Extrinsic calibration higher reprojection error compared to intrinsic calibration

Hello everybody!

I am using two cameras in order to create a stereo system. Firstly, I estimate the intrinsic parameters for each camera (obtaining for both cameras, average reprojection errors of around 0.28 px.) and then I am estimating the extrinsic parameters by importing the previously-obtained intrinsic values. The average reprojection errors for the stereo system are around 1 px. which to me seems a way larger error compared to the error obtained for the intrinsic parameters. Can anybody give me some intuition regarding some causes that influence the stereo calibration to perform poorly than the monocular calibration?

Answer 1

[Matt J]

Some possible causes?

1. The estimates are based on intrinsics that already have some error in them. The error can only increase from there.
2. The quality and accuracy of your measurements
3. The configuration of points used to do the estimation. Some configurations encode more or better information than others. For example, if all your points are coplanar, calibration is ill-posed
4. The distance of the plane you're reprojecting on may be different.