[Insight-users] LBFGS with Centered2DTransform: All points mapped to outside of the moving image

[Luis Ibanez]

Hi Felix,

Thanks for letting us know of your progress.

All ITK optimizers have the same API for setting the
scaling parameters array.

However, the Optimizer parameters themselves
are very unique to every optimizer. Therefore, you
have to familiarize yourself with the optimizer
parameters that control the equivalent of the
"Step length" concept.

This is critical, because it may be responsible from
jumping to a position that is too far in the parametric
space, and therefore results on your image fixed
image not overlapping any longer with the moving
image.


      Regards,


            Luis


-------------------------
On Thu, Oct 8, 2009 at 8:43 AM, Felix Bollenbeck <
Felix.Bollenbeck at iff.fraunhofer.de> wrote:

>
> Dear Luis, dear users
>
> thanks for your hint.
>
> The problems seems to be comming from the scales - the floating image is
> indeed far far away (see below).
>         I have done some reading on the quasi newton methods  maths, but
> completely forgot thinking about parameter scaling, so I had kept the scales
> from the gradientDescent code.
>
>
> Using uniform 1.0 scaling avoids the first error, but then stepsizes are
> 2-3 maginitudes to small laeving the metric value almost unchanged, over the
> iterations.
>
> Does the LBFGS class have a different interface for the scaling parameters
>         and where would I look best for more info on the scale parameters?
>
> Best regards,
>
>         Felix.
>
>
> Initial =
>  Angle (radians)   = 0
>  Angle (degrees)   = 0
>  Center X      = 111.125
>  Center Y      = 116.064
>  Translation X = 11.9944
>  Translation Y = -8.46667
> Number of Pixels=425780
> FixedImage schedule: [1, 1]
>
> MovingImage schedule: [1, 1]
>
> -------------------------------------
> MultiResolution Level : 0
>
> GetValueAndDerivative [0, 111.125, 116.064, 11.9944, -8.46667]
>     // <====================================
>
> *************************************************
>   N=5   NUMBER OF CORRECTIONS=5       INITIAL VALUES F= 15301   GNORM=
> 5.16606e+06
> *************************************************
>    I   NFN    FUNC        GNORM       STEPLENGTH
> GetValueAndDerivative [0.000196306, 111.125, 116.064, 9324.5, 17691.2]
>    // <====================================
> ExceptionObject caught !
>
> itk::ExceptionObject (0x7f9f10)
> // snip
> Description: itk::ERROR: MeanSquaresImageToImageMetric(0x7f2260): All the
> points mapped to outside of the moving image
>
>
>
>  *Luis Ibanez <luis.ibanez at kitware.com>*
> Gesendet von: insight-users-bounces at itk.org
>
> 06.10.2009 20:42
>   An
> Felix.Bollenbeck at iff.fraunhofer.de
>  Kopie
> insight-users at itk.org  Thema
> Re: [Insight-users] LBFGS with Centered2DTransform: All points
>  mapped to outside of the moving image
>
>
>
>
>
> Hi  Felix,
>
> Thanks for the clear description of the problem.
>
> If the Exception is thrown before printing the first
> iteration of the optimizer, it is then likely that the
> problem is in the initialization of the transform.
>
> ...
>
> Printing out the parameters that are received
> by the Metric in its first evaluation will help
> answering the question.
>
> Simply add a std::cout in the GetValueAndDerivative()
> method, of the Metric that you are using.
>
>
>   Please let us know what you find,
>
>
>        Thanks
>
>
>             Luis
>
>
> ---------------------------------------------------------------
> On Wed, Sep 30, 2009 at 9:39 AM,  <Felix.Bollenbeck at iff.fraunhofer.de>
> wrote:
> >
> > Dear ITK user,
> >
> > I want to use the LBFGS for centered 2D registration, instead of gradient
> > descent, hoping for a little speed-up.
> >
> > I exchanged the optimizer in one of the multi-res examples.
> >
> > (The regular step gradient optimization works fine. )
> >
> > The LBFGS version complains though:
> >
> > "All points mapped to outside of the moving image"
> >
> > I assume, I have made two possible errors:
> >
> >
> >         1. Scales are not passed correctly to the LBFGS and back, so that
> > the floating image is indeed way off course
> >
> >
> > and /or
> >
> >         2. Transform and/or Metric do not deliver derivatives as the
> > optimizer expects it, and the registration method assumes that these
> values
> > are zero
> >
> >
> >
> > The exeception is thrown before any optimizer positions are printed.
> >
> > Can you point me in the right direction?
> >
> > Regards,
> >
> > Felix.
> > _____________________________________
> > Powered by www.kitware.com
> >
> > Visit other Kitware open-source projects at
> > http://www.kitware.com/opensource/opensource.html
> >
> > Please keep messages on-topic and check the ITK FAQ at:
> > <http://www.kitware.com/opensource/opensource.html>
> http://www.itk.org/Wiki/ITK_FAQ
> >
> > Follow this link to subscribe/unsubscribe:
> > <http://www.itk.org/Wiki/ITK_FAQ>
> http://www.itk.org/mailman/listinfo/insight-users
> >
> >
> _____________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at
>  <http://www.itk.org/mailman/listinfo/insight-users>
> http://www.kitware.com/opensource/opensource.html
>
> Please keep messages on-topic and check the ITK FAQ at:
> <http://www.kitware.com/opensource/opensource.html>
> http://www.itk.org/Wiki/ITK_FAQ
>
> Follow this link to subscribe/unsubscribe:
>  <http://www.itk.org/Wiki/ITK_FAQ>
> http://www.itk.org/mailman/listinfo/insight-users
>
>
> _____________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at
> http://www.kitware.com/opensource/opensource.html
>
> Please keep messages on-topic and check the ITK FAQ at:
> http://www.itk.org/Wiki/ITK_FAQ
>
> Follow this link to subscribe/unsubscribe:
> http://www.itk.org/mailman/listinfo/insight-users
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20091008/6e9a2b58/attachment.htm>