Hough Transform 2D Circles Image Filter GetCircles patch.

Algorithms
36

FYI, I just had a closer look at how the patch I proposed last week, “PERF: Improved speed of copying and resizing NeighborhoodAllocator” http://review.source.kitware.com/#/c/22862/ would affect the performance of ITK’s Hough Transform.

I observed a 10% reduction (approximately) of the duration of filter->Update(), typically going from 50 to 45 seconds, on the following test program, using VS2017, Release, 64-bit:

#include <itkHoughTransform2DCirclesImageFilter.h>
#include <iostream>
#include <ctime>

int main()
{
  typedef unsigned char PixelType;
  const auto image = itk::Image<PixelType>::New();
  enum { sizeX = 4096, sizeY = sizeX };
  image->SetRegions({ sizeX, sizeY });
  image->Allocate();
  image->FillBuffer(1);

  const auto filter =
    itk::HoughTransform2DCirclesImageFilter<PixelType, unsigned char, double>::New();
  filter->SetInput(image);

  const auto time1 = std::time(nullptr);
  filter->Update();
  const auto time2 = std::time(nullptr);

  std::cout << "Duration: " << (time2 - time1) << " seconds" << std::endl;
}

HoughTransform2DCirclesImageFilter::GenerateData() indirectly does a lot of Neighborhood allocations, by calling GaussianDerivativeImageFunction::EvaluateAtIndex on each input pixel > Threshold.