question about metric setter

pierrectl · July 2, 2020, 10:51am

Hi everyone,
I went through itkMeanSquaresImageToImageMetric.hxx file, and I was wondering what is the effect of changing this->SetUseAllPixels(true) to this->SetUseAllPixels(false)?

github.com

InsightSoftwareConsortium/ITK/blob/master/Modules/Registration/Common/include/itkMeanSquaresImageToImageMetric.hxx#L44


{
  this->SetComputeGradient(true);

  m_PerThread = nullptr;
  this->m_WithinThreadPreProcess = false;
  this->m_WithinThreadPostProcess = false;

  //  For backward compatibility, the default behavior is to use all the pixels
  //  in the fixed image.
  //  This should be fixed in ITKv4 so that this metric behaves as the others.
  this->SetUseAllPixels(true);
}

template <typename TFixedImage, typename TMovingImage>
MeanSquaresImageToImageMetric<TFixedImage, TMovingImage>::~MeanSquaresImageToImageMetric()
{
  delete[] m_PerThread;
  m_PerThread = nullptr;
}

/**

Thank you in advance for your reply!

dzenanz · July 2, 2020, 2:28pm

It probably controls whether to evaluate the metric on all the pixels or just the selected subset (which is usually faster).

pierrectl · July 2, 2020, 2:47pm

ok thank you. So how does this “selected subset” work? I mean, it divides the whole image in a few parts? And how many parts? And what is the shape of these parts? A cube?

dzenanz · July 2, 2020, 3:15pm

Usually randomly sampled. See this discussion.

blowekamp · July 2, 2020, 4:11pm

You are looking at an ITKv3 registration metric. The discussion @dzenanz linked too was for the ITKv4 registration framework. All that I recall about the ITKv3 sampling is that the point set changed each iteration, but I haven’t looked at that code for a very long time!

pierrectl · July 3, 2020, 6:42am

ok, thanks to both of you.

Is there any difference between using the metric v3 or metric v4 regarding registration performance? Actually, I thought I was using v5 of this metric…