Fix and add tests for ellipse fitting by MaximSmolskiy · Pull Request #26694 · opencv/opencv

MaximSmolskiy · 2025-01-01T16:46:08Z

Pull Request Readiness Checklist

The main problem is that fitEllipseNoDirect function on first step finds optimal solution from space of all conic curves (besides ellipse there are at least hyperbola and parabola). It does not impose any special restrictions on solution, so that it must be only ellipse. So, found curve may easily be not ellipse. But all the following steps consider this to be ellipse, which is why we get very strange and incorrect results.

As a solution, we can add fallback to fitEllipseDirect function as soon as we realized that we got not ellipse - as has already been done in fitEllipseAMS function

opencv/modules/imgproc/src/shapedescr.cpp

Lines 641 to 689 in 4d26e16

    
           // Check that an elliptical solution has been found. AMS sometimes produces Parabolic solutions. 
        
           bool is_ellipse = (coeffs(0)  < 0 && \ 
        
                              coeffs(2)  < (coeffs(1) *coeffs(1) )/(4.*coeffs(0) ) && \ 
        
                              coeffs(5)  > (-(coeffs(2) *(coeffs(3) *coeffs(3) )) + coeffs(1) *coeffs(3) *coeffs(4)  - coeffs(0) *(coeffs(4) *coeffs(4) )) / \ 
        
                                           ((coeffs(1) *coeffs(1) ) - 4*coeffs(0) *coeffs(2) )) || \ 
        
                             (coeffs(0)  > 0 && \ 
        
                              coeffs(2)  > (coeffs(1) *coeffs(1) )/(4.*coeffs(0) ) && \ 
        
                              coeffs(5)  < (-(coeffs(2) *(coeffs(3) *coeffs(3) )) + coeffs(1) *coeffs(3) *coeffs(4)  - coeffs(0) *(coeffs(4) *coeffs(4) )) / \ 
        
                                           ( (coeffs(1) *coeffs(1) ) - 4*coeffs(0) *coeffs(2) )); 
        
           if (is_ellipse) { 
        
               double u1 = pVec(2) *pVec(3) *pVec(3)  - pVec(1) *pVec(3) *pVec(4)  + pVec(0) *pVec(4) *pVec(4)  + pVec(1) *pVec(1) *coeffs(5) ; 
        
               double u2 = pVec(0) *pVec(2) *coeffs(5) ; 
        
               double l1 = sqrt(pVec(1) *pVec(1)  + (pVec(0)  - pVec(2) )*(pVec(0)  - pVec(2) )); 
        
               double l2 = pVec(0)  + pVec(2) ; 
        
               double l3 = pVec(1) *pVec(1)  - 4.0*pVec(0) *pVec(2) ; 
        
               double p1 = 2.0*pVec(2) *pVec(3)  - pVec(1) *pVec(4) ; 
        
               double p2 = 2.0*pVec(0) *pVec(4) -(pVec(1) *pVec(3) ); 
        
               x0 = p1/l3/scale + c.x; 
        
               y0 = p2/l3/scale + c.y; 
        
               a = std::sqrt(2.)*sqrt((u1 - 4.0*u2)/((l1 - l2)*l3))/scale; 
        
               b = std::sqrt(2.)*sqrt(-1.0*((u1 - 4.0*u2)/((l1 + l2)*l3)))/scale; 
        
               if (pVec(1)  == 0) { 
        
                   if (pVec(0)  < pVec(2) ) { 
        
                       theta = 0; 
        
                   } else { 
        
                       theta = CV_PI/2.; 
        
                   } 
        
               } else { 
        
                   theta = CV_PI/2. + 0.5*std::atan2(pVec(1) , (pVec(0)  - pVec(2) )); 
        
               } 
        
               box.center.x = (float)x0; 
        
               box.center.y = (float)y0; 
        
               box.size.width = (float)(2.0*a); 
        
               box.size.height = (float)(2.0*b); 
        
               if( box.size.width > box.size.height ) 
        
               { 
        
                   float tmp; 
        
                   CV_SWAP( box.size.width, box.size.height, tmp ); 
        
                   box.angle = (float)(90 + theta*180/CV_PI); 
        
               } else { 
        
                   box.angle = (float)(fmod(theta*180/CV_PI,180.0)); 
        
               }; 
        
           } else { 
        
               box = cv::fitEllipseDirect( points ); 
        
           }

Fix #26078

Fix #26360

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

I agree to contribute to the project under Apache 2 License.
To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
The PR is proposed to the proper branch
There is a reference to the original bug report and related work
There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
The feature is well documented and sample code can be built with the project CMake

s-trinh · 2025-01-03T00:05:26Z

👍 great if it can solves all these issues

I have proposed this pull request: Add getClosestEllipsePoints() function to get the closest point on an ellipse

Feel free to comment if there is something missing or if you think this function is useless, etc.

asmorkalov · 2025-01-03T08:11:34Z

Thanks a lot for the contribution!
fitEllipseNoDirect is called from fitEllipseDirect. Inifinite recursion is possible. Also direct and noDirect functions are called in different branches here and there depending on input. M.b. we need to define "invalid" elipse output, e.g. RotatedRect(0,0,0,0,0) and handle the case on caller side. I'll analyze all cases and return back soon.

MaximSmolskiy · 2025-01-03T14:02:23Z

Yes, test Features2d_Detector_Keypoints_MSER.validation catch infinite recursion - there are all points lie on same line

[237, 47;
 238, 47;
 236, 47;
 240, 47;
 241, 47;
 239, 47;
 234, 47;
 235, 47;
 233, 47;
 222, 47;
 223, 47;
 221, 47;
 224, 47;
 220, 47;
 228, 47;
 229, 47;
 230, 47;
 227, 47;
 226, 47;
 231, 47;
 225, 47;
 219, 47;
 232, 47;
 218, 47;
 213, 47;
 214, 47;
 215, 47;
 212, 47;
 216, 47;
 211, 47;
 217, 47;
 210, 47;
 207, 47;
 208, 47;
 209, 47;
 242, 47;
 206, 47;
 244, 47;
 245, 47;
 243, 47;
 246, 47;
 197, 47;
 198, 47;
 199, 47;
 200, 47;
 201, 47;
 196, 47;
 202, 47;
 203, 47;
 195, 47;
 204, 47;
 205, 47;
 191, 47;
 192, 47;
 193, 47;
 190, 47;
 194, 47;
 189, 47;
 247, 47;
 251, 47;
 252, 47;
 253, 47;
 254, 47;
 250, 47;
 255, 47;
 249, 47;
 256, 47;
 257, 47;
 258, 47;
 248, 47;
 261, 47;
 262, 47;
 260, 47;
 263, 47;
 269, 47;
 270, 47;
 268, 47;
 271, 47;
 267, 47;
 272, 47;
 266, 47;
 265, 47;
 264, 47;
 273, 47;
 259, 47;
 277, 47;
 278, 47;
 276, 47;
 275, 47;
 279, 47;
 274, 47;
 280, 47;
 281, 47;
 188, 47;
 288, 47;
 289, 47;
 290, 47;
 285, 47;
 286, 47;
 284, 47;
 287, 47;
 291, 47;
 292, 47;
 283, 47;
 293, 47;
 294, 47;
 282, 47;
 295, 47]

…llipse-fitting Improve robustness for ellipse fitting #26773 ### Pull Request Readiness Checklist Related to #26694 Current noise addition is not very good because for example it turns degenerate case of one horizontal line into degenerate case of two parallel horizontal lines Improving noise addition leads to improved robustness of algorithms See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

asmorkalov · 2025-01-21T07:54:42Z

@MaximSmolskiy May I close this pr?

…-for-ellipse-fitting Improve robustness for ellipse fitting opencv#26773 ### Pull Request Readiness Checklist Related to opencv#26694 Current noise addition is not very good because for example it turns degenerate case of one horizontal line into degenerate case of two parallel horizontal lines Improving noise addition leads to improved robustness of algorithms See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

…itEllipseAMS Improve robustness for fitEllipseAMS #26810 ### Pull Request Readiness Checklist Related to #26694 Added functionality to add noise to points in degenerate cases and try again for `fitEllipseAMS`. `fitEllipseNoDirect` and `fitEllipseDirect` already have this See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

asmorkalov · 2025-01-22T11:48:36Z

@MaximSmolskiy ?

MaximSmolskiy · 2025-01-23T21:30:06Z

@MaximSmolskiy May I close this pr?

@asmorkalov I think no, because issues are still not fixed

It's pretty obvious to me what the matter is.
There are two main problems:

fitEllipseNoDirect fundamentally does not guarantee a valid ellipse at the output even in non-degenerate cases. Because it finds optimal conic curve from all possible conic curves (at least there are ellipses, parabollas and hyperbolas). So, even in non-degenerate cases we can pass points for which optimal conic curve will be hyperbola or parabola, but then algorithm thinks that it is always ellipse and by formulas valid only for ellipse gets very bad ellipse.
fitEllipse almost always сalls this unreliable fitEllipseNoDirect instead of more reliable fitEllipseDirect where algorithm finds optimal ellipse from only all possible ellipses (and only degenerate cases can cause problems)

So, to avoid problems with infinite recursion in this patch, I can offer to call fitEllipseDirect from fitEllipse - not fitEllipseNoDirect. If this offer suits you, I will prepare corresponding patch in this PR and retest that it fixes issues

asmorkalov · 2025-01-24T08:43:50Z

So, to avoid problems with infinite recursion in this patch, I can offer to call fitEllipseDirect from fitEllipse - not fitEllipseNoDirect. If this offer suits you, I will prepare corresponding patch in this PR and retest that it fixes issues

Please go ahead!

…b.com/MaximSmolskiy/opencv into fix-and-add-tests-for-ellipse-fitting

MaximSmolskiy · 2025-01-24T10:44:48Z

So, to avoid problems with infinite recursion in this patch, I can offer to call fitEllipseDirect from fitEllipse - not fitEllipseNoDirect. If this offer suits you, I will prepare corresponding patch in this PR and retest that it fixes issues

Please go ahead!

@asmorkalov Locally ellipse fitting tests are passed and images for issues from description haven't changed

MaximSmolskiy · 2025-01-24T17:33:27Z

Some tests failed - I'll deal with them

…-for-ellipse-fitting Improve robustness for ellipse fitting opencv#26773 ### Pull Request Readiness Checklist Related to opencv#26694 Current noise addition is not very good because for example it turns degenerate case of one horizontal line into degenerate case of two parallel horizontal lines Improving noise addition leads to improved robustness of algorithms See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

…-for-fitEllipseAMS Improve robustness for fitEllipseAMS opencv#26810 ### Pull Request Readiness Checklist Related to opencv#26694 Added functionality to add noise to points in degenerate cases and try again for `fitEllipseAMS`. `fitEllipseNoDirect` and `fitEllipseDirect` already have this See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

asmorkalov · 2025-05-12T14:33:52Z

I merged 4.x branch, but there is still test failure on mac M1:

[ RUN      ] Imgproc_FitEllipse_JavaCase.accuracy
/Users/opencv-cn/GHA-OCV-2/_work/opencv/opencv/opencv/modules/imgproc/test/test_fitellipse.cpp:105: Failure
The difference between e.size.width and sqrt(2.)*2 is 2.8278916729973163, which exceeds 0.4, where
e.size.width evaluates to 0.00053545174887403846,
sqrt(2.)*2 evaluates to 2.8284271247461903, and
0.4 evaluates to 0.40000000000000002.
/Users/opencv-cn/GHA-OCV-2/_work/opencv/opencv/opencv/modules/imgproc/test/test_fitellipse.cpp:106: Failure
The difference between e.size.height and sqrt(2.)*2 is 2.8278915229961741, which exceeds 0.4, where
e.size.height evaluates to 0.00053560175001621246,
sqrt(2.)*2 evaluates to 2.8284271247461903, and
0.4 evaluates to 0.40000000000000002.
[  FAILED  ] Imgproc_FitEllipse_JavaCase.accuracy (0 ms)

asmorkalov · 2025-05-13T06:44:28Z

Reproduced the issue locally with debug build on nvidia jetson. Will try to debug it.

asmorkalov · 2025-08-29T10:17:12Z

@MaximSmolskiy Could you rebase and fix conflicts after your previous patch?

MaximSmolskiy · 2025-08-29T20:53:57Z

@MaximSmolskiy Could you rebase and fix conflicts after your previous patch?

@asmorkalov I will do it, but I think there is still a lot of work to do here

asmorkalov · 2025-09-09T07:50:38Z

[ RUN ] fitEllipse_Modes.accuracy/0, where GetParam() = 13 Test fails on many platforms.

Details

  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.031352996826171875, which exceeds 0.01, where
  res.center.y evaluates to 55.695201873779297,
  ref_center.y evaluates to 55.726554870605469, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 0
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.16428375244140625, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 74.943496704101562,
  max(ref_size.width, ref_size.height) evaluates to 74.779212951660156, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 0
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:719: Failure
  Expected: (angleDiff(ref_angle1, res.angle)) <= (0.1), actual: 179.89 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 0
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.14357805252075195, which exceeds 0.01, where
  res.center.x evaluates to 7.0617284774780273,
  ref_center.x evaluates to 7.2053065299987793, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 1
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.25177383422851562, which exceeds 0.01, where
  res.center.y evaluates to -42.611194610595703,
  ref_center.y evaluates to -42.359420776367188, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 1
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.7310638427734375, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 200.03376770019531,
  max(ref_size.width, ref_size.height) evaluates to 199.30270385742188, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 1
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:715: Failure
  Expected: (angleDiff(ref_angle2, res.angle)) <= (0.1), actual: 179.951 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 1
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.02429962158203125, which exceeds 0.01, where
  res.center.x evaluates to -84.580696105957031,
  ref_center.x evaluates to -84.556396484375, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 2
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.067471504211425781, which exceeds 0.01, where
  res.center.y evaluates to 4.5729231834411621,
  ref_center.y evaluates to 4.6403946876525879, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 2
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.17604827880859375, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 92.930023193359375,
  max(ref_size.width, ref_size.height) evaluates to 93.106071472167969, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 2
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:719: Failure
  Expected: (angleDiff(ref_angle1, res.angle)) <= (0.1), actual: 179.962 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 2
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.148834228515625, which exceeds 0.01, where
  res.center.x evaluates to 3.876861572265625,
  ref_center.x evaluates to 3.72802734375, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 3
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.035064697265625, which exceeds 0.01, where
  res.center.y evaluates to -94.567138671875,
  ref_center.y evaluates to -94.602203369140625, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 3
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.46600341796875, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 294.24429321289062,
  max(ref_size.width, ref_size.height) evaluates to 294.71029663085938, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 3
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:719: Failure
  Expected: (angleDiff(ref_angle1, res.angle)) <= (0.1), actual: 0.105774 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 3
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.14314651489257812, which exceeds 0.01, where
  res.center.x evaluates to 51.765361785888672,
  ref_center.x evaluates to 51.622215270996094, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 5
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.20017647743225098, which exceeds 0.01, where
  res.center.y evaluates to -1.8791747093200684,
  ref_center.y evaluates to -2.0793511867523193, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 5
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:711: Failure
  The difference between min(res.size.width, res.size.height) and min(ref_size.width, ref_size.height) is 0.553375244140625, which exceeds sizeDiff, where
  min(res.size.width, res.size.height) evaluates to 215.90545654296875,
  min(ref_size.width, ref_size.height) evaluates to 215.35208129882812, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 5
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.480255126953125, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 433.421875,
  max(ref_size.width, ref_size.height) evaluates to 432.94161987304688, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 5
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.12608122825622559, which exceeds 0.01, where
  res.center.x evaluates to 0.96769058704376221,
  ref_center.x evaluates to 1.0937718152999878, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 6
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:711: Failure
  The difference between min(res.size.width, res.size.height) and min(ref_size.width, ref_size.height) is 0.1369476318359375, which exceeds sizeDiff, where
  min(res.size.width, res.size.height) evaluates to 102.49685668945312,
  min(ref_size.width, ref_size.height) evaluates to 102.35990905761719, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 6
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.1242523193359375, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 238.30842590332031,
  max(ref_size.width, ref_size.height) evaluates to 238.18417358398438, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 6
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.01079559326171875, which exceeds 0.01, where
  res.center.x evaluates to -35.456672668457031,
  ref_center.x evaluates to -35.445877075195312, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 7
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.032003402709960938, which exceeds 0.01, where
  res.center.y evaluates to -17.682683944702148,
  ref_center.y evaluates to -17.714687347412109, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 7
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:711: Failure
  The difference between min(res.size.width, res.size.height) and min(ref_size.width, ref_size.height) is 0.7824859619140625, which exceeds sizeDiff, where
  min(res.size.width, res.size.height) evaluates to 168.25273132324219,
  min(ref_size.width, ref_size.height) evaluates to 169.03521728515625, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 7
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.340789794921875, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 439.63427734375,
  max(ref_size.width, ref_size.height) evaluates to 439.29348754882812, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 7
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.066010951995849609, which exceeds 0.01, where
  res.center.x evaluates to 3.6527106761932373,
  ref_center.x evaluates to 3.7187216281890869, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 10
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.01129913330078125, which exceeds 0.01, where
  res.center.y evaluates to 87.745697021484375,
  ref_center.y evaluates to 87.756996154785156, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 10
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.10178375244140625, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 88.948890686035156,
  max(ref_size.width, ref_size.height) evaluates to 89.050674438476562, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 10
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:715: Failure
  Expected: (angleDiff(ref_angle2, res.angle)) <= (0.1), actual: 179.884 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 10
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.071498870849609375, which exceeds 0.01, where
  res.center.y evaluates to -43.936538696289062,
  ref_center.y evaluates to -44.008037567138672, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 13
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:711: Failure
  The difference between min(res.size.width, res.size.height) and min(ref_size.width, ref_size.height) is 0.10516738891601562, which exceeds sizeDiff, where
  min(res.size.width, res.size.height) evaluates to 46.379077911376953,
  min(ref_size.width, ref_size.height) evaluates to 46.484245300292969, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 13
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:719: Failure
  Expected: (angleDiff(ref_angle1, res.angle)) <= (0.1), actual: 179.946 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 13
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.05939483642578125, which exceeds 0.01, where
  res.center.x evaluates to 87.68695068359375,
  ref_center.x evaluates to 87.746345520019531, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 14
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.03304290771484375, which exceeds 0.01, where
  res.center.y evaluates to 65.046295166015625,
  ref_center.y evaluates to 65.079338073730469, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 14
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.2847900390625, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 147.41598510742188,
  max(ref_size.width, ref_size.height) evaluates to 147.70077514648438, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 14
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:715: Failure
  Expected: (angleDiff(ref_angle2, res.angle)) <= (0.1), actual: 179.914 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 14
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.24158477783203125, which exceeds 0.01, where
  res.center.x evaluates to -82.680130004882812,
  ref_center.x evaluates to -82.438545227050781, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 15
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.14835548400878906, which exceeds 0.01, where
  res.center.y evaluates to -19.137044906616211,
  ref_center.y evaluates to -18.988689422607422, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 15
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:711: Failure
  The difference between min(res.size.width, res.size.height) and min(ref_size.width, ref_size.height) is 0.2572021484375, which exceeds sizeDiff, where
  min(res.size.width, res.size.height) evaluates to 204.713134765625,
  min(ref_size.width, ref_size.height) evaluates to 204.4559326171875, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 15
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.54803466796875, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 381.34765625,
  max(ref_size.width, ref_size.height) evaluates to 381.89569091796875, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 15
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:719: Failure
  Expected: (angleDiff(ref_angle1, res.angle)) <= (0.1), actual: 179.938 vs 0.1
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 15
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.x and ref_center.x is 0.083209991455078125, which exceeds 0.01, where
  res.center.x evaluates to 20.930164337158203,
  ref_center.x evaluates to 21.013374328613281, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 19
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:709: Failure
  The difference between res.center.y and ref_center.y is 0.45600008964538574, which exceeds 0.01, where
  res.center.y evaluates to -1.4683732986450195,
  ref_center.y evaluates to -1.9243733882904053, and
  0.01 evaluates to 0.01.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 19
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:711: Failure
  The difference between min(res.size.width, res.size.height) and min(ref_size.width, ref_size.height) is 0.592803955078125, which exceeds sizeDiff, where
  min(res.size.width, res.size.height) evaluates to 167.90231323242188,
  min(ref_size.width, ref_size.height) evaluates to 168.4951171875, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 19
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:712: Failure
  The difference between max(res.size.width, res.size.height) and max(ref_size.width, ref_size.height) is 0.79290771484375, which exceeds sizeDiff, where
  max(res.size.width, res.size.height) evaluates to 431.66806030273438,
  max(ref_size.width, ref_size.height) evaluates to 430.87515258789062, and
  sizeDiff evaluates to 0.10000000149011612.
  Google Test trace:
  /home/ci/opencv/modules/imgproc/test/test_convhull.cpp:656: iteration 19

asmorkalov · 2025-11-12T12:36:34Z

@MaximSmolskiy Could you take a look on the failed tests?

Fix and add tests for ellipse fitting

d3092de

MaximSmolskiy force-pushed the fix-and-add-tests-for-ellipse-fitting branch from e3c66ca to d3092de Compare January 2, 2025 19:30

asmorkalov added category: imgproc bug labels Jan 3, 2025

MaximSmolskiy mentioned this pull request Jan 5, 2025

Fix tests for ellipse fitting #26716

Merged

6 tasks

asmorkalov added this to the 4.12.0 milestone Jan 8, 2025

MaximSmolskiy mentioned this pull request Jan 15, 2025

Improve robustness for ellipse fitting #26773

Merged

6 tasks

MaximSmolskiy mentioned this pull request Jan 20, 2025

Improve robustness for fitEllipseAMS #26810

Merged

6 tasks

Merge branch '4.x' into fix-and-add-tests-for-ellipse-fitting

da77935

MaximSmolskiy added 2 commits January 24, 2025 13:41

Change approach

24e18fc

Merge branch 'fix-and-add-tests-for-ellipse-fitting' of https://githu…

7157ab0

…b.com/MaximSmolskiy/opencv into fix-and-add-tests-for-ellipse-fitting

MaximSmolskiy added 2 commits January 24, 2025 20:50

Fix Imgproc_FitEllipse_HorizontalLine.accuracy test

72f620e

Fix tests

c21bf23

asmorkalov requested a review from vpisarev May 12, 2025 13:53

Merge branch '4.x' into fix-and-add-tests-for-ellipse-fitting

cd94355

Restored DM nornalization.

32b79c1

asmorkalov modified the milestones: 4.12.0, 4.13.0 Jun 27, 2025

MaximSmolskiy added 2 commits August 29, 2025 23:59

Merge branch '4.x' into fix-and-add-tests-for-ellipse-fitting

9fcb222

Fix

7eab37b

asmorkalov modified the milestones: 4.13.0, 4.14.0 Dec 9, 2025

	// Check that an elliptical solution has been found. AMS sometimes produces Parabolic solutions.
	bool is_ellipse = (coeffs(0) < 0 && \
	coeffs(2) < (coeffs(1) coeffs(1) )/(4.coeffs(0) ) && \
	coeffs(5) > (-(coeffs(2) (coeffs(3) coeffs(3) )) + coeffs(1) coeffs(3) coeffs(4) - coeffs(0) (coeffs(4) coeffs(4) )) / \
	((coeffs(1) coeffs(1) ) - 4coeffs(0) *coeffs(2) )) \|\| \
	(coeffs(0) > 0 && \
	coeffs(2) > (coeffs(1) coeffs(1) )/(4.coeffs(0) ) && \
	coeffs(5) < (-(coeffs(2) (coeffs(3) coeffs(3) )) + coeffs(1) coeffs(3) coeffs(4) - coeffs(0) (coeffs(4) coeffs(4) )) / \
	( (coeffs(1) coeffs(1) ) - 4coeffs(0) *coeffs(2) ));
	if (is_ellipse) {
	double u1 = pVec(2) pVec(3) pVec(3) - pVec(1) pVec(3) pVec(4) + pVec(0) pVec(4) pVec(4) + pVec(1) pVec(1) coeffs(5) ;
	double u2 = pVec(0) pVec(2) coeffs(5) ;
	double l1 = sqrt(pVec(1) pVec(1) + (pVec(0) - pVec(2) )(pVec(0) - pVec(2) ));
	double l2 = pVec(0) + pVec(2) ;
	double l3 = pVec(1) pVec(1) - 4.0pVec(0) *pVec(2) ;
	double p1 = 2.0pVec(2) pVec(3) - pVec(1) *pVec(4) ;
	double p2 = 2.0pVec(0) pVec(4) -(pVec(1) *pVec(3) );

	x0 = p1/l3/scale + c.x;
	y0 = p2/l3/scale + c.y;
	a = std::sqrt(2.)sqrt((u1 - 4.0u2)/((l1 - l2)*l3))/scale;
	b = std::sqrt(2.)sqrt(-1.0((u1 - 4.0u2)/((l1 + l2)l3)))/scale;
	if (pVec(1) == 0) {
	if (pVec(0) < pVec(2) ) {
	theta = 0;
	} else {
	theta = CV_PI/2.;
	}
	} else {
	theta = CV_PI/2. + 0.5*std::atan2(pVec(1) , (pVec(0) - pVec(2) ));
	}

	box.center.x = (float)x0;
	box.center.y = (float)y0;
	box.size.width = (float)(2.0*a);
	box.size.height = (float)(2.0*b);
	if( box.size.width > box.size.height )
	{
	float tmp;
	CV_SWAP( box.size.width, box.size.height, tmp );
	box.angle = (float)(90 + theta*180/CV_PI);
	} else {
	box.angle = (float)(fmod(theta*180/CV_PI,180.0));
	};


	} else {
	box = cv::fitEllipseDirect( points );
	}

Uh oh!

Conversation

MaximSmolskiy commented Jan 1, 2025

Pull Request Readiness Checklist

Uh oh!

s-trinh commented Jan 3, 2025

Uh oh!

asmorkalov commented Jan 3, 2025

Uh oh!

MaximSmolskiy commented Jan 3, 2025

Uh oh!

asmorkalov commented Jan 21, 2025

Uh oh!

asmorkalov commented Jan 22, 2025

Uh oh!

MaximSmolskiy commented Jan 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

asmorkalov commented Jan 24, 2025

Uh oh!

MaximSmolskiy commented Jan 24, 2025

Uh oh!

MaximSmolskiy commented Jan 24, 2025

Uh oh!

asmorkalov commented May 12, 2025

Uh oh!

asmorkalov commented May 13, 2025

Uh oh!

asmorkalov commented Aug 29, 2025

Uh oh!

MaximSmolskiy commented Aug 29, 2025

Uh oh!

asmorkalov commented Sep 9, 2025

Uh oh!

asmorkalov commented Nov 12, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

MaximSmolskiy commented Jan 23, 2025 •

edited

Loading