Result different between opencv 3.3 and opencv 4.1

,
1

I have two systems:

  1. Ubuntu with Opencv 3.3
  2. Windows with Opencv 4.1

I found the output of estimateNewCameraMatrixForUndistortRectify from the same input are different between the two systems. Why?

2

both 3.3 and 4.1 are years old and outdated

if you think, you have found a regression,
please at least update that system to current 4.5.x and check again
(then it may be time to raise an issue here )

also, can you show some example input/output ?

3

@berak Thank you for your reply.

Here is the code:

fisheye::estimateNewCameraMatrixForUndistortRectify(_Kmat, _Dmat, oldSize, Matx33d::eye(), _newKmat, balance, newSize);

Here is the output:

Ubuntu opencv 3.3:

_newKmat = (232.72, 0, 233.076, 0, 233.979, 246.974, 0.0, 0.0, 1.0);

Windows: Opencv 4.1:

_newKmat = (136.75358477913400, 0.0, 468.25649224937445, 0.0, 137.49362656177013, 250.19162967081587, 0.0, 0.0, 1.0);

4

i you give us the complete input data, we could try to repeat it

5

_Kmat = (238.2775569220771, 0.0, 333.4472958102587, 0.0, 239.56699550075558, 243.93441548309076, 0.0, 0.0, 1.0)
D = (-0.017895948523542635, -0.11236456139509969, 0.1512045667679157, -0.06372236316052718 )

oldSize = newSize, row = 480, col = 640

balance = 0.0

1 Like
6

Hi @berak, did you get a chance to look into this?

7

sorry for late, using:

cv::Mat_<double>  _Kmat(3,3);
_Kmat << 238.2775569220771, 0.0, 333.4472958102587,
         0.0, 239.56699550075558, 243.93441548309076,
         0.0, 0.0, 1.0;
cv::Mat_<double> D(1,4);
D << -0.017895948523542635, -0.11236456139509969, 0.1512045667679157, -0.06372236316052718;

cv::Size oldSize(640,480);
cv::Size newSize(oldSize);
double balance = 0.01;
cv::Mat _newKmat;

cv::fisheye::estimateNewCameraMatrixForUndistortRectify(_Kmat, D, oldSize, cv::Matx33d::eye(), _newKmat, balance, newSize);
cout << CV_VERSION << endl;
cout << newSize << endl;
cout << _newKmat << endl;

i get:

4.1.0-dev
[640 x 480]
[135.8781091743231, 0, 467.3073767842933;
 0, 136.6134133222692, 250.1263844111433;
 0, 0, 1]

and (!!!)

4.5.3-dev
[640 x 480]
[0.000641802933620525, 0, 639.1742875591813;
 0, 0.0006452760490629078, 562.6380509428207;
 0, 0, 1]

(yea, looks way off to me !)
(but, unfortunately NO idea about fisheye calibration here ;( )

edit:

test here is passing fine

@autocar , can you check ?
also, how / where did you get your calibration data ?

8

you could lobby for some regression tests on that module (open an issue on the github). those, if they existed, should have caught that, assuming it’s a specific bug someone introduced, instead of numerical issues. the parameters look obvious enough that I think the usage here is probably correct.