[MINC-users] more minctracc questions!

[Andrew Janke]

> But, my initial understanding was that these weights scale the axis in
> the search space

Correct.  Well in effect they set the amount by which the algorithm
will "step" in each axis to begin with. As the algoritm converges,
these step sizes will get smaller.

> after doing some experiments with different param values, I thought I
> have to set the rotation weight higher and the translation weight much
> lower in case I have embryos in completely different orientations (in
> this way the algorithm searches for a transformation with more
> rotation and not that much translation!!) again! does it make ANY
> sense?! ...

Yes, perfect sense, the only problem that you will strike is that a
Euler angle decomposition of a rotation space (what minctracc uses) is
not smooth for large rotations (above 45deg or so) so you may have
trouble getting it to converge. There is some code in there for a
quaternion decomposition of rotation space but it is largely untested.
A quaternion decomposition of a rotation space is smooth for large R.

> Also, do you know if the optimization method used optimized one
> parameter (tx, ty, tz, rx, ry, rz) at a time?

they are all optimised simultaneously. Although of course at some
level in the code, they all occur at a finite level...

My suggestion if you have large initial rotations would be to stary
minctracc with smaller rotation weights and instead have a number of
starting transformations that involve various combinations of the
large rotations that you are getting as per FLIRT in FSL.


--
Andrew Janke   (a.janke at gmail.com || http://a.janke.googlepages.com/)
Canberra->Australia    +61 (402) 700 883