[PanoTools] [semi] automatic registration [was hugin beta release]

[Edward Wildgoose]

I've lost my most interesting paper, but yes, exactly what you described is the algorithm.  Harris detectors rings a bell, but perhaps Deriche edge detectors are worth reading up on?  Re: Step 2, I think (intuitively) that accuracy would be greatly improved by a rough match using a frequency based approach.  

In fact this is probably an excellent area to start researching since it seems to be extremely fast.

I think though that with wide angle lenses it may be necessary to do a kind of iterative approach, ie assume some initial alignment (1/3 overlap and horizontally aligned for example), then distort the image as though this were correct, THEN run an initial algorithm to find alignment.  Otherwise I suspect the distortion will be far too large.

OK, so now all we need is lots of free time to code it...

Ed W

-----Original Message-----
From: Pablo d'Angelo [mailto:pablo at mathematik.uni-ulm.de]
Sent: 14 October 2003 07:42
To: PanoTools at yahoogroups.com; PTX mailing list
Subject: [PanoTools] [semi] automatic registration [was hugin beta
release]


On Mon, 13 Oct 2003, Edward Wildgoose wrote:

> > Edward Wildgoos,
> > are you still interested in implementing an automatic registration algo for
> > the pano tools suite?
> 
> Still interested, but I stopped looking for interesting papers on the idea.
> 
> I think the algorithm is fairly clear though (but not trivial).  Basically do rough image matching using a frequency based alignment method (I'm assuming it will be accurate enough given the distortions involved mind...).  Then several papers have referenced using a certain filter (forget the name) to locate hundreds of features in each image.  Match and optimise these between images (pruning the high error matches since these are likely mis registrations between each image).
> 
> I think this is the "right" way to do it, but certainly not trivial...

Well, thats life ;)
My current plan is the the following (its part of a algorithm you
described above) that can be used in concert with the panotools:

1. harris corner detection in an image pair.
   We could probably use/offer other point feature detectors as well.
   Do you remember which feature detection algorithm the other papers
   used?

2. take the corners with the highest score from each pair,
   use correlation to find possible pairs.
   Before using correlation, other constraints could be applied to cut
   down the number or possible corrospondences. This is where a rough
   estimation of the images would be handy.

3. do a local optimisation (using pano tools optimizer)
   
4. remove points with large distance. Maybe I could also take
   one of the next best points 2.) that matches well, and fits into
   the current transform.

This obviously depends how good step 1 and 2 work. to many false
candiates and everything will fail, because the optimisation error is
not a measurement for a bad corrospondence then.

As mentioned in step 2, a hint from of the relative orientation would be
nice to have. I'm not sure if frequency domain algorithms exist that
work on non rectilinear images. Actually, if I understood them
correctly, they also estimage a affine, not a perspective transform.


Well, all this won't be implemented in the next week or so, but I'll
work on it from time to time.

Have to fix bugs and add features to the GUI as well :)

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

------------------------ Yahoo! Groups Sponsor ---------------------~-->
Rent DVDs from home.
Over 14,500 titles. Free Shipping
& No Late Fees. Try Netflix for FREE!
http://us.click.yahoo.com/ArdFIC/hP.FAA/3jkFAA/.Cr1lB/TM
---------------------------------------------------------------------~->

To unsubscribe from this group, send an email to:
PanoTools-unsubscribe at yahoogroups.com

 

Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/