

Bart_van_der_Wolf

pro member
Peter Ruevski said:
What software did you use?
Click to expand...

I used ImagesPlus, in particular the Canon Raw File Conversion in "NoWhite Balance, Mono Bayer Color Data" mode, with output to a 16-bit FITS file (so 12-bits scaled by a factor 16).

I am asking because as John pointed out there should be no difference between the R, G1, G2 and B pixel's noise in non-colorbalanced, non-demosaiced pure raw data.
Click to expand...

Why should there be no difference? The ADC uses different analog gain levels for the R, G, and B sensels, so the ADC quantized noise will also be different for each channel right from the onset all the way up to the end result. It's only after colorbalancing that the black and white points are equal for the channels, not before.

As you can see in this histogram the spread (noise) of all types of pixels (R G1 G2 B) is pretty much the same at around 4-5 counts (this is on a 0-4095 scale). If your data was scaled to 16bits, that would mean multiplying the values by 16. This would give about 64-80 counts of noise for my 30D and you are getting about 50 (on the red) on your 1Ds MarkII - I can believe that.

However the significant difference between the channels should not be there if the data was not white balanced.
Click to expand...

White balancing increases the noise level even more, you are right, that's by the way why it is important to color balance the incoming light with an appropriate filter for best image quality.

Bart
 
J

John Sheehy

New member
Bart_van_der_Wolf said:
Why should there be no difference? The ADC uses different analog gain levels for the R, G, and B sensels, so the ADC quantized noise will also be different for each channel right from the onset all the way up to the end result. It's only after colorbalancing that the black and white points are equal for the channels, not before.
Click to expand...

This is not how things work. The sensor is read out homogenously, as if the camera were B&W. It is up to the conversion software to recogize that the pixels had color filters on them.
 
P

Peter Ruevski

New member
The results

Hi all,

At this link http://www.pages.drexel.edu/~par24/rawhistogram/Mk3Test.html you can find the raw histograms I built from Steve's files.

The results are quite interesting. I have not yet thought through everything but superficially it seems that:
  • The noise performance looks quite good
  • The intermediate ISOs are just as "fake" as in the 5D
  • The data does not appear quite as "raw" as it can be. There is something strange in most of the histograms - probably a scaling of some kind again.
  • For example the value "1019" seems to be taboo ;-) A very drastic dip up to ISO 640 (less than 300 pixels out of 10160800 !?) and a rather visible one all the way to 3200...

By the way all the graphs are scaled to the same so it is very convenient to evaluate noise - high narrow peaks mean low noise, and conversely wide short ones mean high noise.
 
J

John Sheehy

New member
Peter Ruevski said:
The results are quite interesting. I have not yet thought through everything but superficially it seems that:
  • The noise performance looks quite good
Click to expand...

Yes. There is no improvement on the low ISO end over the the mk2, but the high end is markedly improved.

  • The intermediate ISOs are just as "fake" as in the 5D
Click to expand...


  • The fakeness is in a Canon-centric context. Almost all other digital cameras seem to achieve all other ISOs from the base, without a unique optimized initial readout. It's only a problem with Canon because users are used to read noise not scaling directly with ISO.

    [*]The data does not appear quite as "raw" as it can be. There is something strange in most of the histograms - probably a scaling of some kind again.
    [*]For example the value "1019" seems to be taboo ;-) A very drastic dip up to ISO 640 (less than 300 pixels out of 10160800 !?) and a rather visible one all the way to 3200...
    Click to expand...

Those 1019 values are probably only the result of interpolated pixels to replace the mapped-out ones. It is a lot easier for them to interpolate to 1019, relative to the number of instances of real values and other interpolated values when there is a wider noise curve.
 
P

Peter Ruevski

New member
John Sheehy said:
Bart_van_der_Wolf said:
Why should there be no difference? The ADC uses different analog gain levels for the R, G, and B sensels, so the ADC quantized noise will also be different for each channel right from the onset all the way up to the end result. It's only after colorbalancing that the black and white points are equal for the channels, not before.
Click to expand...
This is not how things work. The sensor is read out homogenously, as if the camera were B&W. It is up to the conversion software to recogize that the pixels had color filters on them.
Click to expand...
Exactly. I too am 99.9% sure that the CMOS chip is designed to have exactly the same signal amplifiers on all pixels - it would be too awkward to do otherwise. So ImagesPlus is definitely doing something... ... different... I think I will download it and try it myself...

In fact if there were a market for it, one of these sensors with the last production step of depositing the RGB filter dyes skipped would make a wonderful high resolution and more sensitive black and white camera.

In fact some time ago, someone took one of those $20 one time use CVS digital cameras and physically stripped the Bayer layer... now that is what I call a custom sensor ;-)
http://photo.net/bboard/q-and-a-fetch-msg?msg_id=00CM0R
 
S

Steve Saunders

New member
Looking at your graph Peter (I'm not too good at understanding them) and the way the intermediate ISO's dip the opposite way, does this mean FI that ISO640 is noisier than ISO800?
 
P

Peter Ruevski

New member
Steve Saunders said:
Looking at your graph Peter (I'm not too good at understanding them) and the way the intermediate ISO's dip the opposite way, does this mean FI that ISO640 is noisier than ISO800?
Click to expand...
That is exactly what it means.
In fact there is another very simple criteria for how noisy each setting is. Just look at the sizes of the CR2 files. Smooth uniform data compresses better, so larger files = more noise. Here is the list of your files and their sizes sorted by size:

8,851,978 ISO50.CR2
9,072,747 ISO100.CR2
9,104,771 ISO200.CR2
9,189,957 ISO125.CR2
9,212,504 ISO400.CR2
9,232,967 ISO250.CR2
9,333,133 ISO160.CR2
9,387,834 ISO500.CR2
9,413,326 ISO320.CR2
9,487,567 ISO800.CR2
9,588,031 ISO640.CR2
9,741,850 ISO1000.CR2
10,025,151 ISO1250.CR2
10,055,767 ISO1600.CR2
10,473,186 ISO2000.CR2
10,905,582 ISO2500.CR2
11,408,577 ISO3200.CR2
13,545,598 ISO6400.CR2

Notice the strange inversions. For example 400 is less noisy than 250, 160 and 320... from the graphs it looks like it is even less noisy than 125 too but just barely... I need to do the actual math...

Of course the size trick only applies if the "scene" is exactly the same for all files - and in this case it is. For a real scene "the same" would include focusing and aperture since what is in focus and the depth of field will influence the compressibility of the data.
 
K

KrisCarnmarker

New member
Thanks for the analysis Peter.

I've read about the intermediate ISO's behavior elsewhere, so this is confirmation.

It makes me wonder though, what's the point? Unless there is some specific reason for an exact ISO setting, I see absolutely no reason to use them. Moreover, I'm so used to switching ISO in the "regular" increments, that I find the intermediate ones annoying.

Unless someone can provide a reason to use them, they will get disabled :)
 
N

Nill Toulme

New member
They appear still to be useful in the higher ranges, e.g., 1000 and 1250 cleaner than 1600, and 2000 and 2500 cleaner than 3200. As that neighborhood is where I live a lot of the time, I'm very glad to know that.

Has anybody performed this exercise on the 1DsMkII?

Nill
~~
www.toulme.net
 
B

Bart_van_der_Wolf

pro member
Nill Toulme said:
Has anybody performed this exercise on the 1DsMkII?
Click to expand...

Following the same procedure (1/100th second exposure time, no lens, no noise reduction), here are the dark-noise filesizes of my 1DsMkII:

-L- 10,561,759 bytes
100 10,817,782 bytes
125 11,134,697 bytes
160 11,558,619 bytes
200 10,963,919 bytes
250 11,306,939 bytes
320 11,763,851 bytes
400 11,211,902 bytes
500 11,591,980 bytes
640 12,097,121 bytes
800 11,748,987 bytes
1000 12,194,758 bytes
1250 12,759,277 bytes
1600 13,594,548 bytes
-H- 16,238,014 bytes

As mentioned earlier, the same 'real analog gain' ISO benefits up to ISO 800, and from approximately ISO 1000 we get to the 'unity gain' levels. Once unity gain is reached (the point where a single electron equals a single digital number difference), Poisson statistics will work against us and it becomes better to under-expose and boost the signal in postprocessing. That will avoid adding amplifier noise.

Bart
 
N

Nill Toulme

New member
Thanks Bart. This is extremely helpful particularly as it translates to a very easy rule of thumb for practical application to both cameras: Don't use the intermediate ISO's till you get above 800, then do use them.

Or am I oversimplifying?

Well I see ISO 500 on the 1DsMkII doesn't quite fit the pattern, but it's still probably close enough to let the rule of thumb stand.

Nill
~~
www.toulme.net
 
You must log in or register to reply here.
Top