[Insight-users] interpretation of the gradient step

[Luis Ibanez]

Hi Romain,


Thanks for the detailed description of your problem.

The parameter that you want to use for stopping the
RegularStepGradientDescent is the


             MinimumStepLength


This is the size of the step length at which the optimizer
will consider to have find enough precision in the parameters
of the Transform.

The code will be something like:

        optimizer.SetMinimumStepLength(  0.01 );

Notice the this is to be interpreted in terms of the Transform
parameters, and weighted by the ParameterScaling.


If you were using a TranslationTransform, then the mapping is
quite straightforward. The MinimalStepLength could be interpreted
in millimeters, and then you just need to set it to 1/10th of
the pixel size in your image.



   Regards,



      Luis



--------------------------
Romain Lebel wrote:
> Hi everybody (merry christmas in advance!)
> 
> I'm trying to make my registration process faster. I'm using MeanSquares 
> metric and RegularStepGradientDescentOptimizer.
> 
> If want to stop the registration process for example when images are 
> registred at one tenth of a pixel.
> 
> How can i interpret the current steplength (or maybe another parameter) 
> to have an information like "at how many parts of a pixel my images are 
> registred"?
> 
> To summarize i need an information related to the images themselves that 
> i can interpret in order to stop the registration when it reaches a 
> certain satisfaying value. (metrics can take very different value from 
> one couple of image to another so it is difficult to stop using the metric)
>