[Insight-users] Question about RegularStepGradientDescentOptimizer

[Feng Qiao]

Thanks Luis and KK for your helpful comments. The metric I used was the 
NormalizedCorrelationImageToImageMetric. And the PET images I tried to 
register were really noisy, so I assume that the metric might also be 
pretty noisy. I tried to use the multi-resolution approach to get better 
results. In my test I used 3 levels, for the first level I used a max 
step length of 20 (now according to Luis, that value might be too large) 
and a min step length of 0.1, and each subsequent level continues with 
the current step length and reduce the min step length by a factor of 
10. Here I assume the unit of these numbers is millimeter, is that right?
Also for noisy image metric, is there any more appropriate optimizer to 
use other than the RegularStepGradientDescentOptimizer? The 
GradientDescentOptimizer seems to be too simple without a converging 
condition...
Thanks.

Feng

Karthik Krishnan wrote:

> Hi Feng,
>
> You can also try plotting the metric around the starting position with 
> respect to the parametric space.
>
> Set the metric with the initial transform and invoke
> metric->GetValueAndDerivative( parameters, value, derivative );
> and write the output into an image. You can do this with respect to 
> the translation parameters, and use Paraview to throw contours and 
> visualize the metric. This will tell you if you have local minima.
>
> Alternatively you can write the derivative to the file and plot the 
> deformation field using Paraview again (See "Visualizing deformation 
> fields using Paraview") section of the Software Guide. This field will 
> roughly indicate the direction the gradient descent optimizer will 
> move in.
>
> In general this will tell you if your metric is noisy or smooth. Its 
> hard to converge using the GradientDescentOptimizer if your metric is 
> noisy.
>
> Also, if you are using any of the metrics derived from the 
> HistogramImageToImageMetric, you can plot the joint histogram too.
> See  Examples/Registration/ImageRegistrationHistogramPlotter.cxx
> Please cvs update cause they were added two minutes ago.
>
> Thanks
> regards
> karthik
>
>
>
>
>
>
>
>
>
>
> Feng Qiao wrote:
>
>> Hi,
>> I tried to use the RegularStepGradientDescentOptimizer to register 
>> PET images. When I monitor the registration progress, I found that 
>> the measure value initially goes down during the first couple of 
>> iterations, then it just kept increasing for many iterations (the 
>> metric was set to be minimized). The increment for each step was not 
>> that large though, and it did stop finally. The parameters I  used 
>> was:  max step length =20, min step length = 0.01.
>> I'm just wondering if this is normal...and if so why is the case? Any 
>> help would be appreciated.
>>
>> Feng
>>
>>
>> _______________________________________________
>> Insight-users mailing list
>> Insight-users at itk.org
>> http://www.itk.org/mailman/listinfo/insight-users
>>
>
>