[ptx] Re: how to improve HFOV optimization?

[Littlefields - Rik, Janis, Kyle & Peter]

Hi Pablo,

On Fri Apr 30 09:00:26 BST 2004 Pablo d'Angelo wrote:

 > I've one question. you scale by avgfov, I assume avgfov is the average of
 > the two fov's involved in the control point, right? How does it behave if
 > the hfov's are quite different? Shouldn't the fov of the complete pano be
 > used instead (but if I remember correctly, that is not easily 
available in
 > the optimisation function)?

The avgfov that I use is the average over all images.  I modified
SetAlignParams() to calculate that and leave it lying around in
a global variable.  The optimization function fcnPano picks up
the average fov after it calls SetAlignParams to store parameter
values provided by the optimizer.

The way that I use this is to record the initialAvgFov calculated
from the starting parameter values.  Then on every evaluation
of fcnPano I pick up a new currentAvgFov and scale all of the errors
by initialAvgFov / currentAvgFov;

I am guessing that your concern about different hfov's is that
scaling might change the way that the errors are weighted
between images.  I think this is OK because I use the same
weight for all errors.

--Rik

 > Hi Rik,
 >
 > On Thu, 29 Apr 2004, Littlefields - Rik, Janis, Kyle & Peter wrote:
 >
 > > Pablo,
 > >
 > > I plugged in the fov penalty that I suggested yesterday
 > > >    ...  scale the errors by 1.0/avgfov
 > > > to correct for the shrinking image size as fov gets smaller.
 > >
 > > It seems to work fine.
 > >
 > > Yesterday I gave this example:
 > > > ...single-row series ... with a [nominal] 105mm lens (fov 12.3).
 > > > It optimizes to average error (as reported by PTGui)
 > > > of 0.446 pixels.  When I force fov 10 and reoptimize,
 > > > the average error drops to 0.445 pixels.  At fov 5, the
 > > > error is 0.323 pixels.  It should be apparent why the
 > > > optimizer wants to push fov to zero in this case.
 > >
 > > With the optimizer mod installed, this panorama now
 > > optimizes to focal length 117mm (fov 11.1)  even if
 > > I start from geometries optimized for focal lengths
 > > as far wrong as 20mm (fov 59.1) and 1000mm (fov 1.3).
 > >
 > > Does that sound like what you wanted?
 >
 > Yes, pretty much. Even better, since I would have tried to cure the 
symtoms
 > by adding an additional penalty term, instead of modifiying the old one.
 >
 > I think this should be tested and then applied to the panotools
 > project on sourceforge.
 >
 > I've one question. you scale by avgfov, I assume avgfov is the average of
 > the two fov's involved in the control point, right? How does it behave if
 > the hfov's are quite different? Shouldn't the fov of the complete pano be
 > used instead (but if I remember correctly, that is not easily 
available in
 > the optimisation function)?
 >
 > Sorry, I'm quite busy these days so I can't comment on everything 
happening
 > here.
 >
 > ciao
 >   Pablo