VojtechAdam
GPT sums it up well so I will just paste its answer:
1) Meaning of the Reprojection Error
When you calibrate a camera (or a stereo setup), you collect many images of a known target (e.g., a checkerboard) at different poses. The calibration algorithm estimates:
• Intrinsic parameters (focal length, principal point, distortion coefficients), and
• Extrinsic parameters (rotation and translation between cameras in a stereo setup).
After finding these parameters, the calibration procedure projects the known 3D points of the target back onto the image plane. The reprojection error is the difference (in pixels) between:
1. The actual, detected feature locations in the image.
2. The feature locations predicted by the calibrated camera model.
In simpler terms, if your calibration is perfect, every chessboard corner (or calibration marker) in real-world 3D maps exactly onto the image’s detected corner pixel. In reality, there’s always some small difference.
• Often reported as RMS (root mean square) error in pixels.
• It can be separate for each camera in a stereo setup: “rgb reprojection error,” “left reprojection error,” “right reprojection error,” etc.
2) How to Interpret the Reprojection Error
• Lower is better. A reprojection error near 0.1–0.5 pixels is usually considered very good for typical machine vision cameras. Values up to 1 pixel can be acceptable, depending on your application accuracy needs.
• There’s no universal “golden” threshold; it depends on your lens, resolution, required precision, etc. But anything well below 1 pixel is generally “quite good.”
• If your process or code mentions a “threshold” (e.g., 1 pixel or 1.5 pixels RMS), that typically means anything below that threshold is considered “good enough” for the intended use.
So yes, you can use reprojection error to decide if your calibration is “good.” If the reported errors are reasonably low (e.g., under a pixel or so, depending on your setup), it indicates your camera is well-calibrated. If errors are significantly higher, your poses/data might have been suboptimal, or the camera model might not be capturing distortion well.
3) What Can Cause a Camera to “De-Calibrate”?
Even if you start with a great calibration, a few real-world factors can shift the camera’s geometry or how it captures light:
1. Physical Shock / Vibration
• Dropping or bumping the device, strong vibrations, or shipping the device without adequate protection can shift internal lens/sensor alignment.
2. Temperature Changes
• Significant heating or cooling can slightly expand or contract the lens barrel and sensor mount, altering focal length or introducing slight misalignment.
3. Mechanical Stresses
• Over-tightening mounts, twisting cables, or even repeated insertions/removals from a fixture can cause small changes in orientation or lens position.
4. Aging of Materials
• Over very long periods, adhesives or mechanical parts can wear or settle, gradually shifting alignment.
If any of these occur, the original calibration parameters may no longer match the real camera geometry, and the reprojection error will go up.
Thanks,
Jaka
