Acceptable tracking camera calibration reprojection error
-
Hi
Does anyone know what is the acceptable figures for the reported reprojection error after the calibration is done?
The example in the documentation shows a figure of 0.17.. but im usually getting 0.5 - 0.8 -ish..
Does it matter?
-
We have the calibration script require a re-projection error <= 0.5. We've found this to be a reliable value where our computer vision services perform well but is also attainable with the calibration tool. If you're having trouble getting to this value make sure that the camera is focused and doesn't have and smears on the lens.
-
@Alex-Gardner said in Acceptable tracking camera calibration reprojection error:
We have the calibration script require a re-projection error <= 0.5. We've found this to be a reliable value where our computer vision services perform well but is also attainable with the calibration tool. If you're having trouble getting to this value make sure that the camera is focused and doesn't have and smears on the lens.
I have the voxl-calibrate-camera tool reporting successful calibration even when the reprojection error is >0.5. Is this a new feature? Im on SDK-0.8.
Also, do i need to skew the chessboard in various angles and orientations across the 4 images being collected? I could complete the calibration either by placing the chessboard flat facing the camera, or by skewing in various angles and orientations and both yields similar reprojection error. OpenCV seems to suggest the latter approach is necessary though.
-
Ah,
If you're talking about the calibration for the fisheye tracking lens the rules are different. It's much easier to calculate the intrinsics of the fisheye lenses since there's much more warp in the images, hence why there's only 4 images required for the calibration. A result of this is that the reprojection cutoff will be much higher because it's such a distorted image (we have found that <=5 works for the fisheye).
If you run the calibrator on one of the pinhole stereo cameras you'll see a lot more of the target rectangles, some of which have been intentionally elongated as to force the user to include some skew in the input set. This isn't necessary for the fisheye though.
-
@Alex-Gardner said in Acceptable tracking camera calibration reprojection error:
Ah,
If you're talking about the calibration for the fisheye tracking lens the rules are different. It's much easier to calculate the intrinsics of the fisheye lenses since there's much more warp in the images, hence why there's only 4 images required for the calibration. A result of this is that the reprojection cutoff will be much higher because it's such a distorted image (we have found that <=5 works for the fisheye).
If you run the calibrator on one of the pinhole stereo cameras you'll see a lot more of the target rectangles, some of which have been intentionally elongated as to force the user to include some skew in the input set. This isn't necessary for the fisheye though.
Do you know if the reprojection error is a good and reliable metric to determine if the calibrationis done well?
I have on rare occasions calibrated the fisheye cam, obtained reproj errors of less <0.3, but the calibration was obviously off. The aircraft was very unstable at takeoff, hovering and pure yaw rotations. The instability went away after i performed a recalibration (even if the reproj error ismore than what i had before)
I could confirm this because i kept the old intrinsics fileand i can reproduce the instability simpyly by swapping back to the old file