[ptx] possible optimizer bug, filename ordering [Re: bugs - hugin_snapshot_hugin_2004_07_16-08_33_win32.zip]

[Rik Littlefield]

Pablo,

Thanks for the test cases.  As we discussed offline, these tests 
illustrate that b=0 prevents optimization of b, but they do not 
illustrate the "goes berserk" problem;  b is optimized correctly if it 
is optimized at all.  My build of the optimizer handles these tests the 
same as yours. 

It seems that b=0 prevents optimization in some cases but not in 
others.  I don't know exactly why that is.  Maybe something to do with 
relative scaling of various parameters in the depths of the 
Levenberg-Marquardt algorithm.  Anyway, the safest approach seems to be 
always make a/b/c non-zero if you want to optimize them.

Let me know if/when you want me to look at a test case where the 
optimizer goes crazy.  I suspect that it will turn out to be related to 
control point placement, so I will need access to the images too.

Thanks much,
--Rik

Pablo d'Angelo wrote:

>On Fri, 16 Jul 2004, Rik Littlefield wrote:
>
>  
>
>>>Ups, overlooked that. I'll take a look at it.. Strangely, the same thing
>>>happens if I just use PTOptimizer on the script. If I set the initial b to
>>>some number significantly different from 0 (0.1 or so), it works, but else
>>>it fails on the command line as well. If I set it to a really small number,
>>>like 0.0001, the b estimate goes berzerk and I get a rms error of ~144,
>>>where
>>>
>>>I don't remember the old optimizer did stuff like that, I'm using the new,
>>>panotools 2.7a-2 rpm from bruno's site. Maybe its a bug somewhere inside
>>>the improved optimisation?
>>>
>>>
>>>      
>>>
>>Could always be a bug.  But I just reran one of my standard test cases, 
>>using pano12.dll built from 2.7.0.3 sources, under PTGui, optimizing 
>>fov/a/b/c from starting a/b/c = 0, e-6, e-5, e-4, e-3, e-2, 0.1, 0.2, 
>>0.5, and it converged stably in all those cases. It finally failed to 
>>converge well starting at a/b/c = 1, no great surprise. 
>>
>>Could also be some subtle interaction with other parameters.
>>
>>If this problem persists, send me the PTOptimizer script and I'll try it 
>>in my environment.
>>    
>>
>
>Here are some simple tests, I have made, Its just a 2 image pano with some
>11 control points, maybe the control points are not set nicely for a,b,c
>value estimation.
>
>ciao
>  Pablo
>  
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.email-lists.org/pipermail/ptx/attachments/20040718/6666e642/attachment.htm