Covariance R_ymj

76 views

Skip to first unread message

Henrik

unread,

Jun 20, 2016, 7:37:55 AM6/20/16

to kalibr-users

Hi!

I have a question how the covariance R_ymj of the 2D-3D correspondences are set in equation 11 in your article https://3234f89137bccf2ede29cc86e315c75116020d70.googledrive.com/host/0B64GJ60h3Ai1MVVwWTZwekhtcFU/publications/bib/furgale_iros13.pdf?

And do you know where in the code it is calculated?

Thanks in advance
Henrik

joern

unread,

Jun 30, 2016, 4:26:31 AM6/30/16

to kalibr-users

Hi Henrik,

this covariance characterizes the distribution of the corner detections. While it is set to a default value in the standard toolbox, you are able to fiddle with the value in the updated version. Please see here for details: https://github.com/ethz-asl/kalibr/wiki/Multi-IMU-and-IMU-intrinsic-calibration#3-running-the-calibration where it is called "--reprojection-sigma".

Since the intrinsic calibration uses the same detector and you may use a similar setup in terms of target, camera configuration and lighting, the uncertainty returned by the intrinsic camera calibration can serve as a starting point for this value. However, your camera/IMU calibration dataset may exhibit significantly more motion which in turn may affect the distribution due to motion blur.

Nevertheless, I would advice you to start from these values and tune them if appropriate to (somewhat) account for motion blur.