Reprojection Error

The pixel distance between a 3D point reprojected onto the image and where it was actually observed.

Reprojection error measures how accurate a set of estimated camera parameters, camera pose, and 3D points really are: take an estimated 3D point, reproject it onto the image using the estimated intrinsics and extrinsics, and compare that to the pixel actually observed for the corresponding point — the distance between them, in pixels, is the reprojection error. It's the most common optimization objective and quality indicator in geometric vision: in camera calibration, a smaller average reprojection error on the calibration board's corners is better (a MATLAB documentation example gives 0.19 pixels), and if it's noticeably large, the worst images can be removed and the camera recalibrated; bundle adjustment jointly adjusts all camera poses and 3D points to minimize the total reprojection error; and PnP pose solving, triangulation, and SLAM back-ends are all optimized around it.

ExampleCalibrating a wrist camera with a checkerboard, OpenCV's calibrateCamera returns an overall RMS reprojection error; if a few images show noticeably higher error (usually from misdetected corners), removing them and recalibrating is standard practice.

Also called

Re-Projection Error

Related

Camera Calibration · Bundle Adjustment · Perspective-n-Point · Triangulation · Pinhole Camera Model · Hand-Eye Calibration

Sources

Wikipedia: Reprojection error
MathWorks: Evaluating the Accuracy of Single Camera Calibration
Wikipedia: Bundle adjustment