Fisheye distortion coefficient format

[shreyas23]

Hey @rameau-fr hope you're doing well!

I'm calibrating a set of fisheye cameras and initializing K with some nominal known values while setting the distortion coefficients to 0. I noticed that k4 is always 0 after optimization.

After brief investigation it seems like there's a discrepancy in the format and order of the distortion vector:

via OpenCV in fisheye::calibrate() : Output vector of distortion coefficients (k1, k2, k3, k4).

MC-Calib/McCalib/include/Camera.hpp

Line 43 in 0f893aa

// fx,fy,u0,v0,k1,k2,k3,k4 (Kannala)

setDistortionVector:

MC-Calib/McCalib/src/Camera.cpp

Lines 59 to 63 in 0f893aa

	if (distortion_model_ == 1) {
	intrinsics_[4] = distortion_vector.at<double>(0);
	intrinsics_[5] = distortion_vector.at<double>(1);
	intrinsics_[8] = distortion_vector.at<double>(2);
	intrinsics_[6] = distortion_vector.at<double>(3);

getDistortionVector:

MC-Calib/McCalib/src/Camera.cpp

Lines 85 to 90 in 0f893aa

	else if (distortion_model_ == 1) {
	cv::Mat distortion_vector = cv::Mat(1, 4, CV_64F, cv::Scalar(0));
	distortion_vector.at<double>(0) = intrinsics_[4];
	distortion_vector.at<double>(1) = intrinsics_[5];
	distortion_vector.at<double>(4) = intrinsics_[8];
	distortion_vector.at<double>(2) = intrinsics_[6];

but then during optimization:

MC-Calib/McCalib/include/OptimizationCeres.h

Lines 67 to 70 in 0f893aa

	const T k1 = Intrinsics[4]; //
	const T k2 = Intrinsics[5]; //
	const T k3 = Intrinsics[6]; //
	const T k4 = Intrinsics[7]; //

If I understand correctly, the following changes should be made:

setDistortionVector:
intrinsics_[6] = distortion_vector.at<double>(2);
intrinsics_[7] = distortion_vector.at<double>(3);

getDistortionVector:
distortion_vector.at<double>(3) = intrinsics_[7]

Let me know if this sounds right to you.

[rameau-fr]

Hello @shreyas23, I do not know how to thank you enough for reporting it !!
I think you are right. I will double-check this issue with @BAILOOL, and we will make the change as soon as possible. It is quite a big bug here. It might also be the reason why some persons calibrating large field-of-view cameras were facing issues during calibration.
Please, keep me in touch if you make some tests with your modified version of the code.

[shreyas23]

My pleasure! I did run some tests while fixing it on my end and seemingly had better convergence. I can run some comparisons when I get a chance but the effects likely aren't significant in my scenario as my FoV is not too high.