Browse argyllcms: This Month's Archive • Main Archive Page • Related posts • Previous by Date • Next by Date
[argyllcms] Re: scanin and perspective distortion

From: Pablo dAngelo <Pablo.dAngelo@xxxxxx>
To: argyllcms@xxxxxxxxxxxxx
Date: Wed, 06 Feb 2008 11:01:57 +0100
Gerhard Fuernkranz wrote:
> Pablo d'Angelo wrote:
> > Gerhard Fuernkranz wrote:
> >> I think you should not overestimate small errors introduced by
> >> correcting distortions with an image editing program, particularly if
> >> you do it with 16 bits per channel, in a linear light space (raw RGB
> >> with gamma 1.0). When capturing a target with a camera, there are IMO
> >> other typical sources of error with a significantly larger contribution
> >> to the total error (for instance uneven illumination, natural and
> >> optical vignetting induced by the lens, still some residual glare).
> >
> > Regarding that, how sensitive is argyll with respect to vignetting.
> > During adding vignetting correction to my photo stitcher (details at
> > http://hugin.sf.net/tech/ ), I noticed that most lenses/cameras have a
> > light falloff of about 20-50% in the corners. Actually, since most
> > light falloff is radial from the image center, it might be possible to
> > add simple vignetting model (for example polynomial) to the profiling
> > process.
> 
> Even if we have such a model, I'm not sure, how the model parameters
> should be determined. I guess most people would not want to specify the
> geometrical details of the capturing setup and additionally a couple of
> lens-specific parameters, which one likely doesn't know either in
> advance for the particular camera/lens. So it would be necessary IMO, to
> estimate the model parameters automatically - for instance from a
> captured flat field image (but then we arrive at the alternative below
> anyway).

I agree that an automatic estimation of the parameters would be
required for this process. However, since the "relative brightness"
of the target patches is already known, I believe the vignetting
parameters could be determined fully automatically during profiling.
This could also have the advantage that, depending on the type of
profile estimated, non-linear channel responses could probably be
handled as well.

> > Alternatively one could divide the image by a flatfield image, this
> > would also account for uneven lighting, and can be used to correct the
> > image in a preprocessing step. It will only work for "linear" (with
> > gamma 1.0) images, though.
> 
> The flatfield image and the target would need to be reasonably well
> registered, of course. In order to account for uneven lighting it would
> be necessary that both, the target and the flatfield image are perfectly
> flat, or if not, that they have the same curvature shape. The amount of
> luminance error induced by tilting the target plane by a different angle
> at different regions of the target eventually depends on the exact
> lighting setup; using not a single, but several light sources may help
> to reduce this kind of error.

Good point, I forgot about the curvature. In a practical application I would
try to place a uniformly white paper (if such a thing exists...) over the target
(maybe the back of the target might do as well, if it is perfectly flat). The
camera should of course be placed on a tripod, so that the angles between
light sources, target surface and camera stay the same.

Anyway, there are probably other traps to fall into (different reflectance 
function
of color and flatfield target etc.), so nicely parallel lighting and planar 
targets are
probably the best environment.

Additionally, the lprof documentation has some more hints about a good setup.
http://lprof.sourceforge.net/help/ufraw.html

> Some targets have several white (or nearly white) patches distributed
> across the area. One could also fit a curved surface through these
> points in order to estimate the flat field, instead of capturing a flat
> field separately.

Indeed.  As mentioned above, couldn't the other patches be used for
that as well? I guess the question is whether this can lead to ambiguities
in the estimation process. I have to admit that I don't know the exact
error function that is minimized during profiling.

> It should work for arbitrary gamma, as long as the channel response is a
> true power function, and not an arbitrarily shaped curve (note that
> a^gamma * b^gamma = (a * b)^gamma). If this is not granted, I could
> imagine that an iterative process might work, which scales the luminance
> not in RGB space, but in XYZ, after applying the profile created in the
> previous iteration.

My experience is that the cameras I own do not have a true power function as
channel response (when outputting to jpeg). The camera manifacturers (and most 
raw
converters, too) do not use true power function channel response functions 
depending
on the user settings etc. Unfortunately it is often not mentioned in the 
manuals how
different settings for contrast modify the channel response.

ciao
  Pablo
_________________________________________________________________________
In 5 Schritten zur eigenen Homepage. Jetzt Domain sichern und gestalten! 
Nur 3,99 EUR/Monat! http://www.maildomain.web.de/?mc=021114
Follow-Ups:

[argyllcms] Re: scanin and perspective distortion
From: "Gerhard Fürnkranz"
[argyllcms] Re: scanin and perspective distortion

Other related posts:
