[Insight-users] no. of iterations in 3D registration...

[Kajetan Berlinger]

Hi Luis,

I have a question with respect to registration/optimization as well.
In the example below, I use the regualar step gradient descent method and 
correlation coefficient for a simple rigid 2D-2D registration with 
translation only. Visual inspection of the result is quite satisfying but I 
don't understand the way of the optimizer through its parameter space.

(1) Why is the final result not the translation of iteration 5? This iteration 
obviously yielded the minimum value.

(2) Why is the metric value of iteration 11 (last output of the observer) 
different to the final metric, although they are both based on the same 
translation?

************************************************
Current Iterations: 0
Translation X: -0.826109
Translation Y: 3.91376
Similarity Value: -0.92855
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 1
Translation X: -1.454
Translation Y: 7.86417
Similarity Value: -0.948779
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 2
Translation X: -1.75847
Translation Y: 11.8526
Similarity Value: -0.956318
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 3
Translation X: 0.0892366
Translation Y: 11.0871
Similarity Value: -0.953901
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 4
Translation X: 0.302107
Translation Y: 9.09845
Similarity Value: -0.955402
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 5
Translation X: 0.856397
Translation Y: 9.93077
Similarity Value: -0.956675      <---------------------------------------(1)
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 6
Translation X: 0.410382
Translation Y: 9.70478
Similarity Value: -0.956346
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 7
Translation X: 0.474056
Translation Y: 9.20885
Similarity Value: -0.956433
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 8
Translation X: 0.201551
Translation Y: 8.78964
Similarity Value: -0.956553
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 9
Translation X: 0.44661
Translation Y: 8.74017
Similarity Value: -0.956645
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 10
Translation X: 0.391267
Translation Y: 8.85226
Similarity Value: -0.956657
Pixels considered: 46767
************************************************


************************************************
Current Iterations: 11
Translation X: 0.397056
Translation Y: 8.79002
Similarity Value: -0.956625      <----------------------------- (2)
Pixels considered: 46767
************************************************

************************************************
Image B:
Current DRR Pair: 5
************************************************
Final Result:
Translation X: 0.397056
Translation Y: 8.79002
Iterations: 13
Similarity Value: -0.956572   <-------------------------------- (2)
Pixels considered: 46767
Optimizer: Regular Step Gradient Descent
Metric: Correlation Coefficient
************************************************
************************************************


Thanks for helping me!

Kaj


On Thursday 10 August 2006 09:39, Luis Ibanez wrote:
> Hi Shahab,
>
> As long as you set up the optimizer to minimize, the metric should
> never increase for more than one consecutive iteration. That is,
> the metric can increase, but then it should decrease in the next
> iteration.
>
> There is a very very small chance that a metric can increase in
> two consecutive iterations, but it requires a convoluted configuration
> of the cost function.
>
> Can you post a plot of the metric values versus iteration number,
> just as it is done in the ITK Software Guide examples ?
>
>      http://www.itk.org/ItkSoftwareGuide.pdf
>
> How many iterations to use depends on what kind of optimizer you
> are using and how large you define the step length (or its equivalent).
>
> You may want to look at the tutorial on image registration:
>
> at
>
>        http://www.itk.org/HTML/Tutorials.htm
>
> more specifically at
>
>
> http://www.itk.org/CourseWare/Training/RegistrationMethodsOverview.pdf
>
>
> The effect and step length versus number of iterations is also
> discussed in the Image Registration chapter of the Software Guide.
>
>
>
> Regards
>
>
>
>     Luis
>
>
>
> --------------------------
>
> Shahabuddin Ansari wrote:
> > Hello,
> >
> > In 3D rigid registration, I am using versor transform, mean square
> > metric, and linear interpolator. The observation of the final parameters
> > shows that the metric value first decreases to a minimum and then start
> > back to increase. In my understanding, it should fluctuate around the
> > optimal value even the number of iterations are much higher (say 200)
> > then the required. Please commnets (say 50).
> >
> > Thanks
> > Shahab
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users