[ptx] new optimizer: things that might appear to go wrong

[Terje Mathisen]

ptx-bounces at email-lists.org wrote:

> A couple of people have reported cases where the new optimizer makes big 
> changes to fov and a/b/c, even taking fov to large values > 180 degrees.
> 
> The basic issue seems to be that the new optimizer is sometimes able to 
> find "better" values for parameters than the old one was, even when 
> those parameters lie far from the initial estimates.  "Better" means 
> that the error is reduced for the specified control points.  You might 
> think that this is always good, but sometimes it results in 
> "over-fitting" the data so that error is greatly increased in areas 
> where control points were _not_ placed.

This pretty much have to happen when the total number of parameters to 
optimize (y,p,r for each image, plus any lens parameters) is in the same 
ballpark as the minimum (or at least average) number of control points 
between each pair of images.
> 
> I want to reinforce advice that has been posted in other descriptions of 
> how to use the optimizer:
> 
> 1. Do not optimize a/b/c independently for each image, unless you really 
> have used a different lens for each image.
> 2. It is safer to optimize pitch/roll/yaw first, then optimize lens 
> parameters.

IMHO, by far the safest way to get good results is to start by spending 
a few hours generating near-optimal (a,b,c) values for your lens, and 
then leave those numbers fixed afterwards.

Terje

-- 
- <Terje.Mathisen at hda.hydro.com>
"almost all programming can be viewed as an exercise in caching"