Is reprojection error enough in stereo calibration?

Question

How can I evaluate the quality of calibration procedure when I calibrate one stereo camera? OpenCV uses reprojection error to quantify calibration quality.

Many reason may cause the error and I learned by google search that the possible reasons are followed

lens distortion / yaw error / sensor tilt / pitch error / roll error / baseline error / focal length error.

During the stereo camera manufacturing process, calibration is performed. The initial calibration produces the intrinsic and extrinsic cameras parameters which are used to remove lens distortion and provide image rectification. Is reprojection error enough in stereo calibration? My goal is to calculate the depth map.

Answer 1

The RMS reprojection error is a statistic: a number that summarizes an entire distribution of errors. It is subject to all the usual caveats that apply to any statistic: the intuition you have about its meaning will be correct as long as some assumptions underlying it apply (e.g. that the errors are symmetrically distributed and the tails of the distribution are not fat).

Generally speaking, one should use quality figures dictated by the task at hand. If you cannot anticipate in advance the kind of scenes your stereo rig will be used to reconstruct then yes, RMS reprojection error is probably the best you can do. On the other hand, if your application is, say, quality control, and you are always reconstructing exactly the same kind of object (e.g. a mechanical part coming off a conveyor belt), it would make much more sense to express the precision of your rig directly in terms of its 3D reconstruction error, because in that case you want to sell guarantees of the form "My stereo rig will be off by less than 0.1mm with probability higher than 99.9%", for example.