Different film profiles and the uncanny similarity of the result...

[ThomasH]

As I was scanning Agfa film using different *.icc profiles,
I made an interesting observation. One *.icc profile was
generated by Vuescan from a target made on film strip by
the Swiss Colour Science:

http://www.colour-science.com/

However, this target was made by exposing the reflective
target manufactured by Wolf Faust and its Coloraid.

http://www.targets.coloraid.de/

As soon Wolf has provided the high dynamic range target
for RSX-II and CT-Precisa Agfa called A3, I obtained this
target too and made *.icc scanner profile using Vuescan.

Of course I saw on the histograms of these both targets
that the coloraid target was having much larger dynamic
range, foremost black was darker. To my surprise I observe
almost identical histogram of corrected image while using
these both profiles on the same images. Here is one sample:

http://www.pbase.com/phototalk_thh/agfa_icc_compare

I was expecting some visible difference! Maybe the dynamic
range of the target is after all not the primary factor in
determining its color correcting properties??

In that case the credit at Colour Science goes for figuring
out how to expose the reflective target (what light source
should be used??)

Thomas

 

[Don]

One *.icc profile was generated by Vuescan ....
Of course I saw on the histograms of these both targets
that the coloraid target was having much larger dynamic
range, foremost black was darker. To my surprise I observe
almost identical histogram of corrected image while using
these both profiles on the same images. Here is one sample:

http://www.pbase.com/phototalk_thh/agfa_icc_compare

I was expecting some visible difference! Maybe the dynamic

I have to preface this by saying I haven't looked at the image nor am
I familiar with those profiles but purely from a pragmatic standpoint:

Could it be that your test image has no data in that particular band
of dynamic range in which the profiles differ most?

In that case it's not at all surprising that the histograms of both
images would be similar.

Of course, Vuescan is notoriously buggy and therefore totally unsuited
for this type of testing. Its scans are so badly corrupted as to
completely obscure any differences. Furthermore, each Vuescan version
is likely to produce radically different results.

A useful test would be to repeat the scans with a reliable scanner
program. That would test both the profiles and Vuescan.

Don.

 

[Wolf Faust]

Of course I saw on the histograms of these both targets
that the coloraid target was having much larger dynamic
range, foremost black was darker. To my surprise I observe
almost identical histogram of corrected image while using
these both profiles on the same images. Here is one sample:

You can expect cut-offs at the highlights or shadows with many
profilers if the target used does not reflect the medias density range
decently. Obviously you are not going to the limits with your test
image nor does the rather simple VueScan profile cut-off things easily
(one could say this is an advtange of the more simple VueScan
profile). Try scanning the large gamut IT8 coloraid target with the
"Swiss" profile and watch what happens to the colors in column 4,8,12
and DMin/DMax.

What you should also look at, is the difference of each primary color
of the histogram relative to the other colors (especialy close to the
min/max values). Difference here mainly become visible as unwanted
color cast and when looking at your histgrams, there are well visible
differences in some areas.

I would recommend you better use something like the free lprof
qtProfileChecker to do a quick check of a profile. qtProfileChecker is
part of the lprof 1.09 package and download links can be found on
http://www.coloraid.de. With some mouseclicks you can get various dE
values for test patches provided to the profile.

I was expecting some visible difference! Maybe the dynamic
range of the target is after all not the primary factor in
determining its color correcting properties??

Density range is important once your image goes to the limits... but
after all, the colors are mainly defined by the spectral behaviour of
the dye, light and scanner sensor used. Look more closely at your
histograms and you will see the difference is well visible. And maybe
take a better profiler like lprof... and a test image that really
stretches things to the limits (dark shadows, highlights, highly
saturated colors...).

A final note to an Agfa user ;-) : store your Agfa films and target in
a cold, dark and dry environment. This surely applies to all slide
films, but especialy to Agfa slide films if you want to maintain color
accuracy for a long time. The Agfa specs are not really much helpfull
in this respect to users. Better read the Kodak publ. E-30 on how to
store photographic materials for a long time ;-)

 

[Roger S.]

Hi Wolf, I checked my Vuescan profile with qtprofilechecker (it crashes
a lot!) but don't know how to interpret the results.
This is what I pasted from the program:

Media white (XYZ): 96.43, 100.00, 82.51
Primaries: R:0.63, 0.32 G:0.32, 0.68 B:0.1, 1.9e-005
Estimated gamma : R:1.02, G:0.98 B:1.04

Average dE :-1.#J
Standard deviation :-1.#J
Peak : 0.00
Min : 10000000000.00

From the curves and "estimated gamma" it looks like the midtones aren't

neutral and are a bit too green but the endpoints look good. Does this
seem acceptable or should I try creating a profile with the lcms
profiler? It seems to like to crash as well.

 

[ThomasH]

You can expect cut-offs at the highlights or shadows with many
profilers if the target used does not reflect the medias density range
decently. Obviously you are not going to the limits with your test
image nor does the rather simple VueScan profile cut-off things easily
(one could say this is an advtange of the more simple VueScan
profile). Try scanning the large gamut IT8 coloraid target with the
"Swiss" profile and watch what happens to the colors in column 4,8,12
and DMin/DMax.

What you should also look at, is the difference of each primary color
of the histogram relative to the other colors (especialy close to the
min/max values). Difference here mainly become visible as unwanted
color cast and when looking at your histgrams, there are well visible
differences in some areas.

I would recommend you better use something like the free lprof
qtProfileChecker to do a quick check of a profile. qtProfileChecker is
part of the lprof 1.09 package and download links can be found on
http://www.coloraid.de. With some mouseclicks you can get various dE
values for test patches provided to the profile.


Density range is important once your image goes to the limits... but
after all, the colors are mainly defined by the spectral behaviour of
the dye, light and scanner sensor used. Look more closely at your
histograms and you will see the difference is well visible. And maybe
take a better profiler like lprof... and a test image that really
stretches things to the limits (dark shadows, highlights, highly
saturated colors...).

Thanks for the interesting insights, this week I am swamped with
my work, maybe on weekend I will relax some more with my hobby!

You are probably correct about the dynamic range, I posted
histograms of a very hazy, pastel colored image, which clearly
does not carry the full dynamic range. I picked this image at
random because all the histograms looked similar to me, but
a profile checker will provide a more objective verdict...

A final note to an Agfa user ;-) : store your Agfa films and target in
a cold, dark and dry environment. This surely applies to all slide
films, but especialy to Agfa slide films if you want to maintain color
accuracy for a long time. The Agfa specs are not really much helpfull
in this respect to users. Better read the Kodak publ. E-30 on how to
store photographic materials for a long time ;-)

At least to me all this known!! All my targets are in
air sealed box and in the fridge! Before I use them they
stay in the closed box in the room temerature for 8h.

Thomas

 

[Wolf Faust]

Hi Wolf, I checked my Vuescan profile with qtprofilechecker (it crashes
a lot!) but don't know how to interpret the results.
This is what I pasted from the program:

First time I hear of a such a serious problem with the package and I
find Marti's source always rather well written ;-) I am afraid your
listed results are useless and only show that something is going
completly wrong. Did you try generating a profile using lprof and see
if that profile works with qtprofilechecker. Maybe the problems are
caused by the VueScan profile?

 

[Roger S.]

The crashes are primarily due to where I store files on my hard drive-
the program doesn't seem to like files stored more than 3 levels deep
or so (like normal monitor profiles in WinXP, etc).

Anyway, I made a Vuescan 8 bit raw file with Device RGB as the output
space (not sure if it matters) of my Wolf Faust provia target. I used
the measurement tool and then the profile creator with resolution of 33
and with the local convergence analysis on.

Here are the results. I input the profile name, the target vendor and
type and the measurement sheet into Profile Checker. Is this right?

LCMS
Average dE : 0.42
Standard deviation : 0.36
Peak : 2.70
Min : 0.05
Average Target error: (99% confidence) 2.08
Maximum error induced by target itself: 3.15

I didn't have a measurement sheet for Vuescan because I'm not outputing
a file when I let Vuescan create the profile. Should I use the same
RAW scan measurements to check a Vuescan profile? If I do so, here is
the result:

Average dE : 23.55
Standard deviation :8.38
Peak :50.98
Min : 3.14
Average Target error: (99% confidence) 2.08
Maximum error induced by target itself: 3.15

The CIE diagram looks like a bunch of shooting stars...

This is what I got doing a not raw scan with LCMS (target scan's color
balance set to none and Adobe RGB output space). Same as the first
example but with a different scan.

Average dE : 0.45
Standard deviation : 0.36
Peak : 2.74
Min : 0.03
Average Target error: (99% confidence) 2.08
Maximum error induced by target itself: 3.15

 

[ThomasH]

You can expect cut-offs at the highlights or shadows with many
profilers if the target used does not reflect the medias density range
decently. Obviously you are not going to the limits with your test
image nor does the rather simple VueScan profile cut-off things easily
(one could say this is an advtange of the more simple VueScan
profile). Try scanning the large gamut IT8 coloraid target with the
"Swiss" profile and watch what happens to the colors in column 4,8,12
and DMin/DMax.

What you should also look at, is the difference of each primary color
of the histogram relative to the other colors (especialy close to the
min/max values). Difference here mainly become visible as unwanted
color cast and when looking at your histgrams, there are well visible
differences in some areas.

I would recommend you better use something like the free lprof
qtProfileChecker to do a quick check of a profile. qtProfileChecker is
part of the lprof 1.09 package and download links can be found on
http://www.coloraid.de. With some mouseclicks you can get various dE
values for test patches provided to the profile.

I hope I tested it coirrectly...

This is the info of the CT100 profile made from the swiss film strip:

Profile: VueScan Device Profile
VueScan Device Profile
(c) 2003 Hamrick Software
WhitePoint : D50
E:/Foto/Ag_CT100_LS4000_NS4.icc
Media white (XYZ): 96.43, 100.00, 82.51
Primaries: R:0.57, 0.33 G:0.34, 0.66 B:0.14, 1.8e-005
Estimated gamma : R:2.46, G:2.59 B:2.57

And this is the profile made from your Agfa RSX II target:

Profile: VueScan Device Profile
VueScan Device Profile
(c) 2003 Hamrick Software
WhitePoint : D50
E:/Foto/RSX100_II_LS4000_NS4.icc
Media white (XYZ): 96.43, 100.00, 82.51
Primaries: R:0.63, 0.33 G:0.35, 0.65 B:0.12, 0.019
Estimated gamma : R:2.38, G:2.67 B:2.58

I see also a "De Report," which is very erratic, probablt garbled:
In both cases it looks like this:

Average dE: -1.#J
Standard deviation: -1.#J
Peak: 0.00
Min: 100000000.00

(number of nulls might be wrong, cut and paste
does not work for this data...)

I both cses the curces look very similar, all 3
compoments are virtually atop of each other. They are
actually much closer in the CT100 film strip case.

It appears to me from CIE diagrams that your profile yields
much wider color gamut compared to the profile made from
the film strip. I made screen dumps from lpchecker and added
them to the same gallery:

http://www.pbase.com/phototalk_thh/agfa_icc_compare

Thomas

 

[Wolf Faust]

The crashes are primarily due to where I store files on my hard drive-
the program doesn't seem to like files stored more than 3 levels deep
or so (like normal monitor profiles in WinXP, etc).

My lprof is installed on Windows 2000 SP4, also in a rather "low"
directory. But now that you mention it, there is a problem that causes
a crash if the directory (and subdirectories?) specified for the
reference files is also containing other non-reference files. So it is
not a good idea to put reference file and target scan in one
directory. Rather copy the reference files into the prepared
subdirectory of lprof.

Here are the results. I input the profile name, the target vendor and
type and the measurement sheet into Profile Checker. Is this right?

LCMS
Average dE : 0.42
Standard deviation : 0.36
Peak : 2.70
Min : 0.05

These are rather low error values for a slide film profile ;-) One
thing to remember with these values is: high error values indicated
by the profilechecker are a good inidication that something is wrong.
But good error values aren't necessarily an indication that there is
no problem unless you do test with much more testing patches instead
of using the target patches. I usualy use >1100 test patches to get a
more exact idea about the profile quality with the profilechecker.
Also the average dE is not a good indicator for profile quality... but
let's forget about the details now and simply assume the values apply
to all colors ;-)

Note:

lprof is now on Sourceforge: http://sourceforge.net/projects/lprof
with version 1.10 for Linux. Seems like there are not any RPM or apt
binary archives yet.

 

[Roger S.]

Hmm, based on what you said the profile may be okay but this program
doesn't really tell me.
Back to the real world, neither of the Lcms profiles work as well as
the Vuescan one for a few slide scans I tried. The Vuescan profile
gives much more accurate colors for the scan bakeoff slide I have and
for a couple others I used to test it. The best results come from
using both the Vuescan profile and setting the color balance in VS
using a grey card shot under flash.

 

[Wolf Faust]

It appears to me from CIE diagrams that your profile yields
much wider color gamut compared to the profile made from
the film strip. I made screen dumps from lpchecker and added
them to the same gallery:

http://www.pbase.com/phototalk_thh/agfa_icc_compare

The gamut displayed here basicly shows the RGB scanner primaries
extrapolated from the target data. You can assume that the gamut of
the scanner and thus profile is inside that area, but it will hardly
be that large. So this is hardly of practical use.

Better is another method:

Simply feed the target patches through the lprof profilechecker.
Connect the outmost patches shown in the diagram. Assuming that the
target is produced reaching the medias gamut (or being very close to),
you do get the maximum gamut achived using that scanner with the
tested media. And you do see a good approximation of the target gamut.

Sometimes you can see interesting things. For instance, some Nikon
film scanners have a limited gamut in the high saturated greens. The
profilechecker shows you the reference values of the target for the
green patches and where the profile maps them to. But if the
reference values are outside the RGB triangle or if the patches are
all mapped to an imaginary straight line inside the triangle, one can
assume there is some form of gamut limitation causing the patches to
be mapped to that line. So, here you can expect problems with your
scanner that are most likely hardware limited. The only question
remaining is now, how does the profile deal with the scan limitations.
The design of the coloraid targets here is to avoid spreading the
error caused by colors outside the gamut into good areas inside the
gamut and to avoid errors in color hue. This is basicly done by
adding target patches in areas close to but not outside the gamut of
the troubled scanners in the custom target area.

In order to compare gamuts, you should also use a 3D display of the
gamut. A 2D display can be rather misleading. For instance, you can't
see differences in the also important DMin/DMax area. A 3D view can be
done for instance using ColorThink or ColorShop X. A time limited but
fully function demo of ColorShop can be downloaded from X-Rite. I can
provide ColorShop X files for most coloraid targets if you haven't
received them already with the targets. Simply load the data into the
scratchpad of ColorShop and drag them into the 3D viewer. Be warned:
the 3D gamut displayed this way might be smaller than it really is
because ColorThink simply connects patches that seem to be on the
gamut boundary. Especialy with some green colors on the coloraid
target (see comment above) this leads to a wrong display that can be
avoided by only using the color patches that are really on the gamut
boundary (ie. patches in column 4,8,12,DMin, DMax,...)

The Swiss target hardly reaches the gamut of the media and it is
unlikely the gamut is large enough reach the scanners gamut limit. So
here you most likely only see a good approximation of the target
gamut. This doesn't mean you can't scan also the colors outside that
gamut, but it's very likely you run into visible problems.