[argyllcms] Re: Determining proper error value for -r

Ben Goren wrote:

Before  I go  any further,  the only  really important  part. What
``average deviation as a percentage'' does the following sample of
values  represent? These partial  readings are  of the  same color
(paper white, in this case) in different patches on the same sheet
of paper, and the last line is the standard deviation.

  LAB_L   LAB_A   LAB_B   400     410     420     430     440
  94.69   2.28    -9.05   27.72   49.83   84.45   104.77  109.47
  95.23   2.32    -9.36   27.92   50.59   86.03   106.81  111.64
  95.37   2.32    -9.29   28.09   50.88   86.36   107.08  111.88
  95.3    2.29    -9.35   27.62   50.39   86.09   107     111.85
  95.22   2.31    -9.21   28.2    50.71   85.86   106.5   111.29
  95.2    2.34    -9.25   27.45   50.21   85.82   106.61  111.38
  95.37   2.28    -9.35   27.92   50.73   86.4    107.23  112.02
  95.23   2.31    -9.3    27.88   50.41   85.84   106.69  111.54
  0.22    0.02    0.11    0.25    0.34    0.61    0.77    0.81

(That's the readings of but one  ``color'' patch from a chart with
a  few dozen  such. If  you  need the  full  set  of the  spectrum
readings, just give a holler.)


Well, the whole thing is a little "loose", because
it's not actually that critical in the scheme of things.
I calculate the average deviation (absolute) as
0.08, 0.02, 0.05 for L, a and b respectively. Since
the range of L is 0 to 100, this translates
roughly into a percentage.

I'm  pretty confident  that  I  should be  able  to  get a  ``good
enough'' profile to  make my mom happy, but only  if I can conquer
-r. I think  I know enough to  be able to do  that, now...but I'll
have to  spend another  afternoon at their  place to  do it. Since
they're  good people  and the  food  is wonderful,  that's not  so
terrible...I'd just rather spend the time there on something other
than the computer....


I think this is complete overkill. The value is not that critical.
Unfortunately the default interpretation of the -r value in the
0.53 release was too low by a factor of 10, hence the mention
of it since that release. In the 0.54 release (yes I'm still trying to
finish it off), you will get perfectly acceptable results using the
default in 99% of situations. In the 0.53 release, if you set -r to
1.3 or so, it should be fine for 99% of all situations too.

What  I'm  envisioning for  the  future  is a  two-pass  profiling
process, not  noticeably different  for the  end user  from what's
currently used for high-end profiling. I  think it has a chance at
giving  better  results for  high-end  profiling,  and I'm  almost
certain it'll  make profiling  poorly-behaved devices  pretty much
straightforward.


Sorry, I don't agree. While there may be some merit in a more
complicated approach for high end profiling (still to be proven),
I can't see how it is relevant to low end devices. There's
no point trying to get that extra 1 delta E accuracy
somewhere in the colorspace, if the device is banding, has
inks that change by 3 delta E in the first 24 hours, or
has a repeatability of +/- 2-6 delta E.

..snip.. 
So, in short, the end user does much the same thing as before, but
Argyll is doing  the extra step of figuring out  where to be fuzzy
and where to be precise.


I think these are all reasonable ideas, if there was an indication
that it would make some difference. I'm not so confident. My attempts
at trying to establish a test environment to check such effects,
has simply lead me to the conclusion (at the moment), that the
random effects often overwhelm my ability to draw conclusions.

For instance, I created a test set of 7000 RGB patches for
my Epson 1800. I then used that as a reference to test the effects
of varying some parameter for a smaller test set (say 1000 patches).
If I'm lucky, I see that the average error of the profile (measured
against the 7000 patch reference), changes from 3.3 delta E to
3.1 delta E, with the worst case error changing from 8.7 to 9.2. What
am I to make of this ? The change is so small compared to the magnitude
of the basic error, that I have little confidence in drawing conclusions.
It's not like the errors drop from 3 to 1 or something obvious. I'm lucky if
there is a shallow curve. If I'm really luck, the same trend is visible
if I try a chart with 1001 patches in it (ie. that the results aren't being
very sensitive to every incident factor).

A statistician might say something like "you need to repeat your
reference set 10 times, and your experimental test 10 times and
average the results, to reduce the concealing effect of
the random errors on any systematic effect", but of course
this isn't very practical advice.

I'm sure these sorts of issues can be improved on, with sufficient
knowledge, ingenuity and persistence, but it's the sort of project
that would normally take some time and resources.

Graeme Gill.



Other related posts: