[ptx] automatch - some progress picking control points

[Pablo d'Angelo]

On Wed, 14 Jan 2004, Bruno Postle wrote:

> On Wed 14-Jan-2004 at 07:16:30AM +0100, Pablo d'Angelo wrote:
> > 
> > I have started to write a c++ version of it.
> 
> Cool, I bet it's faster too (though I have the perl version down to
> about 5 seconds for each pair of images).

I haven't timed it yet. but Its really fast, even with several hundred
points. Most of the time is wasted by printing dots to the screen ;)

> > Unfortunately I have some bugs somewhere, so I haven't checked it in yet.
> > It creates lots of bogus points.
> 
> For any pair of images, there are always many hundred geometrically
> equivalent triangles of control points simply by chance.

Jep, I've seen that. Actually, I thought about network approach as well, but
I don't have much time to spare right now. Another idea would be to select
only triangles that lead to a consistent relationship (transform) between
the images. One could for example create a histogram of the relative
rotation of the triangle candidates (since we're avoiding symmetric
triangles), and only use the triangles from the highest bin.

Additionally, one could try to exploit feature descriptors (if available).

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