proj-imim:
Parallel displacement panorama assembly
Philo 2001-May-04 21:46:25
I went on some tries with multirow (3x3) microscope images stitching, with
50% overlap and quite a lot of barrel distortion. Following previous
discussion (sorry, don't have my archives here), my first optimization try
used the following setting :
Rectilinear lens, 30° fov, pitch=yaw=0, initial d and e shift parameters set
to +/- half the dimension of images (50% overlap)
First optimisation of roll and d/e of all images except central one, with
lens distortion parameters a=b=c=0. Correct convergence giving correct
placement of images, but of course bad stitching because of barrel
distortion.
Second optimisation of a, b and c: rapid convergence to complete nonsense,
all images are reduced to small circle when assembled. The results are very
similar if I optimize r,d,e and a,b,c simultaneously.
My second strategy was to use small fov values (so that spheric and
rectilinear projection are locally similar):
Rectilinear lens, 1° fov, d=e=0, initial yaw and pitch angles set to +/-
0.5° (50% overlap)
First optimisation of roll, yaw and pitch of all images except central one,
with lens distortion parameters a=b=c=0. Correct convergence giving correct
placement of images.
Second optimisation of a, b and c: rapid convergence to values that
compensate distortion nicely. Final stitching is OK
It took me some time to figure out the cause of the first strategy
failure... it is simply that the d/e shift were created to compensate
optical center offset, and distortion correction is performed centered. So
with huge d/e offset used here, correction was very asymetrical and didn't
at all corrected barrel distortion.
Maybe this can be helpful to someone...
Philo
http://www.philohome.com