[ptx] Optimizing, Stitching and Seaming

[Yili Zhao]

Hi,
  I have some questions and ideas about pt algorithm.
  I remember that Pablo said that the optimizing is done by the 
Levenberg-Marquardt algorithm. From my math knowledge, when I solve a
nonlinear optimizing problem, I need a objective function to be minimized
or maximized, and need approximate values of the parameters to be optimized 
because nolinear optimizing is sensitive to initial values, otherwise, it will
obtain local minia. That is why we need to set the yaw, roll, pitch and hfov 
before optimizing. If I do not consider lens distortion, what is the objective
function to be solved?
  After optimizing, I get the yaw, roll and pitch for every image. From my math
knowledge, I can construct three rotational matrixs based on yaw, roll and pitch,
then a whole transform matrix can be obtained by multiplying the three matrixs. 
Then I can transform every image in its 3D space. If the project format is sphere,
I can project every image in 3D space to its spherical coordinates.
  Because not only one pixel will be projected to the same coordinates,seaming is need. 
If blending is used, maybe the ghost will appear. If seaming is used, the left part of 
the seam can be copied from the left image, and the right part of the seam can be copied 
from the right image, and the ghost can be eliminated. I find a page http://www.gisdevelopment.net/aars/acrs/1991/psq/ps016a.shtml describes some things on this.
  Comments are welcome, thanks.

Yili