cv::cuda::HoughCirclesDetector and cv::HoughCircles do not give the same results

[chacha21]

Hello,

System information (version)

OpenCV 3.1.0, Windows 7 64 bits, Visual Studio 2010, Cuda 8

Detailed description

I have noticed that for a similar input, cv::cuda::HoughCirclesDetector and cv::HoughCircles do not give the same results. I am not sure if it is normal, because the documentation does not tell that it should be the case (algorithms may be different).
However, it would be convenient, either to fix it, or explain how to circumvent it by adapting the parameters.

Steps to reproduce

Image provided as attachement, and here is a sample code :

cv::Mat img = cv::imread("test.png", cv::IMREAD_GRAYSCALE);

  float dp = 1;
  float minDist = 100;
  int cannyThreshold = 122;
  int accThreshold = 50;
  int minRadius = 1;
  int maxRadius = std::max(img.size().width, img.size().height);

  std::vector<cv::Vec3f> cpuCircles;
  cv::HoughCircles(img, cpuCircles, cv::HOUGH_GRADIENT, dp, minDist, cannyThreshold, accThreshold, minRadius, maxRadius);

  std::vector<cv::Vec3f> gpuCircles;
  cv::Ptr<cv::cuda::HoughCirclesDetector> gpuDetector = cv::cuda::createHoughCirclesDetector(dp, minDist, cannyThreshold, accThreshold, minRadius, maxRadius);
  cv::cuda::GpuMat gpuMat;
  cv::cuda::GpuMat gpuResult;
  gpuMat.upload(img);
  gpuDetector->detect(gpuMat, gpuResult);
  gpuCircles.resize(gpuResult.size().width);
  if (!gpuCircles.empty())
    gpuResult.row(0).download(cv::Mat(gpuCircles).reshape(3, 1));

  printf("CPU circles :\n");
  for(std::vector<cv::Vec3f>::const_iterator it = cpuCircles.begin() ; it != cpuCircles.end() ; ++it)
    printf("(%f,%f,%f)", it->val[0], it->val[1], it->val[2]);
  printf("\n");
  printf("GPU circles :\n");
  for(std::vector<cv::Vec3f>::const_iterator it = gpuCircles.begin() ; it != gpuCircles.end() ; ++it)
    printf("(%f,%f,%f)", it->val[0], it->val[1], it->val[2]);
  printf("\n");

Output result :
CPU circles :
(165.500000,329.500000,56.925388)
GPU circles :
(165.500000,328.500000,100.000000)(165.500000,328.500000,56.000000)(165.500000,3
28.500000,60.000000)(165.500000,328.500000,65.000000)(165.500000,328.500000,73.0
00000)(165.500000,328.500000,87.000000)(165.500000,328.500000,151.000000)(165.50
0000,328.500000,178.000000)(165.500000,328.500000,180.000000)(165.500000,328.500
000,183.000000)(165.500000,328.500000,185.000000)(165.500000,328.500000,187.0000
00)(165.500000,328.500000,190.000000)(165.500000,328.500000,224.000000)

[chacha21]

I did not mention why I would like to see the same results :
I want the cuda implementation to be an optimization called under the hood if the hardware is compatible. Thus, I need the results to be equivalent so that it can be transparent to the user.

[alalek]

GPU-based algorithms with "bit-exact" results are usually very ineffective (except very simple cases).

Usage questions should go to User Q/A forum: http://answers.opencv.org

[chacha21]

I am not sure to understand what you mean.
What I wanted to highlight is that there is not the same number of circles. The CPU returns 1 circle, the GPU 14 circles.
If there was the same number of circles, I wouldn't care that the floating values of their centers would not be the same exact floating values.

[alalek]

I mean that in general your should not expect "the same" results from CPU and GPU algorithms.
There is usual consequence: different floating values -> different threshold result -> different number of results.

If you have evidences that there is real bug in implementation then please open a ticket with details.

[chacha21]

I think I have found evidences of a real difference in implementation, but it this a bug ?
When using a dummy image with just white concentric circles on black background:

the CPU version reports 1 circle because the "minDistance" parameter is used while trying to find the best radii.
the GPU version reports multiple circles, because the "minDistance" is used before trying to find the best radii.

Those are two different interpretations of the minDistance parameter, because the doc does not tell how it should handle concentric circles.

[zachary-zheng]

@chacha21 @alalek Hello, I have encountered the same phenomenon that GPU gave more circles result, and it seems the parameter ''minDist Minimum distance'' is valid. The centers of the detected circles are obvious larger than minDist.
why?