[Insight-users] Question about MI

[Henning Meyer]

You can use optimizer->SetScales( ... ); to adjust the optimizer
behaviour. Have a look at chapter 8.6.1


Henning

2006/6/30, Jonathan Wong <jon.the.wong at gmail.com>:
> Well I'm actually using Versor3dRigidTransform and optimizer because the
> image is actually 3D. However, mainly the misallignment comes from rotation
> and translation around the Z, with very small (negligible) rotations in the
> other directions.
>
> I am using the optimizer, but the problem arises, when rotation along the Z
> is really different from the fixed image. Say 110 or 165 degrees. In my code
> so far, Mattes is much more stable than Viola, but I'm trying to figure out
> a set of parameters to use to obtain the ideal metric automatically by
> helping the Optimizer by pre-intializing the optimizer with angles close to
> the actual angle.
>
> Any other suggestions or improvements on how to do this?
>
>
> Jon
>
> On 6/28/06, Henning Meyer <tutmann at gmail.com > wrote:
> > Even if Mattes MI just takes a random subset it should be a stable
> > result, which does not change to much when running it over an over
> > again with the same initial parameters.
> > Furthermore it should be quite continious, which I guess you asked for.
> > But if it is just translation and rotation in 2D then why not take an
> > optimizer - it could to it automatically quite fast.
> >
> > Henning
> >
> > 2006/6/28, Jonathan Wong <jon.the.wong at gmail.com>:
> > > Thanks, Henning. Actually, I want to try to semi-automatically obtain
> the
> > > registration transform, since I know there is only rotation and
> translation
> > > along the z-axis. I don't have a problem with translation, but it seems
> that
> > > depending on how far I intialize from the true angle of rotation, it
> will
> > > converge to different angle values, which may or may not be correct.
> > >
> > > Therefore, I was going to try a series of values, in a specific region,
> and
> > > find the lowest metric and use that resulting angle as the actual
> transform.
> > > However, since Mattes MI Metric supposedly takes a random subset of
> samples
> > > the metric should be somewhat different each time it's run during a
> process
> > > (if it runs in a loop).I was wondering if the metric values from running
> > > registration overagain with different initial values would have metrics
> that
> > > would be relative to one another, so I can effectively use the lowest
> metric
> > > to essentiall find the global minimum.
> > >
> > > Hopefuly I was more clear than previously.
> > >
> > > Jon
> > >
> > >
> > > On 6/28/06, Henning Meyer <tutmann at gmail.com> wrote:
> > > > If I got you right you won't need an optimizer.
> > > > You just want to have the metric value of you transforms (given in
> > > angles).
> > > > You could just create an instance of your metric and you transform and
> > > > then retrieve the metrics value.
> > > >
> > > > Henning
> > > >
> > > > 2006/6/27, Jonathan Wong <jon.the.wong at gmail.com >:
> > > > > Since MI takes a random sub set of samples from the images. Each
> time i
> > > run
> > > > > MI, it could give me a different value.
> > > > >
> > > > > Currently I want to try to intialize different angles to the MI
> > > registration
> > > > > method and compare their metrics to see which is indeed the correct
> > > > > registration?
> > > > >
> > > > > Would I have to do something with the optimizer, so that the metrics
> > > will be
> > > > > a relative value to one another, or can i simply instantiate a new
> > > > > optimizer?
> > > > >
> > > > > Thanks
> > > > > Jon
> > > > >
> > > > > _______________________________________________
> > > > > Insight-users mailing list
> > > > > Insight-users at itk.org
> > > > > http://www.itk.org/mailman/listinfo/insight-users
> > > > >
> > > > >
> > > > >
> > > >
> > >
> > >
> >
>
>