[ITK-dev] [ITK] 2D/3D registration in v4

[Matt McCormick]

Hi David,

On Thu, Nov 12, 2015 at 9:51 PM, David Burns <david.mo.burns at gmail.com> wrote:
> I think this issue relates to the way the metrics compute gradients. It can
> be reproduced using
>
> Examples/RegistrationITKv4/ImageRegistration1.cxx
> and
> Examples/RegistrationITKv3/ImageRegistration1.cxx
>
> by setting Dimensions to 3, and using 2D images as inputs.
>
> I found using: metric->ComputeGradientOff()
> will allow the registration to proceed for ITKv3 using the
> MeanSquaresImageToImageMetric, but there is no such function for the v4
> version.
> Using a metric that does not require gradients (eg Mattes) will allow the
> registration to proceed in both v3 and v4.
>
> I humbly suggest the registration framework and metrics should be able to
> handle this situation. Also of note to those interested, the v3 registration
> on my system goes about twice as fast per iteration.

Patches would be welcome to improve the situation! Details on patch
submission are here:

  http://www.itk.org/Wiki/ITK/Git/Develop


In terms of performance, we are creating a system of benchmarks so
that we can quantify performance to make improvements and identify
regressions.  The work is being added to this repository:

  https://github.com/InsightSoftwareConsortium/ITKPerformanceBenchmarks

The basic infrastructure will be in place in a few weeks time, after
which you create a pull request for your use case. This will help make
it possible to improve performance on your use case.

Thanks,
Matt