[ptx] HFOV, focal distance...

[Terje Mathisen]

ptx-bounces at email-lists.org wrote:

> On Fri 03-Jun-2005 at 15:06 +0200, Terje Mathisen wrote:
> 
> [a, b & c parameters]
> 
>> I measured it by printing a really big rectangular spreadsheet on an
>> A0+ printer, i.e. the printout is about 1.4 x 1.0 m.
>>
>> I then took a photo of this printout, after placing the camera on a
>> tripod and using a small mirror to make sure that the camera was
>> pointing at the actual center of the image.
> 
> 
> This isn't strictly necessary, the rectangular grid object can be at any
> angle to the camera and you will still get accurate results.

Yeah, I studied the math (did some work on making an automated
intersection recognizer that could automatically generate control pairs
for all possible combinations), but realized that it was easier to
verify good results when I knew that yaw should be zero (actually, I
believe I did lock this down) and roll & pitch should be close to zero.

The really hard part is to get good results close to the corners, but
otoh the really important area is the 50-70% distance from center range,
since this is where most of the (central) stitch areas will end up!
> 
>> On the single image I placed 100+ horizontal and vertical line control
>> points, then optimized everything (y,p,r,a,b,c) (remembering to select
>> rectiliear output!) except HFOV.
> 
> 
> This technique can actually determine the horizontal field of view at
> the same time as everything else, though if you want a really accurate
> value you do need to assemble a full 360 degree panorama.

It is always easier when the number of free variables is less, so I also
did multiple rounds with and without locking down roll & pitch.

It is amazing how tight the grid needs to be to get good all-over
results when the real distortion doesn't really match a third-order poly
too well. :-(

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