Bill Atkinson's target vs i1Profiler's target

[Tarkowsky]

In a previous topic I was strongly suggested by many to use i1Profiler targets which are supposedly better then Bill Atkinson's targets especially if one has to use i1Profiler to create ICC profiles.
So I decided to do a simple  test with my Epson 7900 printing the biggest targets (5202 patches for BA and 6000 patches for i1Profiler) on Canson Rag Photographique paper and compare the two profiles using ColorThink and also printing QuadTone 21 gray steps to check neutral values along the gray axis and the darkest black printable.
To measure such big targets  I used a i1iSis XL and I measured three times each target averaging the results.
As for the 21 QuadTone  gray steps I used a i1Pro following the same procedure. .

Here are the results:
 
Gamut:
BA 5202 patches = 617'045
i1P 6000 patches = 610'026

Density:
BA 5202 patches = 1.57
i1P 6000 patches = 1.56

Black:
BA 5202 patches Lab = 18.9  0.2  0.4
i1P 6000 patches Lab = 19.2  0.1  0.0

As we can see BA target gives a slightly bigger gamut and  density.
Checking  for better neutral values along the gray axis i1P 6000 patches gives a better neutralization.

I printed some images (using intent relative colorimetric only) and couldn't see much differences. but I should do more test.

Difficult to say if  there is a winner.

One can notice that BA targets are old and  quite simple in their inner structure whereas i1Profile targets are the outcome for X-rite 6 years long quest for the ultimate target (PM5 was released in 2005 if I am not wrong).
Well if there has been an improvement it's is almost infinitesimal.


Tark

[digitaldog]

When I tested the two options, I could see a better print from the i1P produced profile target than the older BA target. Reds were cleaner (less orange).

[shewhorn]

When I tested the two options, I could see a better print from the i1P produced profile target than the older BA target. Reds were cleaner (less orange).

Hmmm... I'm curious to know why there would be much, if any difference at all? With 5202 patches under i1Profiler and the Atkinson patches, only 72 of those patches are really different. Both Bill and X-Rite from what I have observed are pretty much selecting equally spaced patch values unless you're doing an optimization in i1Profiler. They both use values that are numerically equally spaced (Bill uses 16 bit precision, X-Rite truncates the precision or rounds down) counting in a trinary sequence. At 5202 patches you'd have 17^3 values or 4913 patches + 289 additional patches. In Bill's targets those 289 patches are actually just dedicated to greyscale of which there are only 256 discrete values, he just doubles up on a few numbers to make the chart visually appealing. Once I started looking at the actual numbers in the Atkinson profiles I became baffled as to why everyone thinks they're so great or magical (when I first got into this I was lead to believe (by reading the forums) that there was something special about Bill's profiles). Aside from the visual aesthetic of the charts and how they are laid out, there isn't really anything special about them (which got me thinking that maybe a lot of people were seeing a placebo effect?... I'll admit to being guilty of that from time to time :lol: ). I generated the numbers below in a spreadsheet and they are pretty much exactly the same numbers that the Atkinson profiles use (the Atkinson profiles only go out 3 decimal places).

0
15.9375
31.875
47.8125
63.75
79.6875
95.625
111.5625
127.5
143.4375
159.375
175.3125
191.25
207.1875
223.125
239.0625
255

And i1 Profiler would use this set (just truncating the precision, again, I generated the numbers in a spreadsheet and when I crossed checked them in the pxf file... they are identical)

0
15
31
47
63
79
95
111
127
143
159
175
191
207
223
239
255

So... if you generate a 5202 patch chart (which for testing purposes would be the comparison I would want to make) in i1Profiler, 4913 patches are going to be identical. The difference in your charts is going to come down to 289 patches. Actually if you further that analysis you will see that i1Profiler also generates a string of greyscale patches. For 5202 patches, i1Profiler specifically has 217 steps of greyscale patches so ACTUALLY... the difference between an Atkinson profile and an i1Profiler generated profile comes down to 72 patches out of 5202. 5130 of those patches are going to be the same. If you look at those 72 patches under i1Profiler, they are specifically applying a slight bias around the greyscale so if anything the difference you should see with the i1Profiler targets would (in theory) be more neutral black and whites as it has a comparison base for what is exactly x,x,x (where x is the same number) and what is slightly outside of x,x,x.

I actually have my own set of charts that I use, the numbers for which I've generated in a spreadsheet. I'm doing something very similar to what X-Rite is doing with i1Profiler (sometimes they make a difference and sometimes they don't).

Cheers, Joe

[digitaldog]

Hmmm... I'm curious to know why there would be much, if any difference at all?

Because for one, I tested the 1728 patch target from Bill versus the same number of patches generated by i1P.

[shewhorn]

Because for one, I tested the 1728 patch target from Bill versus the same number of patches generated by i1P.

But why the difference? That's what's piqued my curiosity. 1728 patches from Bill vs. i1P should be almost identical (the integers will be identical, the i1Profiler generated targets won't have the floating point). If i1Profiler can properly handle 16 bit reference data then the results should be identical. That they're not leads me to believe that i1Profiler can't handle 16 bit reference data (it is taking 16 bit measurements, it's just not seeing the 16 bit reference so that introduces some error, specifically I think it's truncation error).

If you look in /Library/Application Support/X-Rite/i1Profiler/ColorSpaceRGB/TestCharts you'll find that for every chart i1Profiler stores two files for each chart, a txt, and a txf. The txt file is a 16 bit CGATS .5 file and contains 16 bit reference data BUT... the txf file is in CXF format it only uses 8 bit reference data. (I know you know all this... just filling in the details for those who don't). Because i1Profiler only generates 8 bit charts I have a sneaking suspicion that it can't deal with 16 bit reference data (even if measured in another app like Measure Tool) so here's what I think is happening...

Let's say you have reference data of 127.99 and this is printed in the chart as 127.99. You then take a measurement and the value is read as 128.9, save it as a CGATs file and drop it on the measure module to import it. I think i1Profiler is taking the reference and truncating the 16 bit float to an 8 bit integer so now i1Profiler sees the reference number as 127, not 127.99. Now of course it's comparing an 8 bit reference to a 16 bit measurement. Had the chart been printed and referenced at 8 bits instead of a measurement of 128.9 you'd have a reference (and patch) that was 127 and you'd get a value of 127.9 but now by using the Atkinson profile we've introduce an error of 1.9 not 0.9 and over an entire profile... I betcha that adds up to be rather significant.

Just a theory.

I know that if I generate a chart in Color Port with 8 bit data using identical numbers to what i1Profiler produces there is absolutely zero difference.

So to sum up... I think i1Profiler is introducing truncation error when it's fed a file that uses 16 bit reference data.

Cheers, Joe

[Tarkowsky]

If you look at those 72 patches under i1Profiler, they are specifically applying a slight bias around the greyscale so if anything the difference you should see with the i1Profiler targets would (in theory) be more neutral black and whites

That's exactly what I noticed and its the only aspect where i1P target based profiles outperform BA profiles: better grayscale neutralization.

What is a bit disappointing is that both target s as well as all commercial charts used for profiling devices have the same approach on how to sample the printer space:
the generation of  a regular or semi-regular orthogonal grid of printer test values.

Some years ago (2004?) I remember reading an article by Gill W. Graeme where he stated

 "It seems that mathematically such a rectangular sampling is not the most efficient way of exploring a multidimensional space, so an alternative based on a simplex sampling device space has potential for improving the efficiency of device profiling requiring fewer test point for a given profile accuracy, or returning a higher profile accuracy for a given number of test points".

Alas nobody seems to have done further researches in that particular direction and as far as I know simplex sampling has been put aside somehow.
That's a pity.

Tark

[digitaldog]

But why the difference? That's what's piqued my curiosity. 1728 patches from Bill vs. i1P should be almost identical (the integers will be identical, the i1Profiler generated targets won't have the floating point).

I don’t know. Now I’m wondering....

My target was scrambled, Bill’s isn’t. I regenerated a non scrambled target to match Bills in i1P, compared the two in ColorThink and see that the differences are tiny (Max dE:   0.49). Could it be the patch size? Maybe better sampling from the iSis? And X-Rite claimed before I tested this that the new target generator was ‘better’ and I did see better reproduction of reds in the two profiles. But perhaps the better comment was compared to MeasureTool’s generator. Its interesting and a bit odd that when I ask for 1728 patches unscrambled, I get Bill’s target!

In terms of txt versus txf, I’m using the txt of course in ColorThink. It can’t accept that txf doc.

[shewhorn]

"It seems that mathematically such a rectangular sampling is not the most efficient way of exploring a multidimensional space, so an alternative based on a simplex sampling device space has potential for improving the efficiency of device profiling requiring fewer test point for a given profile accuracy, or returning a higher profile accuracy for a given number of test points".

Fascinating. Do you know where to find the original article (I'd be interested in reading it)? Had to do a bit of searching on that one. The math for the time being is over my head (love math, I'm just not very good at it). I'm curious to know what advantage that would really have? Obviously I don't know what I'm talking about here but here's my line of thinking... given a sufficiently large set of patches and given the linearity of modern printers is there really an advantage to be gained? Perhaps with a smaller patch set it might be better optimized (say, less than 500 patches on something like an x300 series Canon or x900 series Epson)? Assuming we don't know the response characteristics of the printer (we discover that after we take the measurements) I'm wondering how much of a difference the distribution of patch values would have? It seems as if both X-Rite and Argyll have gone in the direction of optimizing patch sets based on the response of an existing profile. In my own tests however the benefits of such optimizations (both for i1Profiler and Argyll) are trivial if you're operating from a larger patch set (over 1000 patches). Seeing as it's faster to just run a chart with more patches instead of starting with a smaller patch set, scanning it in, and then generating another patch set for optimization from a workflow standpoint (at least in terms of working with professional large format inkjet printers, I imagine there are exceptions for different technologies) it doesn't seem to be all that beneficial. (btw, not making concrete factual statements here, I'm just thinking out loud)

Cheers, Joe

[shewhorn]

I don’t know. Now I’m wondering....

.....(snip)

In terms of txt versus txf, I’m using the txt of course in ColorThink. It can’t accept that txf doc.

[SPOCK] Fascinating [/SPOCK] Should be easy enough to modify the txt CGATS.5 file in a spread sheet to strip off the floating point precision to see what you get.

Cheers, Joe

[Alan Goldhammer]

Some years ago (2004?) I remember reading an article by Gill W. Graeme where he stated

 "It seems that mathematically such a rectangular sampling is not the most efficient way of exploring a multidimensional space, so an alternative based on a simplex sampling device space has potential for improving the efficiency of device profiling requiring fewer test point for a given profile accuracy, or returning a higher profile accuracy for a given number of test points".

Alas nobody seems to have done further researches in that particular direction and as far as I know simplex sampling has been put aside somehow.
That's a pity.

Tark

You have the author's name flipped.  It's Graeme Gill and he is the one-man team behind ArgyllCMS.  I think he's incorporated this thinking into Argyll.

[Tarkowsky]

Do you know where to find the original article (I'd be interested in reading it)?

Here is the link where you can buy the article.
http://www.imaging.org/IST/store/epub.cfm?abstrid=32145

Tark

[Jalok]

It seems as if both X-Rite and Argyll have gone in the direction of optimizing patch sets based on the response of an existing profile. In my own tests however the benefits of such optimizations (both for i1Profiler and Argyll) are trivial if you're operating from a larger patch set (over 1000 patches).

I've done a simple test: build a profile from a 513-patch target, another from 1026 patches, and a third one from a 513-patch optimization target over the first profile. The best of all three profiles was the one built from the 1026-patch target. The 513-patch optimized profile came in second, just a little better than the original 513-patch profile. Another test I made some weeks ago was to build a 1000-patch optimized profile over an original 2000-patch original one. The quality increment were marginal, but still visible on some saturated shadows of a given image which i1Profiler extracted some patches from. I wish it could be extracted more than 30 patches. Maybe it could be more valuable than the generated optimization patches.

[digitaldog]

I've done a simple test:

Mimics what I saw too. If you start with a small patch sample, optimization is useful. But if you start with a larger sample it doesn’t. I was really hoping for more from this technology.

[Rhossydd]

I was really hoping for more from this technology.

This begs the question is there much more to be had anyway?
From what I've seen, and other's experiences I've read, i1Pr profiles aren't very much different to those from other good profiling software (PMP5, MP, Argyll). Do you really think there's much lacking from current profiles ? Yes, they're a little better in some respects, but there will be a lot of photographers who wouldn't see the difference in their prints.

Yes, the creation software could be easier to use, faster, more versatile and more reliable (all of which we'd hoped to see in i1Pr), but is there much more we can expect from the final end profile ?

[digitaldog]

This begs the question is there much more to be had anyway?

Considering the title of the thread, yes. Bill Atkinson did this years ago when he built his profiles for the Epson’s whereby he used off the shelf software and code of his own to produce iterations that improved gray balance of the iterative profile by examining the gray axis of the first profile, building a new target and readjusting the gray axis of the 2nd profile. Some of us were hoping to see this in i1P, to the point we could perhaps build a special B&W weighted profile using a perceptual intent tweak for such output.

[Rhossydd]

Considering the title of the thread

Quite......from what we're read here there seems not a lot of difference which charts one uses now.
Maybe a tiny difference ? Is that a consistently repeatable improvement(difference) ? or are we getting into the realms of measurement and calculation differences ?

Some of us were hoping to see this in i1P, to the point we could perhaps build a special B&W weighted profile using a perceptual intent tweak for such output.

Did you think your previous profiling system was lacking very much and in need of improvement ?

[digitaldog]

Did you think your previous profiling system was lacking very much and in need of improvement ?

I think Bill’s approach proved there could be improvement in gray balance for ink jets who’s behavior is less than ideal for such output.