[ptx] autopano: first attempt at semi-automatic panorama stitching

[Ed Wildgoose]

>Whats interesting about it is that is has an easy to understand description of
>a feature indexing algorithm called "Smith"'s edge and corner detection
>algorithm. I got it implemented today and it works vastly better than my old
>one. In the PhD thesis its compared to Harris's algorithm (which is what most
>more modern algorithms base their idea on) and while it is quite good in
>general, Harris's is more stable against slight image movements (which is what
>I need for the panorama thing :)
>
>  
>
Harris corners seems to be the basis for all modern feature detecting.

I wonder though if something like the following algorithm wouldn't help:

Find all the feature points on an image.  Apply a standard clustering 
algorithm to group them into smallish clusters.  This now means that we 
have effectively subdivided the image into seperate processing areas.

Matthew uses a more advanced technique for blanking out large chunks of 
each image where there doesn't appear to be much happening.  I believe 
that this is the key to allowing faster image correlation.  You are 
literally left with only islands of detail and this presumably vastly 
reduces the amount of image to cross correlation and probably also 
reduces the number of accidental false registrations.

Now for each cluster of points attempt to correlate.  Assuming that we 
know that the pictures are rotated the correct way around then  we can 
perhaps apply knowledge of the lens params and assumed image positioning 
to warp the images appropriately and bring differences in rotation, 
distortion to a much smaller measure

Now attempt to cross correlate the clusters of points using both image 
correlation based on surrounding pixels, and also knowledge of relative 
positions of each interest point  to confirm matching.  Matthew's paper 
mentions looking for a first good match and then building on that to 
help reduce the search space (at least that is how I read it?).  Anyway, 
I say the last bit quickly as though it were straightforward, although 
of course this is the meat...

Perhaps this last step could make use of the hugin registration point 
optimiser.  This seems to be pretty robust on the images I tried it on, 
despite small changes in rotation (what are you using anyway?  Just 
standard image correlation?).  So do something like this:

Take a cluster of points.  Pick one of the inner points and cross 
correlate with a random cluster.  If you find a strong single 
correlation then proceed to add extra points from each cluster and cross 
correlate either rejecting when there is multiple possible correlations, 
or else using the points so far to work out rotation, distortion and use 
this to decide which correlation is most likely....  Hmm, complicated 
optimisation if we are not careful...

Anyway, the key thing from Matthew's paper seemed to be to try and find 
islands of interest points and correlate those rather than the whole image?

>I also found an interesting paper today that might be an alternative to the
>[iccv2003] panorama paper you posted me:
>http://citeseer.nj.nec.com/mikolajczyk02affine.html
>  
>
This does actually look really neat