[MINC-users] minctracc - different results on 32 vs 64 bit
Claude LEPAGE
claude at bic.mni.mcgill.ca
Tue Dec 9 15:14:54 EST 2008
Hi Vlad,
> we came a cross an interesting behavior of minctracc - it produces
> slightly different results depending on the architecture.
>
> I run mritotal on 9 subjects and the differences between 32 and 64 bit
> version runs are up to 0.3 degree in determining rotations, 0.1 mm in
> terms of shifts and up to 0.01 in scale.
I know this. I once told Louis about it. It is due to rounding errors
when determining the degrees of freedom. If you run minctracc with -debug,
you will see that the lattice size in x, y, z may be off by 1 in 32 vs 64
bits, thus causing different number of sampling points and registration.
Of course, you can go and dig into the source code if you want to fix it,
but to be fair, I don't know which one of the two results, 32 or 64 bits,
is the correct one.
Claude
More information about the MINC-users
mailing list