von Kries or Bradford white point transform?
Oliver Knoll
a) What is the difference of the von Kries and the Bradford white point
transformation (except for different values in the 3x3 matrices)? Which
one is "better" for what purpose?
b) Is it correct that when transforming RGB with e.g. D65 to XYZ(D65)
and then to L*a*b*(D50) using a L*a*b* white point D50 I _have_ to use
one of the transformations above?
c) Is it correct that when transforming RGB -> XYZ -> L*a*b* using the
same white point I do not have to use one of the transformations above
(just to make sure...)?
Oliver
Sent via Deja.com http://www.deja.com/
Before you buy.
Chris Cox
wrote:
> Just when I thought everything was clear...
>
> a) What is the difference of the von Kries and the Bradford white point
> transformation (except for different values in the 3x3 matrices)? Which
> one is "better" for what purpose?
You use the matrices to convert into and out of cone response space. For
vonKries style transforms you use linear scaling of the values -- and that
could be concatenated into a single matrix. But the Bradford transform
uses non-linear scaling of the (pseudo) cone responses -- this cannot be
made into a single matrix transform.
The Bradford transform is more complicated, but produces better results
than pure vonKries style transforms.
> b) Is it correct that when transforming RGB with e.g. D65 to XYZ(D65)
> and then to L*a*b*(D50) using a L*a*b* white point D50 I _have_ to use
> one of the transformations above?
No, but to get colors with similar appearances you SHOULD.
Otherwise you end up with color casts (yellow, blue, etc.)
> c) Is it correct that when transforming RGB -> XYZ -> L*a*b* using the
> same white point I do not have to use one of the transformations above
> (just to make sure...)?
Correct.
Chris
Frank Wood
news:ccox-ya02408000R...@news.slip.net...
>
> > b) Is it correct that when transforming RGB with e.g. D65 to XYZ(D65)
> > and then to L*a*b*(D50) using a L*a*b* white point D50 I _have_ to use
> > one of the transformations above?
>
> No, but to get colors with similar appearances you SHOULD.
> Otherwise you end up with color casts (yellow, blue, etc.)
>
>
> > c) Is it correct that when transforming RGB -> XYZ -> L*a*b* using the
> > same white point I do not have to use one of the transformations above
> > (just to make sure...)?
>
colours here, although I haven't seen that stated. If not, the D65 or D50,
or any other illuminant specification isn't relevant, in terms of
specification. Any colour has a location on any diagram, RGB, or whatever,
and can be accurately specified, and reproduced if it lies within the system
boundaries.
If we're talking about doing a colour analysis under one illuminant, and
trying to display it with a different illuminant, that's what a TV system
does all the time. Commercial CRTs set their white, in Europe, at D65 (this
may have changed since I was actively involved). But the studio luminaires
run at 3,200K at best. So there's translation to be done, anyway. But in any
analysis/display chain, whether TV, film or whatever, the system parameters
are driven by the display. This is what defines the white point, and, in the
end, the colour space.
Transformations from one colour space to another are a mathematical problem,
and sometimes a tedious one. But ANY colour space can define a particular
colour, even if it can't reproduce it accurately. The PAL analysis uses
negative numbers, and they make the errors less obvious, causing luminance
errors rather than chrominance ones, on the display.
Frank Wood
fr...@woodf-l.dircon.co.uk
Chris Cox
<fr...@woodf-l.dircon.co.uk> wrote:
> "Chris Cox" <cc...@slip.net> wrote in message
> news:ccox-ya02408000R...@news.slip.net...
> >
> > > b) Is it correct that when transforming RGB with e.g. D65 to XYZ(D65)
> > > and then to L*a*b*(D50) using a L*a*b* white point D50 I _have_ to use
> > > one of the transformations above?
> >
> > No, but to get colors with similar appearances you SHOULD.
> > Otherwise you end up with color casts (yellow, blue, etc.)
> >
> >
> > > c) Is it correct that when transforming RGB -> XYZ -> L*a*b* using the
> > > same white point I do not have to use one of the transformations above
> > > (just to make sure...)?
> >
> I feel I've missed something. Presumably, we're talking about reflected
> colours here, although I haven't seen that stated. If not, the D65 or D50,
> or any other illuminant specification isn't relevant, in terms of
> specification. Any colour has a location on any diagram, RGB, or whatever,
> and can be accurately specified, and reproduced if it lies within the system
> boundaries.
We're talking about reproducing a color appearance.
The appearance of a sample (color) is VERY dependent on the illuminant used
to view it, and somewhat dependent on the surrounding colors, background,
and intensity of the illumination.
> If we're talking about doing a colour analysis under one illuminant, and
> trying to display it with a different illuminant, that's what a TV system
> does all the time.
True, but they do it very poorly.
> Commercial CRTs set their white, in Europe, at D65 (this
> may have changed since I was actively involved). But the studio luminaires
> run at 3,200K at best. So there's translation to be done, anyway. But in any
> analysis/display chain, whether TV, film or whatever, the system parameters
> are driven by the display. This is what defines the white point, and, in the
> end, the colour space.
Not even close.
That would only be true in a VERY close system (like early TV broadcast).
Even today, video is stored and edited in a variety of formats and
colorspaces - each with their own strengths and weakness for various tasks.
Chris
Charles Cowens
>
> > Just when I thought everything was clear...
> >
> > a) What is the difference of the von Kries and the Bradford white point
> > transformation (except for different values in the 3x3 matrices)? Which
> > one is "better" for what purpose?
>
Chris Cox wrote:
> You use the matrices to convert into and out of cone response space. For
> vonKries style transforms you use linear scaling of the values -- and that
> could be concatenated into a single matrix. But the Bradford transform
> uses non-linear scaling of the (pseudo) cone responses -- this cannot be
> made into a single matrix transform.
> The Bradford transform is more complicated, but produces better results
> than pure vonKries style transforms.
>
Just to amplify on Chris Cox's answer--correctly, I hope--the Von-Kries
style white point compensation that would be most typical would be the
one described in the ICC spec's Annex A where you scale the X, Y, and Z
values by the ratios of D50 X over Actual White Point X, D50 Y over
Actual White Point Y, and D50 Z over Actual White Point Z, respectively.
Charley Cowens
Frank Wood
As I said, I came late to this thread. And we have argued before.
But a colour can be uniquely defined, without regard to its subjective
appearance.
>
> The appearance of a sample (color) is VERY dependent on the illuminant
used
> to view it, and somewhat dependent on the surrounding colors, background,
> and intensity of the illumination.
True, although I would question the intensity factor, unless it implies a
change of colour temperatrure.
>
>
> > If we're talking about doing a colour analysis under one illuminant, and
> > trying to display it with a different illuminant, that's what a TV
system
> > does all the time.
>
> True, but they do it very poorly.
They do it as well as their system constraints permit. I seem to remember
having this argument earlier.
>
>
> > Commercial CRTs set their white, in Europe, at D65 (this
> > may have changed since I was actively involved). But the studio
luminaires
> > run at 3,200K at best. So there's translation to be done, anyway. But in
any
> > analysis/display chain, whether TV, film or whatever, the system
parameters
> > are driven by the display. This is what defines the white point, and, in
the
> > end, the colour space.
>
> Not even close.
That, I can't buy! When it comes to the crunch, the judgment of the editor,
or whoever, is right. But it's purely subjective, and rightly so. Whether
it's a screen illuminated by a film projector, or a TV screen, what the
customer sees is the end product.. You or I may quibble about its accuracy
in colour reproduction, as engineers. But that has no relecvance to the
customer's acceptance of the product.
Frank Wood
fr...@woodf-l.dircon.co.uk
Chris Cox
wrote:
Actually, NO.
The Von Kries style transforms use a 3x3 matrix to convert XYZ values into
a cone repsonse space, then scale the tristimulus values, then a 3x3 matrix
to convert back. As long as the scaling is linear (diagonal matrix) then
you can concatenate the 3 matricies into a single 3x3 matrix. There are
several VonKries style transforms in use -- each using a different set of
cone primaries (some measured, some exerpimentally derived). Again, the
Bradford transform (and some other recent transforms) is similar, but uses
non-linear scaling and cannot be done with a single matrix.
Scaling the XYZ coordinates by themselves is the WORST way to do white
point compensation. But it is a very common mistake.
Chris
Charles Cowens
>
> > Just to amplify on Chris Cox's answer--correctly, I hope--the Von-Kries
> > style white point compensation that would be most typical would be the
> > one described in the ICC spec's Annex A where you scale the X, Y, and Z
> > values by the ratios of D50 X over Actual White Point X, D50 Y over
> > Actual White Point Y, and D50 Z over Actual White Point Z, respectively.
>
> Actually, NO.
>
> The Von Kries style transforms use a 3x3 matrix to convert XYZ values into
> a cone repsonse space, then scale the tristimulus values, then a 3x3 matrix
> to convert back. As long as the scaling is linear (diagonal matrix) then
> you can concatenate the 3 matricies into a single 3x3 matrix. There are
> several VonKries style transforms in use -- each using a different set of
> cone primaries (some measured, some exerpimentally derived). Again, the
> Bradford transform (and some other recent transforms) is similar, but uses
> non-linear scaling and cannot be done with a single matrix.
>
> Scaling the XYZ coordinates by themselves is the WORST way to do white
> point compensation. But it is a very common mistake.
>
Chris-
Thanks for clearing that up for me. I'll have to stop calling the ICC
Annex A method a Von Kries type transform.
Charley Cowens
DannyRich
>
>Scaling the XYZ coordinates by themselves is the WORST way to do white
>point compensation. But it is a very common mistake.
>
>Chris
>
Funny,
I seem to remember that CIELAB and RLAB always do pretty well in the color
appearance studies published in the literature and all they do is scale the
tristimulus values. Maybe its because linear algebra obeys the associative law
and scaling a cone fundamental that is a linear transform of a tristimulus
value can be re-written as a scaling of the tristimulus values.
Chris Cox
dann...@aol.com (DannyRich) wrote:
> snip
> >
> >Scaling the XYZ coordinates by themselves is the WORST way to do white
> >point compensation. But it is a very common mistake.
> >
> >Chris
> >
>
> Funny,
>
> I seem to remember that CIELAB and RLAB always do pretty well in the color
> appearance studies published in the literature and all they do is scale the
> tristimulus values.
you must be reading different literature.
In all the papers and books I've read, LAB and XYZ are included as baseline
implementations and always score the worst.
> Maybe its because linear algebra obeys the associative law
> and scaling a cone fundamental that is a linear transform of a tristimulus
> value can be re-written as a scaling of the tristimulus values.
Again, that only works if you're using linear scaling.
The most recent models use a non-linear scaling.
Chris
DannyRich
>you must be reading different literature.
>In all the papers and books I've read, LAB and XYZ are included as baseline
>implementations and always score the worst.
>
>> Maybe its because linear algebra obeys the associative law
>> and scaling a cone fundamental that is a linear transform of a tristimulus
>> value can be re-written as a scaling of the tristimulus values.
>
>Again, that only works if you're using linear scaling.
>The most recent models use a non-linear scaling.
>
>Chris
XYZ is not a color appearance space so it is irrelevant. The nonlinear
adaptations transforms are only applied to the Blue channel and this is because
in an appearance metric, brightness and colorfulness come into play rather than
just luminance (lightness) and chroma, and brightness is nonlinear. But the
studies from CIE Technical COmmittee 1-27, of which I am a member, have
reported consistently that the advantages of the nonlinear transforms are
slight and while CIELAB rarely wins any contents, the difference in performance
is usually only just less than statistically significant. The nonlinear
adjustment to the Blue channel proposed in the Bradford model, carries with it
the disadvantage that it is not easily inverted. Thus one can transform from A
to D50 but going back to A is not an exact transform but can be approximated to
within the experimental error of the appearance data - which is considerable.
For an example of the application of a simple von Kries transform check out the
papers on the OSA Uniform Color Scales, the acknowledged, most uniformly space
color order system in the world today.