[semi] automatic registration [was hugin beta release]

[Pablo d'Angelo]

On Tue, 14 Oct 2003, Edward Wildgoose wrote:

> > I want to add the resulting point into a PT project, so that the expert user
> > can fix a bad automatic stitch, for example there are changes in the
> > overlap area that the automatic algorithm doesn't cover.
> 
> For sure.  Although in practice the user might be overwhelmed by the high number of points.  I guess it is easy to remove 90% of them later though!

Hmm, actually I didn't want to keep thousands of control points for the
whole panorama. Optimizing with them would be quite slow I guess.

> > Correlation of the candidate points (#2) could also be done in the frequency
> > domain.
> 
> At first glance I thought this was a great idea.  But thinking some more,
> I think it buys us nothing.  At least if you use the frequency domain on
> the original then you have more data.  Actually we gain nothing by running
> it on the extracted points because these may be misaligned as well... Oh
> well.

Hmm its the question if a feature based, optimizing algorithm should be used
at all or not.

> > > In fact this is probably an excellent area to start researching since it
> > > seems to be extremely fast.
> 
> > Hmm, we'll see how fast it is. I'm a bit sceptic.
> 
> Erm, Isn't it just one FFT on each image?  I'm a bit hazy on the algorithm mind and I don't have any papers in front of me...

It think its a bit more than that, if rotation is included (fourier-mellin
transform).

> Should be really quick if so.  Much faster than the point method.

Yes, probably faster. The drawback of these methods is that they always use
all the image information, whereas the feature based methods can hopefully
made so that they ignore moving people etc.

> Scanning few a couple of papers it can even be made to do sub-pixel
> accuracy on distorted images...  (Need to consider that certain
> distortions are simple operations in frequency space).

I haven't seen any paper that include the lens distortion estimation in a
frequency based image registration. Can you send that one to me?
Not even a paper how how the lens distortion transformations look in
frequency space. But then again I might have not looked well enought.

> It would be an excellent phd/msc subject to do matching on distorted
> images whilst incorporating the lens correction factors.  Should be
> doable, and would mean that you move everything into the optimisation
> routine.  

Hmm, I already have a msc and phd subject ;)
Acutally, one wouldn't need an optimisation routine with pure frequency
based approach, since all the wanted parameters can be calculated directly
(IIRC). Maybe its needed to incoporate the lens distortion?

> I think this is a red herring for now though.  Point matching is better,
> and inspectable by a human (later).  However, the point is that it may be
> useful to have a rough initial match, hence either try frequency based
> technique, OR (easier) ask the user to drag and drop them in rough
> alignment... (seems a good solution to me...)

acutally a low precision frequency method shouldn't be too hard to
implement, I hope. But currently I don't have time for much of this stuff,
since I'm quite busy with my thesis.

So I'm quite happy with a feature point based algorithm, because then I can
use the pano tools optimizer, which is known to work quite well :)

> P.S. Got any more references on your Harris point detectors?

A Combined Corner and Edge Detector,
http://citeseer.nj.nec.com/ncontextsummary/35939/0
unfortunately the document is not in the citeseer database.

But I'll just use the implementation in the vigra image processing library
I'm already using inside hugin. it also includes an interface to fftw, so
we got the basic tools to start implementing something ;)

PS: cc'ed to the ptx list. maybe somebody with image/signal is lurking
around there. There are at least to astro physicists who probably
learned a lot of math at the university :)

ciao
  Pablo
--
http://wurm.wohnheim.uni-ulm.de/~redman/
Please use PGP