[Insight-users] Registration: X translation is always wrong

[itkvtk123 at gmx.net]

Here you find an example of the problem.
I'm trying to register a clinical image with a rotated and translated version of itself.
See

http://img411.imageshack.us/img411/632/18632413.png

The real offset is

X = -15
Y = 30
R = -24 (degrees)

My multiresolution registration finds the parameters

X = -1.68 (wrong)
Y = 30.13 (correct)
R = -23.99 (correct)

So all parameters are perfect except X and I don't understand why X is always wrong.
I'm using Gradient Descent with the MeanSquares Metric.

I plotted the way the optimizer takes through parameter space (both images same plot with different perspective):

http://img38.imageshack.us/img38/5980/53532608.png
http://img196.imageshack.us/img196/5332/22686740.png

Red is the path the optimizer takes and the green cross indicates the real value it should find.

As you can see it directly goes to the correct Y and R values, but in matters of X, it takes the wrong direction (see images: the optimizer goes to the right handside but it should go to the left handside (to real X value), if you know what I mean).

Any comment appreciated.

Merci
-- 
Neu: GMX Doppel-FLAT mit Internet-Flatrate + Telefon-Flatrate
für nur 19,99 Euro/mtl.!* http://portal.gmx.net/de/go/dsl02