Re: aligning stacks and combining them?
JPW <johnpwatkins <at> gmail.com>
2011-10-01 18:00:00 GMT
On Sep 30, 1:59 am, kfj <_... <at> yahoo.com> wrote:
. . .
> Hugin started out as a GUI from which to operate a bunch of command
> line tools. A lot of this heritage is still apparent, and keeping the
> components separate also has certain advantages, i.e. each can evolve
> by itself and be used individually and by other projects as well. The
> helper programs, like nona for 'warping' and enblend/enfuse for
> combining the warped images, ar CLI-driven. Since for most purposes
> it's perfectly sufficient to use them with their default settings,
> probably noone has made an effort to present their interface
> graphically. Only when people try and do things out of the ordinary,
> additional parameters become necessary. Trying to keep a GUI
> consistent with the moving targets (--exposure-mu and --exposure-
> sigma, for example, are quite recent. if I remember correctly) takes
> someone to do it - hugin has rather slim resources, and there are a
> lot of more pressing needs. You'll just have to bear with this status
> quo for now.
I don't mean to sound like I'm complaining. Once I understand how it
works I can probably learn to use the CLI.
> The difference between any of your large figure trials should be
> minimal, but there should be a measurable (if not immediately
> noticable) difference to the default of .25 - since you've had a look
> at the gaussian curves in the manual, imagine that doubling the sigma
> widens the curve to double breadth, so your lowest try at 5 is already
> 20 times as wide as the default, and a further increase to 50 does not
> make much of a difference, just as incresing it further to 500
> wouldn't either.
Makes sense.
> It does depend on the level of noise and the peak intensities. If
> you've one image withe a bright pixel where the correcponding pixels
> are dark, this pixel is closest to well-exposedness and would be
> weighted as 'best' - not what you want, that's why you raised the
> sigma to 'cloud' enfuse's judgement. The more images you have, the
> less obvious the effect of one brighter pixel becomes, since all
> images are still included in the result, even though the less well-
> exposed are included less. Using enfuse with the default settings
> does, as you observed, help some, but the increased sigma should make
> the effect better. This may be hard to appreciate with the naked eye.
> Try this: Load the result from the default settings and that with
> increased sigma as layers into an image processor. Set the layer mode
> to 'difference' and then flatten the image. It will look pretty much
> black at first glance now, and this demonstrates how small the
> difference is - it's hard to see with your naked eye. Next, set the
> levels so as to have the white point quite low down (lower quarter or
> less, just play with it) to brighten the difference image. You'll now
> see the difference between the two versions magnified, and I'm quite
> confident the differnce will look quite noisy and show how the added
> parameter worked.
Actually after I answered you I took a closer look and could see the
difference. It was on a (physically) smaller scale, and more subtle
than the jpeg noise. Also, just looking at the histograms of the
difference of the images, the gains do indeed slow quickly as sigma
increases. Most interesting to me is that you see a similar
diminishing increase in benefit (in terms of noise) as the number of
fused images go up. 2 images are OK. 3 is a good minimum for good
results, and in fact with default sigma, is far superior to just two
with a sigma of 50. 3-5 seems optimum. More than five appears to not
be worth the effort (for a typical application, I suppose.)
> You also have to take into account that warping the images (like,
> shift them all so that they precisely overlap) will already reduce the
> noise a bit, since the pixels in the warped images will already be put
> together from several original pixels by the interpolation stage. And
> to appreciate the denoising effect fully, you should have the output
> at 'optimal size'. If you really want to have a closer look at problem
> areas, set the output crop to that area (in the openGL preview) and
> select a larger-than-optimal output size in the stitcher tab. It'll
> work like a magnifying glass.
Good idea.
> And while going on about the exposure sigma all the time, it is
> equally important that you should set the saturation weight to zero,
> as you have done (and as I managed to convey to you without any
> typos 
- often the noise isn't only of varying brightness, but also
> of varying coulour. Having enfuse judge a pixel best just because
> noise made it more saturated isn't what you want either, and the
> default settings use the saturation as quality criterion as well,
> though less so than the exposure.
Makes sense.
> Did you follow up the links from the Wikipedia article I recommended?
Yes but not completely.
> I think that
>
> http://research.edm.uhasselt.be/~tmertens/papers/exposure_fusion_redu...
>
> is really an excellent explanation of the process.
Agreed. That really helped. Pictures, diagrams, examples, etc. I need
to read about gaussian pyramids. The comparison with other techniques,
and the comparison of changes in contrast, saturation, and exposure
weighting was especially interesting.
> Once you grasp the
> concepts, it all falls into place. And there are a few GUIs for enfuse
> as well - try and browse for 'enfuse GUI' - with which you can more
> playfully explore the parameters.
>
> Kay
Thanks and have a good trip.
--
--
You received this message because you are subscribed to the Google Groups "Hugin and other free panoramic
software" group.
A list of frequently asked questions is available at: http://wiki.panotools.org/Hugin_FAQ
To post to this group, send email to hugin-ptx <at> googlegroups.com
To unsubscribe from this group, send email to hugin-ptx+unsubscribe <at> googlegroups.com
For more options, visit this group at http://groups.google.com/group/hugin-ptx
RSS Feed