One important measure is the reprojection error of your calibration targets. If you do a standard calibration, you get the internal camera matrix and the relative pose for all targets to the camera. You can then use this information to project all patterns into your images and check if the projected point is close to its former detection. (the rmse of theses pixel distances is also the return value of calibrateCamera)
A good calibration is well below 1 px as a rule of thumb.
No. That will be just random noise. You could take your set of calbration images and run the calibration several times on subsets of these images. (e.g. always remove one image and run the calibration on the other N-1 images). You should get rather similar values for fx that will diverge after some decimals. That approach could give you a first insight in how many decimals are important.
Asked: 2015-04-09 18:04:19 -0600
Seen: 496 times
Last updated: Apr 10 '15
