[Insight-users] Multiresolution registration slower than regular one

[Luis Ibanez]

Hi itkvtk123,

The statement that

   "multi-resolution registration is faster than single-resolution one"

is certainly a simplistic one.   :-)


The expectation is that because the multi-resolution approach
solves the mis-alignment in a progressive way, it should take
less computation than the single-resolution, where many small
step will be required.

In practice however, the devil is in the details.

It all depends on what type of parameters you set in both:

 a) The single-resolution registration, and
 b) Each one of the levels of the multiple-resolution registration.


For example, a common mistake is to start a second or
third resolution level with optimizer jumps that are so large
that they will make you lose the initialization provided by the
previous resolution level.

Another common mistake, for example, is to use the optimizer,
and setup a number of iterations that is too large along with
convergence parameters that are too tolerant and will not
really stop the optimizer before it reaches the maximum
number of iterations alloted for it. In this conditions the
optimizer will be running unnecessary iterations that do not
contribute to the refinement of the Transform parameters.


In summary, here is what you want to check:

A) Does the optimizer uses all the iterations that
     you alloted for it ? (in each of the resolution levels).

B) If you plot the parameters of the Transform, do
     they perform big jumps in the transitions between
     on resolution level and the next ?



   Regards,


         Luis



------------------------------------------------------------------------------------
On Wed, Oct 14, 2009 at 6:08 PM,  <itkvtk123 at gmx.net> wrote:
> Hey guys,
>
> both on the ITK website and in the manual it says, that the multiresolution registration is both faster and more robust than the regular registration.
> I implemented both now and the multiresolution registration is always about 10 times slower than the regular registration.
> I use the same image data, the same optimizer, the same metrics and the same optimizer scales for both registrations.
> The multiresolution registration is run with 3 levels.
> (when I set the number of levels to 0, both registration methods run the same speed)
> Offtopic: Why can I set the number of levels to 0? It does not make sense! Does it? Does ITK use 1 level then (regular registration)?
>
> Where is the problem that the multires so slow?
>
> Best regards,
> itkvtk123
>
>
>
>
>
> --
> GRATIS für alle GMX-Mitglieder: Die maxdome Movie-FLAT!
> Jetzt freischalten unter http://portal.gmx.net/de/go/maxdome01
> _____________________________________
> 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
>