[semi] automatic registration [was hugin beta release]

Mike Runge mike at trozzreaxxion.net
Tue Oct 14 15:12:05 BST 2003


Pablo,

I don't have any experiences or skills with feature recognition and so I'm
maybe not allowed to jump into that discussion?! So I didn't send this mail
to list. But I have a idea (probably not new ;-) ), that maybe is worth of
having a look after:

What do you think about letting the user pick f.e. 8 points in a so called
'reference pair'.
Use this pixel positions of e.g the left picture for all left pictures in all
pairs of the same row.
Use aft (maybe with larger search region) with a initial position taken from
the corresponding point of the right picture of the pair.
Run Optimizer and kick out the 2-4 worst matching points of each pair or all
points, that are with a larger difference than a specific value.

The same could be happen (with different ref pairs) for each row or for
vertical mapping (pic1-row1 to pic1-row2).

The downsides (as far as I can see):
- usable gui required to specify ref-pairs and tell the algo what ref-pair to
use for which pictures. Maybe not easy to understand what to do to get best
results?!
- requires more or less identical moves/tilts between the individual pics.

What I like:
- User can (more or less) control how many points are used and in what
regions they are created.
- required techniques are more or less available.

OK,
I'm not a programmer and I don't know what kind of magig you are able to
do;-)
Maybe this idea is not powerful enough and not the most clever way - it's
just driven by my personal user perspective and by my limited knowledge about
recognition stuff.
This kind of idea would only be usefull for people, that create the
individual pictures carfully following a specific method (same overlap,
etc.). But I don't think that many people do multirows od sphaerical panos
without tripod/nodal point adapter.

best, mike  


More information about the ptX mailing list