• 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!

Impact of 14-bit ADC in the EOS-1D Mk III?

Doug Kerr

Well-known member
Now that the EOS-1D Mk II has been available for a while, can anyone give me a concise review of the apparent impact of having 14 analog-to-digital conversion of the sensel outputs?

By the way, I have never been clear as to whether, in "12-bit ADC cameras", all processing is done to a precision of 12 bits (up to the point that the format calls for an 8-bit representation), or has (for example) the processing been done on a 16-bit precision basis?

Thanks.
 
Now that the EOS-1D Mk II has been available for a while, can anyone give me a concise review of the apparent impact of having 14 analog-to-digital conversion of the sensel outputs?

More accurate gradients, lower noise, better shadow definition.

By the way, I have never been clear as to whether, in "12-bit ADC cameras", all processing is done to a precision of 12 bits (up to the point that the format calls for an 8-bit representation), or has (for example) the processing been done on a 16-bit precision basis?

The ADC outputs 12-bit quantized data. One can only assume the other processing components to have at least the same accuracy.

Bart
 

Doug Kerr

Well-known member
Hi, Bart,

More accurate gradients, lower noise, better shadow definition.

And you experience this in practice?

The ADC outputs 12-bit quantized data. One can only assume the other processing components to have at least the same accuracy.

Of course. My question was, "is the processing ever at better precision than the ADC?".

Thanks.

Best regards,

Doug
 
And you experience this in practice?

I'm strong, I've resisted the Mark III temptation, so far
wink.gif
. I base my judgement on on-line samples from actual users (ones with a more credible background). It is very tricky having to rely on other people's experiences, since the exact shooting/processing conditions will impact the technical quality of the images, but a pattern is visible. That pattern is fully in line with the theoretical expectations.

The real proof of the pudding would follow from a more controlled evaluation of Raw files, I might need to rent a 1DMk3 body to compare with my (same sensel) pitch 1DsMk2. There are other differences tough (e.g. micro lenses and AA-filter) that will also improve overall image quality, so it might be difficult to differenciate between the various contributing IQ factors.

My question was, "is the processing ever at better precision than the ADC?".

Yes, I understood that, but I think only a manufacturer can answer that question for sure, unless one disassembles a 1DMk3 and studies the components in detail (which may be difficult with ASICs). One would expect some components to use 16-bit precision to avoid accumulated errors, but that again cannot be stated for sure by us mere mortals.

Bart
 

John_Nevill

New member
Bart / Doug,

I have a 1DMkIII for the next couple of days, let me know what you need, imagewise, and I may be able to oblige.

BTW, I also have a gretag colorchecker and a few IT8 targets.
 
Bart / Doug,

I have a 1DMkIII for the next couple of days, let me know what you need, imagewise, and I may be able to oblige.

BTW, I also have a gretag colorchecker and a few IT8 targets.

Thanks for the offer John, I hope it won't take too much of your time but this is what would help:

1. Black/read noise frames (2 Raws for each ISO incl. L and H, additional sRaws would be helpful). These are 1/8000s shots without lens (body cap on) and viewfinder closed. Camera noise reduction should be switched off. They would preferably be taken at approx. 4 second intervals (to eliminate potential card writing and LCD noise sources). Two Raws are needed per ISO for subtraction to facilitate systematic (e.g. hot/dead pixel) noise reduction.

2. A sequence of 1/3rd stop increasing exposure pairs (the pairs are needed to eliminate the influence of systematic noise, e.g. dust and uneven lighting), at each ISO but ISO L, 100, 800 for starters would already be useful to reach conclusions (based on those, a higher ISO could be selected for unity gain level performance). The subject should be slightly OOF shots of a uniformly lit surface (to suppress surface structure influence). I use a piece of opaline glass in front of the lens to get super diffuse lighting, but if you don't have that a flat structureless surface will suffice.

I'd suggest starting by determining an average exposure level as indicated by the camera meter or an incident lightmeter. Then, with the viewfinder closed, use Manual f/5.6 exposure +2EV and up to probably +5EV (to make sure all channels fully saturate) by varying the exposure time. The f/5.6 will minimize lens vignetting, and the 1/3rd stop sequence will allow to accurately determine the ADC's analog gain, and the sensel's saturation level (for Dynamic Range evaluation). When your subject is of neutral color, then the LCD clipping indicator may give guidance for how far you need to go with the sequence before total saturation occurs (maybe +4EV is already enough).

3. It would answer a few questions if e.g. the ISO 800 black frame and 1/3rd stop exposure sequence is also executed with the highlight tone priority activated, but at exactly the same manual exposure as without that function.

4. A simple shot of the ColorChecker. This would help the evaluation in IRIS to determine the Bayer CFA pattern order. A shot allowing to create a gamut evaluation requires a much more accurately lit setup than needed for the CFA determination, but I wouldn't object if you could manage it.

I know from experience that it takes some time to do the shooting (and a lot more time to do the number crunching afterwards), so starting with an ISO 800 sequence (with and without highlight tone priority) would already be a significant addition to the earlier black frame noise sequence made available by Steve Saunders and put into charts by Peter Ruevski. However, black frames and exposure frames taken with the same camera, at the same (temperature) conditions, and in image pairs, is preferable for a more scientific analysis method.

We'll have to find a way then to (temporarily) host the Raw files to share them, but getting the shots is the first concern.

Don't hesitate to PM me if you have specific questions regarding the shooting procedure. Maybe John Sheehy and Doug Kerr or others have additional 'requirements', but I didn't want to burden you with too much test shooting so I restricted my request to the essentials. A nice shot at various ISOs of a blue sky will help in getting a better feel for subtle gradient transitions, but it is hard to standardize such a setup for individual comparison between different cameras.

Cheers,
Bart
 
Last edited:

John Sheehy

New member
My question was, "is the processing ever at better precision than the ADC?".

I assume you're talking about the DIGIC JPEG engine (RAW should be final at digitization, except for the slight scaling that Canon tends to do, and bad pixel replacement).

I don't think Canon reveals information like that in any detail.

If people are seeing better results, I don't think the extra two bits are particularly meaningful; they could just as well be zeros for all intents and purposes, but they force more precise calculations because of the assumption that they are meaningful.
 

Doug Kerr

Well-known member
Yes, I understood that, but I think only a manufacturer can answer that question for sure, unless one disassembles a 1DMk3 and studies the components in detail (which may be difficult with ASICs). One would expect some components to use 16-bit precision to avoid accumulated errors, but that again cannot be stated for sure by us mere mortals.

Indeed.

Thanks for your insights, which all make sense to me.

Best regards,

Doug
 

Doug Kerr

Well-known member
Hi, John,
Bart / Doug,

I have a 1DMkIII for the next couple of days, let me know what you need, imagewise, and I may be able to oblige.

Actually, what I am most interested in right now has to do with the "true" ISO speed of the machine (at any given "ISO" setting). Even a rather flawed test of that requires an external exposure meter (whose calibration is "trusted").

A test of the whole metering/exposure chain (again, a flawed test) can be made with the camera alone. (This does not isolate the matter of exposure meter calibration from the matter of the true ISO speed, but generally we find the meter calibration to be nearly conformal with the ISO standards.)

If you could take a test shot to this end it would be very helpful.

Here is the suggested drill:


a. This is best done by sunlight so that secondary implications of white balance are minimized.

b. Use AWB (so as to get a "neutral" exposure result).

c. Make sure that the in-camera processing is the "most benign" (probably means setting contrast to the minimum, but I haven't really looked into the matter of processing parameters in the 1D3).

d. Set the camera t "ISO 100". (Tests at other settings would be useful too, but aren't really required.)

e. Set the metering to CWA and use one of the metered exposure modes (P, Tv, Av, etc.).

f. Shoot a frame-wide image of a neutral test target ("gray card") - the reflectance doesn't matter (so one of the white balance targets would be fine). Take the JPEG output.

g. Determine the "average" RGB triplet value in the center of the image, and let us know what that is.

Thanks so much for the help.

Best regards,

Doug
 

Michael Fontana

pro member
Canon service center stated me the other day, that a 1 DS-3-RAW would be 100 MB, due to the 14bits, too.
I nearly can't belief that....

And no, they didn't had a RAWsample :-(
 

John_Nevill

New member
OK Gents, leave it with me and I will see what I can do!

I'll aim to do most of the light based shots with midday sun, so it'll will be weather dependent

There's no problem with hosting, I have a few Gb of ftp space, so as the images start to unfold i'll deposit them, create a temp guest account and pm those who are interested in the files.

I'll also see if I can sweet talk the MkIII owner in letting me keep it until next weekend.

And finally, I'll have chat with another colleague to see if we can double up on another MkIII.

Phew!, but before all this its my 20th wedding Anniversay today, so I'd better get the immediate priorities right, or I might have a broken MkIII :eek:)
 
Canon service center stated me the other day, that a 1 DS-3-RAW would be 100 MB, due to the 14bits, too.
I nearly can't belief that....

And do, they didn't had a RAWsample :-(

They were wrong. 21.1 'effective' Megasensels * (14/8) bits = 36.9 megabytes image data before compression, plus some for the JPEG thumbnail and header data. But even with compression, that'll challenge your memory cards pretty fast.

That image data will result in 5616x3744 pixel RGB output pixels, so 63.1 Megabytes for a 24-bit RGB file and 126.2 Megabytes for a 48-bit RGB file. Some Raw converters will extract even a few more pixels from the Raw file data, when they ignore 16 pixel multiple dimensions needed for efficient JPEG compression.

Bart
 
OK Gents, leave it with me and I will see what I can do!

Thanks in advance for even considering it, much appreciated.

I'll also see if I can sweet talk the MkIII owner in letting me keep it until next weekend.

If it helps, we could promise to collectively chant positive spells to ward off evil focus issues, as a sign of gratitude. I'm sure he'd like that.

Phew!, but before all this its my 20th wedding Anniversay today, so I'd better get the immediate priorities right, or I might have a broken MkIII :eek:)

Congratulations! 20 years is becoming an exception rather than rule these days. Have a nice celebration.

Bart
 

Andrew Rodney

New member
Now that the EOS-1D Mk II has been available for a while, can anyone give me a concise review of the apparent impact of having 14 analog-to-digital conversion of the sensel outputs.

I could be wrong but I suspect this is more about marketing than anything else. At least with scanners, a higher bit unit didn't necessarily mean a bitter or superior scan. Its useful to capture more than 8-bits of data, no question. Especially when you edit that data. But can we see a difference in 12 bit versus 14 bit when all other spec's are the same? Doubt it.
 

John_Nevill

New member
I've done the 1DMkIII black / read noise frames and they will soon be available (~1hr) via ftp, pm for details.

Files are:

2G0Y9224.CR2 through to 2G0Y9239.CR2 are RAW pairs at ISO L, 100, 200, 400, 800, 1600, 3200 and H.

2G0Y9240.CR2 through to 2G0Y9255.CR2 are sRAW pairs at same ISO range.

BTW, CS3, Bridge and LR won't read the sRAWs, Silkypix does.
 
I could be wrong but I suspect this is more about marketing than anything else.

If only it were marketoid disinformation, it then could be ignored without consequence. Fortunately, it does make a difference.

Imagine an image, full tonal range with shadow noise (Read/dark noise and photon shot noise) and highlight noise (photon shot noise dominated). Let's assume a 12-bit image with a shadow noise standard deviation level of 5, and a highlight noise of approx. 64 (these are realistic levels for Canon DSLRs). Now, what happens when this data is scaled to 16-bits? The noise levels becomes 80 and 1024, or if you were to scale to 8-bits it would become 0 and 4.

A similar 14 bit ADC result would have 5 shadow noise and 128 highlight noise. Scaled to 16-bits that would become 20 and 512, or scaled to 8-bit it would become 0 and 2.

Those noise levels would still be before demosaicing(!), and in linear gamma, but the pattern is clear. The demosaiced, gamma adjusted, and sharpened result from a 14-bit quantization will have lower noise. Also note that these are noise levels of a single standard deviation so with a confidence interval of 68.3%, the real spread can be larger and at higher ISO the noise levels become also more prominent, especially in the shadows. In addition, the tonal differences are quantized more accurately (16384 levels versus 4096), so within the noise error range.

Bart
 
Last edited:

Asher Kelman

OPF Owner/Editor-in-Chief
I could be wrong but I suspect this is more about marketing than anything else. At least with scanners, a higher BIT spec unit didn't necessarily mean a bitter or superior scan. Its useful to capture more than 8-bits of data, no question. Especially when you edit that data. But can we see a difference in 12 bit versus 14 bit when all other spec's are the same? Doubt it.

Hi Andrew,

A good point for scanners! The MFRs quote 16 BIT and give theoretical O.D. that could be achieved of 4.0 or more! In fact few of these scanners do better than 2.7 to 3.2. Perhaps the best go to 3.4.

The point is not merely to have a $35-$700 16 BIT A to D converter! That's easy but delusional!

What's important is the whole chain: scanning motor, optics, physical alignment of the track and mirrors, voltage stability, quality of the voltage going to the A to D silicon and and the processing afterwards. Still, one gets remarkable quality with flatbed scanners for under $100! Unless one is a fastidious and skilled printer, most people will be happy, in fact even delighted with modest prints viewed at an optimal distance. So really there's not a lot of limitation for folk with modest resources.

With cameras it's the same but there's a huge difference between adequate and stunning. As always, you get what you pay for.

What's remarkable is that now camera MFRs are being exposed within weeks of imperfect product whereas before I think they had more time to do silent updates.

Asher
 

Doug Kerr

Well-known member
O.D.?

Hi, Asher,

A good point for scanners! The MFRs quote 16 BIT and give theoretical O.D. that could be achieved of 4.0 or more! In fact few of these scanners do better than 2.7 to 3.2. Perhaps the best go to 3.4.

Remind me what O.D. is.

Thanks.

Best regards,

Doug
 

John Sheehy

New member
If only it were marketoid disinformation, it then could be ignored without consequence. Fortunately, it does make a difference.

Certainly it could make a difference, with sufficiently low noise levels.

Imagine an image, full tonal range with shadow noise (Read/dark noise and photon shot noise) and highlight noise (photon shot noise dominated). Let's assume a 12-bit image with a shadow noise standard deviation level of 5, and a highlight noise of approx. 64 (these are realistic levels for Canon DSLRs). Now, what happens when this data is scaled to 16-bits? The noise levels becomes 80 and 1024, or if you were to scale to 8-bits it would become 0 and 4.

A similar 14 bit ADC result would have 5 shadow noise and 128 highlight noise. Scaled to 16-bits that would become 20 and 512, or scaled to 8-bit it would become 0 and 2.

Those noise levels would still be before demosaicing(!), and in linear gamma, but the pattern is clear. The demosaiced, gamma adjusted, and sharpened result from a 14-bit quantization will have lower noise. Also note that these are noise levels of a single standard deviation so with a confidence interval of 68.3%, the real spread can be larger and at higher ISO the noise levels become also more prominent, especially in the shadows. In addition, the tonal differences are quantized more accurately (16384 levels versus 4096), so within the noise error range.

That sounds like a case of "what one might think", but in reality, it doesn't hold up. The reality is that noise has to be below the ~1 - ~1.2 ADU range before more bits become significantly useful. The 1Dmk3 has a read noise of 4.9 14-bit ADUs. With all the multiple steps involved in white balancing, demosaicing, hue/saturation adjustments, tone curves, sharpening, noise reduction, etc, a conversion can run into posterization much more easily, and having more meaningful bit depth from the beginning is superfluous. Having a few more bits in the rendering engine than meaningful bits in the RAW data is much more valuable, I think, than having extra noisy bits in the RAW data.

What 14-bit and 16-bit RAW files may help to do, is *force* the converters to use more precision. We don't know too much about how converters operate internally, but I have doctored 12-bit RAWs converted to uncompressed DNG, and observed the results in ACR. Even though ACR won't explicitly render highlight areas beyond a certain point with -4 "exposure", it is operating on the full range of RAW data internally. I can tell this because I can alter pixels that were in the 0-4095 range to be in the range near 65,535 (as the DNG is actually 16-bit, utilizing the 12 LSBs), and differences in pixel alteration near the top range of values can be seen as differences in the image, in the way the demosaicing fades gradients away from the clipped areas into non-clipped areas. This suggests that ACR may actually be using no more precision than is in the original RAW. That may also explain why it's anti-banding code works well against gross banding, but doesn't do as much, relatively, for milder cases.

So, my theory is that if the RAW converters always promoted images to the highest possible levels from the beginning, then there would be no illusion of 14-bit files with mostly noise in the 2LSBs giving better results. You would have already had better results with 12-bit files.

And furthermore, there is no major difference in a 1Dmk3 RAW, and one with the 2 LSBs zeroed (or set to binary"10" to maintain the proper mean values). Do you have a case of a file that deteriorates without the RAW data from the 2 LSBs?

The evidence for your hope is lacking, IMO.
 

John Sheehy

New member
Bart what about the issue of posterization. The MK 3 files seem to have more tolerance to adjustment before posterization occurs.

I think it is important to note that no digital camera in production has posterization in the RAW data. There is far too much noise for that to happen. To get clear posterization of RAW data, Noise levels need to be down around 0.5 ADU or even lower. The lowest any DSLR does right now is about 1.25 (12-bit or equivalent), or even 0.9 if my data source can be believed (K10D at ISO 100; I don't have any RAWs from the camera that have totally black areas in them, and the lowest standard deviation I can find in the ones I have is ~2.9 ADU, which may or may not have to do with texture).
 
To test the efficacy of the extra two1DmkIII bits, at my request a friend shot two identical shots using a 1DmkIIN and 1DmkIII "with a polarizer on the lens" to induce a smooth gradient across the sky (i.e. the objective was to get the sky to go from normal intensity to very dark as the polarization angle changed across the sky).

Normally, if you use a CPL, because of the associated gradient, you will see banding, especially as the sky gets very dark. I had "fully" expected the the extra two bits to reduce the banding by a factor of four (by "fully" I mean betting more than a cup of coffee on my level of conviction). Initially the 14-bit A/D and reduced noise at high ISO settings were the key drivers in my decision to purchase the 1DmkIII camera (which didn't actually take place because of all the reported AF issues reported by sports-shooters and birders).

Well, quite a few folks have looked at both of the files and cannot see a difference on a monitor regardless of magnification. I've printed out sections of each file and can't see any difference in banding between the two cameras. Needless to say that my expectations for reduced banding were crushed and left me with my head hanging low. From my perspective, based on the visual evaluations that I (and a few others) have done the extra two bits are more Madison Avenue than substance.

BTW, I can't envision a picture so screwed up that multiple manipulations of Curves, HSL, and whatever else you can think of would actually benefit from the extra two bits; this theory has never been tested by me because I can't find a picture bad enough. However, it is entirely possible that "if" you took a picture under florescent illumination and REALLY REALLY REALLY worked at getting perfect WB across the "entire" tone curve then the extra two bits might be of value; I seriously doubt this because the extra two bits are well below the noise level of the camera (but hey, I learn something new every day).

Regards,

Joe Kurkjian
 

Andrew Rodney

New member
From my perspective, based on the visual evaluations that I (and a few others) have done the extra two bits are more Madison Avenue than substance.

While some could say this isn't totally scientific, it illustrates what happens when the rubber meets the road. It also confirms, unless otherwise dismissed, my feeling that this is more about marketing. We are in agreement.
 
Top