CMOS vs CCD and colour profiles a quick and dirty comparison

[ErikKaffehr]

Hi,

Here is a quick and dirty comparison between:

Leaf Credo 50 (Sony CMOS)
Phase One IQ 360 ( DALSA CCD)
Sony A7rII (Sony CMOS)

This comparison was using LR as it allows the use of DNG Colour Profiles

The protocol:

- An 'EIP' file was chosen from each aperture sweep.
- The 'EIP' was unzipped and the raw was saved in a new folder as DNG
- A colour profile was built for each using Adobe DNG Profile Editor on each
- The images were converted the newly created profiles
- WB was on the third brightest white patch
- Fourth brightest patch was adjusted to have L value around 51 in Lab.
- All three was opened in Photoshop

The IQ360 was downsized to match the Credo image.

Results, see below.

Why I did not use Capture One? The main reason is that I wanted to use an easy to use profile generator. I might retry it with profiles built by DCamProf.

Best regards
Erik

[Christoph B.]

So what you're telling us is that if your file down the threads of a screw it becomes a nail...

[ErikKaffehr]

Hi,

I don't know what the story tells. If you feel that the colour reproduction as shown is very similar it would indicate that colour profiles are more important than sensor type or CFA design.

If you feel that colour reproduction is significantly different between IQ360 and Credo 50 it would indicate CMOS/CCD/CFA differences cannot be handled colour profiling. At least not by Adobe DNG Profile Editor used as a profiling tool.

So the message is really what you see...

Best regards
Erik

So what you're telling us is that if your file down the threads of a screw it becomes a nail...

[Bart_van_der_Wolf]

Hi,

I don't know what the story tells. If you feel that the colour reproduction as shown is very similar it would indicate that colour profiles are more important than sensor type or CFA design.

Hi Erik,

The most obvious visual difference is in the Leaf's orange/yellows. But the DNG profiler uses the native profile for a camera and adjusts on that basis. So if the Leaf already had a different look, and many say it does, very pleasing for skin tone, then the modified profile will carry some of that over to the modified profile.

Since the 2 CMOS devices have a different look, and a CCD and a CMOS device have relatively similar looks, I'd say the profile makes the day, and the effect of the silicon based sensor design is of secondary importance.

A more labor intensive task would be to build a profile for each of the devices from scratch, for all three, all having the same 'look' applied. The Argyll / DCamProf route would enable to do that, but it takes time to do it right.

If you feel that colour reproduction is significantly different between IQ360 and Credo 50 it would indicate CMOS/CCD/CFA differences cannot be handled colour profiling. At least not by Adobe DNG Profile Editor used as a profiling tool.

One might stumble over specific issues for a specific camera, due to CFA differences that cannot be fully 'profiled-away'. The specific illuminant may play a role, while another illuminant may give no problems. So, no easy task (as most controlled testing is never easy). But it would be a typical task for a studio photographer who works with multiple platforms and controlled lighting, and wants to have a similar 'look' from all capture devices.

Cheers,
Bart

[ErikKaffehr]

Hi Bart,

That is on my "to do" list...

Thanks for chiming in!

Best regards
Erik

A more labor intensive task would be to build a profile for each of the devices from scratch, for all three, all having the same 'look' applied. The Argyll / DCamProf route would enable to do that, but it takes time to do it right.

[ErikKaffehr]

Hi,

I have used DcamProf to generate DCP profiles based on DT's test shots for the IQ360, Credo 50 and Sony A7rII. This should shed some light on the following questions:

• Would CMOS or CCD give similar colour rendition with the adequate profiles?
• Do MFD and A7 sensors have different CFA-s?

It should be noted that there was "hue shift discontinuity" problem generating the IQ360 profile.

LR CC was used for these conversion but I may look into using C1 in a few days.

Here the IQ360 represents a DALSA CCD sensor while Credo 50 is a representant for the Sony CMOS based MFDB-s.

Please note that these profiles are linear, good for reproduction but a bit flat. I will add curves later.

So, what I do see?

To me it seems that the IQ-360 produces some more saturated colour than the others, at least on the orange and the blue shaft on the brush. The Sony A7rII and the Credo seem to be very close.

The value of these comparisons is to see what happens if we use "same math" on the raw images.

Processing was following:

1) White balance on third brightest grey patch
2) Adjust fourth grey patch to about L = 52 (in Lab mode)

Best regards
Erik

[Jack Hogan]

To me it seems that the IQ-360 produces some more saturated colour than the others, at least on the orange and the blue shaft on the brush. The Sony A7rII and the Credo seem to be very close.

The value of these comparisons is to see what happens if we use "same math" on the raw images.

Hi Erik, good work on these profiles.  I am thinking that LR is definitely not the appropriate tool if you want 'same math' on the raw images, because we know that it virtually never, ever does the same thing even with all sliders zeroed out.  I would suggest RawTherapee, which does take dcp profiles and can be easily setup for 'same math'.

As to color saturation, imho it normally means nothing (although I don't know whether the comment applies to a dcamProf context).  For instance I was astounded to find out that when a dcp profile is used the linear matrix to XYZ has no negative terms, and often quite a few zeros in it - presumably not to incur in clipping.  This results in totally desaturated images.  They make it up later with non-linear profiles in HSV space, yikes!  I hope your raw converter runs floating point internally

Jack

[ErikKaffehr]

Hi Jack,

Thanks for your input!

The suggestion to use "RawTherapee" is very good suggestion. I will make some work with RT and share. The reason I used LR here is that it is the raw converter that I am most familiar with.

AFAIK, very few raw converters use modern concepts, like floating point. I don't know for how many hundred years FP has been around, but not using FP says a lot about the ancient state of image processing…

Best regards
Erik

Hi Erik, good work on these profiles.  I am thinking that LR is definitely not the appropriate tool if you want 'same math' on the raw images, because we know that it virtually never, ever does the same thing even with all sliders zeroed out.  I would suggest RawTherapee, which does take dcp profiles and can be easily setup for 'same math'.

As to color saturation, imho it normally means nothing (although I don't know whether the comment applies to a dcamProf context).  For instance I was astounded to find out that when a dcp profile is used the linear matrix to XYZ has no negative terms, and often quite a few zeros in it - presumably not to incur in clipping.  This results in totally desaturated images.  They make it up later with non-linear profiles in HSV space, yikes!  I hope your raw converter runs floating point internally

Jack

[Bart_van_der_Wolf]

I have used DcamProf to generate DCP profiles based on DT's test shots for the IQ360, Credo 50 and Sony A7rII.

Hi Erik,

The color rendering is getting closer to the same look, with mainly a slightly lower contrast result for the IQ360. If that would be adjusted, the differences would be even closer. Which shows how well the performance of different technologies and CFAs can be profiled away.

As Jack said, there may still be under-the-hood adjustments going on in different Raw converter engines, but that's something that's harder to find out. As long as one sticks to a single converter, an approximately unified look between cameras can be achieved. Another benefit of DCamProf is that it additionally allows more finetuning of the profile than several other commercial products.

Cheers,
Bart

[ErikKaffehr]

Hi Jack,

Here are the files converted in RawTherapee. These conversions use a new set of profiles with "acr-curve".

The third image is the Credo converted with the Sony A7rII profile, compared with the Credo 50 profile. Not a lot of difference…

I am not very familiar with RT, so mistakes are possible.

Best regards
Erik

Hi Jack,

Thanks for your input!

The suggestion to use "RawTherapee" is very good suggestion. I will make some work with RT and share. The reason I used LR here is that it is the raw converter that I am most familiar with.

AFAIK, very few raw converters use modern concepts, like floating point. I don't know for how many hundred years FP has been around, but not using FP says a lot about the ancient state of image processing…

Best regards
Erik

[eronald]

Hi Erik,

 As long as one sticks to a single converter, an approximately unified look between cameras can be achieved

Cheers,
Bart

Yeah, sure, until you find a metameric failure. Which, if you look for it, you will.

it's like aliasing - unescapable.
Edmund

[ErikKaffehr]

Hi Edmund,

I am sure you are right, there will always be cases of metameric errors.

Just to say, aliasing is not unescapable. We can stop down until diffraction kills resolution.

Best regards
Erik

Yeah, sure, until you find a metameric failure. Which, if you look for it, you will.

it's like aliasing - unescapable.
Edmund

[Bart_van_der_Wolf]

Yeah, sure, until you find a metameric failure. Which, if you look for it, you will.

Sure Edmund, but that has very very little to do with CCD vs CMOS, and only applies to Bayer CFA filters (which create a different trichromatic sensitivity from that of our eyes). Metamerism will always be an issue as long as our cameras and displays are RGB devices, and we use a single profile for different illuminants.

And then it still remains to be seen how big the differences are between different Bayer CFA dye absorption spectra. Many of them have roughly the same characteristics. Also, creating a profile for a specific controlled illuminant (like in studio use) attempts to tackle most of the differences, in fact that is what profiling does, make metameres look the same under a given illumination.

DCamProf also allows to unify things like the way clipping is handled, how contrast is applied, how saturated the colors are rendered and their roll-off towards the limits, etc. That also creates a specific 'look', and can thus be unified.

Cheers,
Bart

[torger]

I'd say that metameric failures are not that common with modern cameras. What you get is that in certain ranges the color filters are worse at separating colors (meaning the profile needs to "stretch" more in that area leading to more noise), but as filters are overlapping they can separate at least some. Sure, with some narrow band artifical color you can trigger metamerism, but it's a narrow case.

Commercial profiles generally prioritize robustness over a specific look design, and thus the profile's stretching is quite weak and thus closer to a linear matrix profile, and thus more of the sensor's color filter character is kept. If you wish you can force cameras to look very similar, but it will lead to quite aggressive stretches and a less robust profile. Still, even if you relax the stretching a fair amount modern camera color filters are today so similar that the looks become very similar anyway.

Between Kodak and Dalsa CCDs there was a big difference between color filters. Between various modern CMOSes today, not so much, unless you look at extremely saturated colors and narrow band colors. The Dalsa CCDs have color filters that are "modern" and neutral, while Kodak seems to have tried to mimic a subjective film response.

From my profile work I've noted that "pleasing skin tones" is not so much about hue or color separation, but how contrast is applied (the eye/brain's color perception is modulated with contrast) and how you rolloff to the whitepoint (highlight rendering). This is about profile and raw converter design more than anything else. When designing those aspects you have very little color science to rely on so it's very much an art.

I don't totally dismiss that CCD could have some different look aspect than CMOS, but suspect if so that it's a pixel peep effect, a texture difference due to sharper aliasing, a different noise quality, and less if anything about color. Sure if everyone loved the Kodaks which indeed have much different color filters, but Dalsa with the more neutral/modern and similar to modern CMOS filters receives much more love. In any case in my work I haven't found the CCD vs CMOS difference to be worth any specific attention.

[Jack Hogan]

Good show, Eric, you are right, not too different at first glance.

Notice the forward matrices have no negative terms?  That means that they 'desaturate' the raw color information when transferring it linearly into XYZ space.  The non-linear profile takes over from there, massaging the colors until they look as they should (it's an offer they can't refuse). I am not sure I understand the logic behind it 100% (perhaps Edmund or Torger can explain it), but it seems to me that it goes something like this: 'let me squeeze all the colors I can into XYZ, even if colorimetrically incorrectly; I will then try to make them look right by bending, pushing and pulling them to where I know they should be (move over and stay there, damn it)'.

It seems to me that after such an energetic tour de force, trying to determine the difference between 'CMOS vs CCD' CFA color recipes by looking at the final result is a bit like trying to figure out the type of potato used in Pringles by tasting them

Jack

[torger]

Good show, Eric, you are right, not too different at first glance.

Notice the forward matrices have no negative terms?  That means that they 'desaturate' the raw color information when transferring it linearly into XYZ space.  The non-linear profile takes over from there, massaging the colors until they look as they should (it's an offer they can't refuse). I am not sure I understand the logic behind it 100% (perhaps Edmund or Torger can explain it), but it seems to me that it goes something like this: 'let me squeeze all the colors I can into XYZ, even if colorimetrically incorrectly; I will then try to make them look right by bending, pushing and pulling them to where I know they should be (move over and stay there, damn it)'.

Yes, that's correct. The reason negative terms are avoided is to avoid clipping of extreme range colors. Matrix-only profile always have negative terms, but then also robustness problems in extreme range, for example that deep saturated blues are clipped to black. Due to this I'm personally not very fond of matrix-only profiles.

Actually it's not only about avoiding non-negative factors, in DNG pipeline you must fit all into ProPhotoRGB if you want to avoid clipping. This desaturation and then stretching afterwards with the LUT is not as brittle as it may sound though, as the stretching is pretty linear. It's a purely DNG pipeline technical issue, and in my opinion it takes away some of the elegance of DNG profiles (I'm quite sure Adobe didn't realize this clipping problem until long after the profile format was decided, and indeed the early Adobe profiles have clipping issues), but in terms of actual result it's not a problem. It just makes it messier for us making camera profiling software.

[Jack Hogan]

Yes, that's correct. The reason negative terms are avoided is to avoid clipping of extreme range colors. Matrix-only profile always have negative terms, but then also robustness problems in extreme range, for example that deep saturated blues are clipped to black. Due to this I'm personally not very fond of matrix-only profiles.

Actually it's not only about avoiding non-negative factors, in DNG pipeline you must fit all into ProPhotoRGB if you want to avoid clipping. This desaturation and then stretching afterwards with the LUT is not as brittle as it may sound though, as the stretching is pretty linear. It's a purely DNG pipeline technical issue, and in my opinion it takes away some of the elegance of DNG profiles (I'm quite sure Adobe didn't realize this clipping problem until long after the profile format was decided, and indeed the early Adobe profiles have clipping issues), but in terms of actual result it's not a problem. It just makes it messier for us making camera profiling software.

Thanks for the clarification, Torger.  Would you still be against non-desaturated forward matrices to XYZ (with optionally a tweaking Profile if desired) in a fully floating point environment?

Jack

[eronald]

Thanks for the clarification, Torger.  Would you still be against non-desaturated forward matrices to XYZ (with optionally a tweaking Profile if desired) in a fully floating point environment?

Jack

Jack,

 Maybe there is something I am not understanding here. I thought XYZ is basically standard observer cone-response observer space. So whatever the input device, you will need to "squeeze linearised device rgb into XYZ" before, I guess, adopting a whitepoint etc. If your rgb input device is weird, in an ideal world yes you'd probably need negative coefs but you can probably recover the information without negative coefficients and apply some color transform later ...whitepoint adoption means that massaging the data heavily in a color space is mandatory anyway. I'm old and was always dumb, so what would I know ...maybe I should dig out a book

Edmund

[ErikKaffehr]

Hi Jack,

Thanks for comments!

With the introduction of CMOS in MFD there used to be a strong suspicion that CCD would offer superior colour than CMOS. I was a bit skeptical about that, as both are in my view simple photovoltaic devices. But, I thought CFA designs may play a major role.

It has also been a lot of statements on CMOS CFA-s being less orthogonal the CCD CFA designs.

In that sense it may be a good thing to see how basic colour rendition was affected by CFA designs and how those differences could be handled by colour profiles.

These experiments may indicate that colour profiling can achieve a good match between a DALSA CCD and and two Sony CMOS sensors.

But, this is all related to the eye of the beholder. It is well possible that what is "pretty similar" to me may be "widely different" to another observer. I don't know that much about "Pringle chips", but I am essentially sure that MFD chips are made of Idaho Potatoes.

Best regards
Erik

Good show, Eric, you are right, not too different at first glance.

Notice the forward matrices have no negative terms?  That means that they 'desaturate' the raw color information when transferring it linearly into XYZ space.  The non-linear profile takes over from there, massaging the colors until they look as they should (it's an offer they can't refuse). I am not sure I understand the logic behind it 100% (perhaps Edmund or Torger can explain it), but it seems to me that it goes something like this: 'let me squeeze all the colors I can into XYZ, even if colorimetrically incorrectly; I will then try to make them look right by bending, pushing and pulling them to where I know they should be (move over and stay there, damn it)'.

It seems to me that after such an energetic tour de force, trying to determine the difference between 'CMOS vs CCD' CFA color recipes by looking at the final result is a bit like trying to figure out the type of potato used in Pringles by tasting them

Jack

[Jack Hogan]

Maybe there is something I am not understanding here. I thought XYZ is basically standard observer cone-response observer space or at least a 5 degree version of same. So whatever the input device, you will need to "squeeze linearised device rgb into XYZ" before, I guess, adopting a whitepoint etc.
If your rgb input device is weird, in an ideal world yes you'd probably need negative coefs but you can probably recover the information without negative coefficients and apply some color transform later ...whitepoint adoption means that massaging the data heavily in a color space is mandatory anyway. I'm old and was always dumb, so what would I know ...maybe I should dig out a book

You are the man where color is concerned, Edmund, I am just learning as I go along.  Given what we know about color matching etc., aren't there some negative numbers in there?  Every Forward Matrix I ever saw before (process) 2012 had them and did a decent job of converting white balanced raw CC24 data straight to XYZ sort of colorimetrically.

Enter THE PROFILE.  Suddenly, the 'profiled' Forward Matrix produces totally desaturated XYZ data, i.e. pretend that the raw data has been mapped to a teeny weeny gamut, way smaller than sRGB, via perceptual intent.  It is then the job of THE PROFILE to take teeny weeny XYZ data and blow it back up to your output gamut accurately again.  Did I mention that on the way there is a trip from XYZ to ProPhoto, then to HSV, then to ProPhoto, then to XYZ, then to ProPhoto, then to HSV, then to ProPhoto, then to XYZ, then to final color space?  A smart converter may be able to cut a couple of those conversions, but not m uch more than that.  I trust Torger when he says that the stretching and conversions are not detrimental.  But I can't see how this would work accurately unless one is running floating point math. 

Jack