There are different methods to evaluate the accuracy. Here we use the reprojection error to determine the accuracy. First, we use homography to calculate the camera's pose with a know chessboard. Then the corners on the chessboard are projected to the image. Next, we remove the distortion of the image with the calibration information. And the corners are detected in the image, which are corresponding to the corners on the chessboard. Finally, we calculate the Normalised Mean Square Error (NMSE) to determine the accuracy. If the output error value is small, then the calibration is good. Otherwise, we need to recalibrate the camera.
Our example:
Install OpenCV, C++
A few updates to the instructions above were needed.
$> cd ~
$> git clone https://github.com:chengKID/Evaluation_Calibration.git
$> cd Evaluation_Calibration
$> mkdir build && cd build
$> cmake ..
$> make
$> ./calib_evaluation./calib_evaluation xxxx/ yyyy zzzz_1 zzzz_2 wwww
where xxxx/ stands for diractory of the camera.txt && (Note: do not forgette the "/")
yyyy stands for checkboad square size in meter
zzzz stands for the number of valid corners, e.g. for 15x10 zzzz_1 is 15 and zzzz_2 is 10
wwww stands for choosed an image, which has been calibrated
Green points in image stand for reprojected points
Red points in image stand for corner points in rectified image