auto-finetune

[Peter Suetterlin]

> I did a lot of work last weekend using hugin-20031001.tar.gz. I stitched a
> multirow pano with 2 rows (22 photos, 11 photos each) - so it was mainly
> picking points (~ 4-5 points per pair, overall about ~ 140 points).

Me too :-)
Just, my point count is some 400 points by now...
> 
> hugin was stable and auto-finetune works well. I used the defaults for region
> size. I would say auto-finetune(aft) had a quote in finding the correct
> matching point about 70%. In some cases aft used permanently a wrong point -
> placing the cursor different didn't helped in that cases. In 3 or 4 cases aft
> was not able  to find a 'matching' point at all.

Indeed, hugin is extremely usable, and especially aft makes the work
extremely easy.
However, if I understand the way of aft correct, placing the cursor
slightly different in the second image should not really change the
result of the aft process.  Else it is doing something wrong. It
should find the local maximum of the correlation...

> In 60% of the cases, where aft fails in first instance, picking
> sligtly different in a second try works. In the other 40% I had to
> give up matching theses points in 'fit-to-screen-view'.

Hmm, may be that is the reason?  How large is your hugin window?  And
what are the aft settings that you use?

> Picking another point or working in
> 100% view is the woraround. Nearly all the times aft works fine in 100%
> viewsize - but this mode is uncomfortable to work with (scrolling, overview
> of picked point', etc.)
> My pictures have a size of 1536x2048 (portrait).   

Mine too, but I prefer working in 100% mode (therefore my requests for
simultaneous scrolling).  But maybe the zoom-view Pablo is working on
will make that obsolete ;^>

> My idea (beneath make the algorithm better ;-) ) how handle that from a user
> standpoint (weather it's feasable or not) would be the following:

Hmm, I guess indeed improving the algorithm is rather the way to go...

> Ok, my ideas are driven from my work with PTOpenGui ;-) Starting with
> automatically created points would be much smarter (and maybe much more
> interestiong to solve for programmers ;-) ).

For that you would have to have a quite good idea of the actual FOV,
the lens distortion, and the approximate positioning of the subimages.

  Pit