Sony A7RII color accuracy: LR vs C1

[bjanes]

Color scientist and photographer Jim Kasson has been publishing an informative series of articles on his blog. The first article is here. After having provided a background for understanding his methods, he compares the color accuracy of two leading raw converters, Adobe Lightroom and Phase One Capture One when rendering Color Checker images taken with the Sony A7RII. The results are more similar than dissimilar, which should be comforting to users of these two raw converters.

Bill

[AlterEgo]

The results are more similar than dissimilar, which should be comforting to users of these two raw converters.

the results (his conversions) are :

http://s17.postimg.org/p0unhbbjj/C1vs_ACR.jpg



yes, they are similar in the sense that both C1 and ACR do not render (in his example, not what is possible) red as green, green as blue and blue as red... otherwise they are simply very dissimilar for somebody who expects more than above.... so it is all about the level of your expectations...

[Mark D Segal]

Which begs the question about which is "more right than the other". To know this, the comparison that needs to be made is between the file values of the GMCC and the values of the output of the raw converter, but then at what stage? The default settings of the raw converter before any editing I suppose? Do we know to what extent the vendors taste for their view of satisfactory default rendering causes the observed differences? Perhaps Jim answers these questions in his posts - I haven't had time to read it in any detail yet. But until I do so, I wonder how indicative of anything important all this really is, when we have so much editing flex in both these applications, and unless we are doing forensic or medical or commercial work where colour accuracy may be important; otherwise, we are aiming for credible colour appropriate to artistic intent, which will not necessarily be accurate colour.

(Typos corrected)

[ErikKaffehr]

Hi,

What I think we see is that different raw converters go into different directions. I did some comparisons with my P45+ and my Sony Alpha 99 and I have found that both C1 and LR were decently accurate. But in the case of C1 I have seen it taking a lot of liberties with blues.

http://echophoto.dnsalias.net/ekr/index.php/photoarticles/79-p45-colour-rendition?start=3

Best regards
Erik

Which begs the question about which is "more right than the other". To know this, the comparison that needs to be made is between the file values of the GMCC and the vlues of the output of the raw converter, but then at what stage? The default settings of the raw converter before any editing I suppose? Do we know to what extent the vendors taste for their view of satisfactory default rendering causes the observed differences? Perhaps Jim answers these questions in his posts - I haven't had time to read it any detail yet. But until I do so, I wonder how indicative of anything important all this really is, when we have so much editing flex in both these applications, and unless we are doing forensic or medical or commercial work where colour accuracy may be important; otherwise, we are aiming for credible colour appropriate to artistic intent, which will not necessarily be accurate colour.

[AlterEgo]

But in the case of C1 I have seen it taking a lot of liberties with blues.

a simple matrix + trc profile for C1 with a little work (with blues) in C1 ColorEditor (matrix converted to LUT to enable that)

target is imaging resource raw for A7R2 (for ISO100)



that gives in idea how far generic/standard (not C1 CH profiles for certain cameras) are from the real life... linear scientific curve was used

[ErikKaffehr]

Hi,

I did some comparisons between the two a year ago or so. In a way I guess that C1 has higher contrast with it's default "film curve". What I have found was a bit that C1 was pushing blues and to a lesser extents greens a lot into saturation, while LR was more conservative.

Something I have noticed is that C1 sort of goes into eye candy, lot of noise reduction  and a bit of oversaturation, sunny greens and pretty hefty sharpening while LR has conservative defaults. So C1 gives more "pop". On the other hand, C1 is much better at suppressing aliasing artefacts than LR6/ACR.

LR has a well defined colour conversion pipeline. It may not affect users, but I think that a well defined and documented pipeline is quite beneficial for software vendors developing software. But, I got the impression that neither software essentially uses "best practices", relying to much on old coding.

There is an area where LR6/ACR is a great looser for me, and that is demosaic. LR makes a mess of demosaic artefacts and C1 is somewhat better. 

Best regards
Erik

Which begs the question about which is "more right than the other". To know this, the comparison that needs to be made is between the file values of the GMCC and the values of the output of the raw converter, but then at what stage? The default settings of the raw converter before any editing I suppose? Do we know to what extent the vendors taste for their view of satisfactory default rendering causes the observed differences? Perhaps Jim answers these questions in his posts - I haven't had time to read it in any detail yet. But until I do so, I wonder how indicative of anything important all this really is, when we have so much editing flex in both these applications, and unless we are doing forensic or medical or commercial work where colour accuracy may be important; otherwise, we are aiming for credible colour appropriate to artistic intent, which will not necessarily be accurate colour.

(Typos corrected)

[Mark D Segal]

Hi,

I did some comparisons between the two a year ago or so. In a way I guess that C1 has higher contrast with it's default "film curve". What I have found was a bit that C1 was pushing blues and to a lesser extents greens a lot into saturation, while LR was more conservative.

Something I have noticed is that C1 sort of goes into eye candy, lot of noise reduction  and a bit of oversaturation, sunny greens and pretty hefty sharpening while LR has conservative defaults. So C1 gives more "pop". On the other hand, C1 is much better at suppressing aliasing artefacts than LR6/ACR.

LR has a well defined colour conversion pipeline. It may not affect users, but I think that a well defined and documented pipeline is quite beneficial for software vendors developing software. But, I got the impression that neither software essentially uses "best practices", relying to much on old coding.

There is an area where LR6/ACR is a great looser for me, and that is demosaic. LR makes a mess of demosaic artefacts and C1 is somewhat better. 

Best regards
Erik

Erik, what do you mean by "makes a mess of demosaic artifacts". Is this something you actually see looking at prints? I print my files to 13 * 19 inches from my Sony a6000 (APS-C sensor) as demosaiced and prepared in Lightroom, often involving some cropping too; frankly I can look pretty attentively and closely at them without seeing such disturbances. Not to say this means they aren't there just waiting to be discovered with high enough magnification, but I'm raising the question of the threshold of practical concern.

[ErikKaffehr]

Hi Mark,

You make a good point. Although the demosaic artefacts I see a lot of the time are visible pixel peeping on screen they may have a less visibility in prints. I may do a bit to much pixel peeping.

It may be said that aliasing should be handled properly at the sampling level, but photographers prefer visual acuity over correct rendition any time it seems. So, raw processing software needs to take care of issues that should be handled in sampling. I can repost quite a few samples and a few new ones would be anyone be interested.

Yes, I am a bit frustrated by this, so I may over react…

Best regards
Erik

Erik, what do you mean by "makes a mess of demosaic artifacts". Is this something you actually see looking at prints? I print my files to 13 * 19 inches from my Sony a6000 (APS-C sensor) as demosaiced and prepared in Lightroom, often involving some cropping too; frankly I can look pretty attentively and closely at them without seeing such disturbances. Not to say this means they aren't there just waiting to be discovered with high enough magnification, but I'm raising the question of the threshold of practical concern.

[Mark D Segal]

Hi Mark,

You make a good point. Although the demosaic artefacts I see a lot of the time are visible pixel peeping on screen they may have a less visibility in prints. I may do a bit to much pixel peeping.

It may be said that aliasing should be handled properly at the sampling level, but photographers prefer visual acuity over correct rendition any time it seems. So, raw processing software needs to take care of issues that should be handled in sampling. I can repost quite a few samples and a few new ones would be anyone be interested.

Yes, I am a bit frustrated by this, so I may over react…

Best regards
Erik

Why not post some samples - would be interesting to see.

I'd like to pause a moment though to reflect on two concepts I think you consider may be trade-offs to some extent: "visual acuity" and "correct rendition". I have some idea of what "visual acuity" means and how it occurs (sharp lenses and edge contrast between naighbouring pixels), but I am more at a loss on the definition of "correct rendition", and how these concepts could play against each other. Perhaps you could explain......

[sandymc]

LR has a well defined colour conversion pipeline. It may not affect users, but I think that a well defined and documented pipeline is quite beneficial for software vendors developing software. But, I got the impression that neither software essentially uses "best practices", relying to much on old coding.

That was true for earlier versions of LR. Later versions, certainly from V4 on, use "content aware processing", for which there is no documentation. And has varied quite significantly version to version.

[ErikKaffehr]

Hi Sandy,

Thanks for the info. I am aware of the content aware rendition stuff, but I thought it did not affect colour.

Best regards
Erik

That was true for earlier versions of LR. Later versions, certainly from V4 on, use "content aware processing", for which there is no documentation. And has varied quite significantly version to version.

[AlterEgo]

I am aware of the content aware rendition stuff, but I thought it did not affect colour.

I guess this was illustrated clearly = http://blogs.adobe.com/lightroomjournal/2012/02/magic-or-local-laplacian-filters.html

[ErikKaffehr]

Hi Mark,

I will post some more examples, but I need some time to prepare. So I just attach a few screendumps:

The first one is a test shot with my Hasselblad 555/ELD and Planar 100/3.5 at 5 meters distance at f/5.6. It generates a generous amount of colour artefacts, but even removing colour it would generate grayscale aliasing. This is an example of the lens clearly outresolving the sensor. Correct OLP-filtering would help, but I would say better resolution is the right medicine.

The second one is just from a walk in the town, shot with A7rII and a Canon 16-35/4 at f/8. The central part has a lot of Moiré.

The third example is taken from Imaging Resource's test samples from the Canon 5Ds/5DsR. The label on the beer bottle renders with a lot of monochrome moiré in the 5DsR shot, which is barely visible in the 5Ds shot.

So what I mean is that we (including myself) do a bit to much pixel peeping, optimising capture and sharpening for microcontrast but much ignoring lower frequencies which are more significant for perception of sharpness in print.

Just to say, I guess that Jeff Schewe's presets are quite optimal for the A7rII, I would think. Ideal sharpening would, in my humble opinion, aim to:

• Keep MTF for low frequencies as close as possible to 100%, but not exceeding 100%.
• Not really push pixel level sharpness, as this will also push artificial detail and it will not be visible in print.

The last sample shows a decent level of sharpening, LR6 with no sharpening and sharpening in FocusMagic at radius 2 and 75% intensity. MTF gets a bit above 100% at 0.3 cycles/pixel.

My concern is pretty much optimising images for large prints, say larger than 30"x40". The crops I print for testing are corresponding to around 40x60".

Here is a comparison of four raw converters: http://forum.luminous-landscape.com/index.php?topic=94812.0

For some reason the enclosed image download instead of opening, it seems.

Best regards
Erik

Why not post some samples - would be interesting to see.

I'd like to pause a moment though to reflect on two concepts I think you consider may be trade-offs to some extent: "visual acuity" and "correct rendition". I have some idea of what "visual acuity" means and how it occurs (sharp lenses and edge contrast between naighbouring pixels), but I am more at a loss on the definition of "correct rendition", and how these concepts could play against each other. Perhaps you could explain......

[Peter_DL]

the results (his conversions) are :
http://s17.postimg.org/p0unhbbjj/C1vs_ACR.jpg

Once the deltaE results are "poisoned" by a tone curve, which Jim Kasson correctly notes in part 11,
I think it does not really make sense to resort to delta-ab (or dCh) and corresponding chromaticity plots.

Lab is not perceptually uniform, particularly not for vertical trajectories (constant ab).

This can be easily illustrated in Photoshop: take any image to Lab mode to increase Lab lightness and see how the colors are fading. HSB-saturation as well as the perceived saturation are getting reduced.

--

[Mark D Segal]

Hi Mark,

I will post some more examples, but I need some time to prepare. So I just attach a few screendumps:

The first one is a test shot with my Hasselblad 555/ELD and Planar 100/3.5 at 5 meters distance at f/5.6. It generates a generous amount of colour artefacts, but even removing colour it would generate grayscale aliasing. This is an example of the lens clearly outresolving the sensor. Correct OLP-filtering would help, but I would say better resolution is the right medicine.

The second one is just from a walk in the town, shot with A7rII and a Canon 16-35/4 at f/8. The central part has a lot of Moiré.

The third example is taken from Imaging Resource's test samples from the Canon 5Ds/5DsR. The label on the beer bottle renders with a lot of monochrome moiré in the 5DsR shot, which is barely visible in the 5Ds shot.

So what I mean is that we (including myself) do a bit to much pixel peeping, optimising capture and sharpening for microcontrast but much ignoring lower frequencies which are more significant for perception of sharpness in print.

Just to say, I guess that Jeff Schewe's presets are quite optimal for the A7rII, I would think. Ideal sharpening would, in my humble opinion, aim to:

• Keep MTF for low frequencies as close as possible to 100%, but not exceeding 100%.
• Not really push pixel level sharpness, as this will also push artificial detail and it will not be visible in print.

The last sample shows a decent level of sharpening, LR6 with no sharpening and sharpening in FocusMagic at radius 2 and 75% intensity. MTF gets a bit above 100% at 0.3 cycles/pixel.

My concern is pretty much optimising images for large prints, say larger than 30"x40". The crops I print for testing are corresponding to around 40x60".

Here is a comparison of four raw converters: http://forum.luminous-landscape.com/index.php?topic=94812.0

For some reason the enclosed image download instead of opening, it seems.

Best regards
Erik

Thanks very much Erik, and also for the reference back to the four converters thread. I think the examples in this current thread to some extent indicate the importance of good sharpening technique apart from anything else. In the other thread, AccuRaw seemed to provide the "cleanest" rendition of the lot. One needs a huge amount of enlargement and close introspection to see these differences, but "clearly" you show they exist.

[AlterEgo]

Once the deltaE results are "poisoned" by a tone curve, which Jim Kasson correctly notes in part 11,
I think it does not really make sense to resort to delta-ab (or dCh) and corresponding chromaticity plots.

that screenshot illustrates the JK's statement about difference between 2 raw specific conversions : with ACR and with C1 , not between target measured with spectrometer and raw conversion using a certain converter + profile + conversion parameters... JK stated (@dpreview - the origin of the screenshot is there) = "At least with this C1 profile and ASP, this seems to me to be much ado about very little" ... so this illustrates what "very little" actually is... very big.

[bjanes]

Once the deltaE results are "poisoned" by a tone curve, which Jim Kasson correctly notes in part 11,
I think it does not really make sense to resort to delta-ab (or dCh) and corresponding chromaticity plots.

Lab is not perceptually uniform, particularly not for vertical trajectories (constant ab).

This can be easily illustrated in Photoshop: take any image to Lab mode to increase Lab lightness and see how the colors are fading. HSB-saturation as well as the perceived saturation are getting reduced.

Peter,

Your comments and those of Jim Kasson that you cited are quite valid. If a tone curve that only affects luminance is employed, changes in luminance will affect the DeltaE and this is partially overcome by using DeltaCab. The eye is more sensitive to hue than chroma (saturation), and a saturation boost is desired by many photographers to make images pop and a DeltaCab due to a chroma boost may be desired rather than detrimental in some situations. Imatest takes this into consideration with a metric called DeltaCab(corr), which discounts the chroma error.

To illustrate, here is a Colorcheck result for my Nikon D800e using the Adobe Standard profile and default settings. Illumination was from 4700K Solux bulbs. Saturation is boosted as indicated by a chroma (saturation) value of 113.2. On the chart, I interpret a chroma boost as affecting *a and *b to the same extent so that the error vector is radially projected from the white point. A hue error deviates from this radial line. This may not be entirely correct, but that is my interpretation.



One may reduce the chroma error by reducing saturation by 10 on the ACR saturation slider.



Boosting saturation by +10, gives a large DeltaCab, which is partially taken into account by the DeltaCab(corr). The error vector is largely ratially situated from the white point, indicating the chroma boost with relatively little shift in hue. The saturation boost is hardly noticeable on my wide gamut monitor (NEC PA241w, covering approximately AdobeRGB), consistent with Norman Koren's statement that the eye is not that sensitive to high chroma ranges.



Regards,

Bill

[Peter_DL]

Peter,

Your comments and those of Jim Kasson that you cited are quite valid. If a tone curve that only affects luminance is employed, changes in luminance will affect the DeltaE and this is partially overcome by using DeltaCab. The eye is more sensitive to hue than chroma (saturation), and a saturation boost is desired by many photographers to make images pop and a DeltaCab due to a chroma boost may be desired rather than detrimental in some situations. Imatest takes this into consideration with a metric called DeltaCab(corr), which discounts the chroma error.

To illustrate, here is a Colorcheck result for my Nikon D800e using the Adobe Standard profile and default settings. Illumination was from 4700K Solux bulbs. Saturation is boosted as indicated by a chroma (saturation) value of 113.2. On the chart, I interpret a chroma boost as affecting *a and *b to the same extent so that the error vector is radially projected from the white point. A hue error deviates from this radial line. This may not be entirely correct, but that is my interpretation.

https://bjanes.smugmug.com/Photography/D800e-Macbeth/i-Rprdm6Q/0/O/Img_0003_PV2012_def_colorerror.png

One may reduce the chroma error by reducing saturation by 10 on the ACR saturation slider.

https://bjanes.smugmug.com/Photography/D800e-Macbeth/i-ZqNGjTd/0/O/Img_0003_PV2012_def_satNeg10_colorerror.png

… interesting approach by Imatest to average-out and to deduct the overall chroma change,

and you are right Bill that the hue error can still be interpreted in terms of angular deviations
(ignoring some potential twists which might not be relevant here for the small deltas).

Regarding the boost of Lab-chroma it is essentially not possible to say if the Raw converter really adds color saturation (maybe as a side effect from the tone curve), or if it is a combined effect from the tonal changes as such and the Lab metrics:  Lab-chroma is not Exposure-invariant. An increase of Exposure, camera exposure or RGB-linear-scaling, lets the colors walk along vertical diagonal lines in a 3D Lab plot. This is better separated in HSB where hue & saturation remain constant.

So when you undo the increased Lab-chroma here (with the lighter color) by reducing saturation, you might end with a "less than neutral" saturation (reduced HSB saturation). Or, in other words, two colors which sit at the same point in a 2D (L)ab plot, and which are different enough in Lab lightness, can have a different perceived saturation and HSB-saturation.

I've tried to illustrate this in the image below (ppRGB). The inner squares are lighter/brighter versions of the outer squares. On the left side the Lab ab channels are kept constant, on the right side the HSB-hue& saturation are kept unchanged.

The analytical pitfalls here with Lab to interpret the single deltas (beyond just deltaE), finally mirror the challenge and the idea to have color-neutral tone curves which is around since long. There once was Simon Tindeman's approach, and more recently there is torger's DCamProf's neutral tone reproduction operator.


Best regards, Peter

--

[bjanes]

I've tried to illustrate this in the image below (ppRGB). The inner squares are lighter/brighter versions of the outer squares. On the left side the Lab ab channels are kept constant, on the right side the HSB-hue& saturation are kept unchanged.

Best regards, Peter
-

Peter, thanks for the comments and illustration. BTW, your image does not show on my browser (Win 10, Chrome), but the image can  be downloaded and viewed in Photoshop.

Regards,

Bill