[ptx] new reference for stitching algorithms
    Aseem Agarwala 
    aseem at cs.washington.edu
       
    Mon May  9 20:17:22 BST 2005
    
    
  
Hi everyone, thanks for the interest.  I'll reply to several 
people/topics at once.
Pablo:
I will release the command-line version in 1-2 days.  As to my future 
work, I'm working in the same domain, but have moved onto other 
problems.  My latest work involves creating panoramic video from a 
single, panning video camera:
http://grail.cs.washington.edu/projects/panovidtex/
Yes. I used the wxWindows image loading library, which ignores the 4th 
channel.  My command-line tool doesn't use wxWindows.
Smartblend:
 From my guesses as to what I've read on the web, smartblend uses a 
dynamic programming approach to finding seams, similar to this paper:
http://graphics.stanford.edu/~jedavis/panorama/
However, it's not clear how to handle regions with more than two 
overlapping images using dynamic programming, so I believe graph cuts 
are more general.
Ippei:
I've never tried to compile on a mac, and don't have one around.  As far 
as the graph cut code that isn't GPL, I'm afraid I can't easily resolve 
this issue, since the code isn't mine.  However, that code has been on 
the web a long time, and likely will remain forever.
Speed:
My code has two parts.  The first involves graph cuts to find optimal 
seams.  The second is a gradient-domain blend.
The graph cut part is currently quite slow on many, large images. 
However, I've recently implemented a hierarchical graph cut algorithm 
that would help.  I suggest people try my stuff (once I release the 
command-line version); if there is enough demand & interest, I can make 
it fast.
The gradient-domain part is actually already fast.  Andrew, I use sparse 
matrix techniques so that the memory allocated is always on the order of 
the number of pixels.  Also, I use a hierarchical preconditioner that 
causes the algorithm to converge pretty quickly.
Thanks for all the interest!
Aseem
-- 
Aseem Agarwala
http://agarwala.org
    
    
More information about the ptX
mailing list