cameraIntr​insicsFrom​OpenCV does not seem to account for differences in distortion model

The function cameraIntrinsicsFromOpenCV in the Computer Vision toolbox seems to give out wrong distortion coefficients.

According to the MATLAB documentation, the radial distortion is expressed as follows:

However, the OpenCV documentation states that the radial distortion is expressed as:

And thus, the results of MATLAB's undistort functions do not seem to be right when using the camera intrinsics from OpenCV. Is there a way to undistort points manually?

Answer 1

[Giridharan Kumaravelu]

The two equations mentioned in your question are essentially the same if the higher order radial distortion coefficients (k4-k6) in OpenCV are set to zero.

By default, OpenCV's calibration tools like cv.calibrateCamera calibrate for two radial distortion coefficients (k1 and k2) only. Often times, this is sufficient to model normal lenses.

Unless you have explicitly calibrated for more than 3 radial coefficients in OpenCV, you should be able to accurately import those camera parameters from OpenCV using cameraIntrinsicsFromOpenCV.

[Min seok Kang]

Thank you for your answer. In my case, I calibrated for all six coefficients, as I was using the Microsoft Azure Kinect hardware. The calibration function in the Azure Kinect SDK gives me six coefficients. So, the undistort and projection functions do not work well enough for my purposes. Is there an option to utilize all six coefficients, or do I need to code my own function for undistorting the image?

[Giridharan Kumaravelu]

Unfortunately, CVT camera calibration tools currently do not support using more than 3 radial distortion coefficients. Isn't there a way to limit the number of radial distortion coefficients in the Azure Kinect SDK? If not, could you directly calibrate the camera using CVT's Camera Calibrator app or using the estimateCameraParameters function?

[Min seok Kang]

In the case of the Azure Kinect SDK, unfortunately, there is no option to choose the number of distortion coefficients. I guess I should calibrate the camera myself using the toolbox, or write code to do the math to undistort the points.