Weird calibration alignment (cone instead of cylinder)

[Elhanan Ben Yishay]

Hi Andrew,

I just did a re-calibration of the arena after changing the position of some cameras - I am using a circular, cylindrical arena (about 50d x 30h), and I get really good re-projection errors from the MultiCamSelfCal (<0.5). However, after doing the alignment to the arena, the traces appear to form a cone shape as I moved the LED from the bottom to the top of the arena (see attached photo). I would guess it's something with the intrinsic calibration of each camera rather than the extrinsic, but I am not sure.

Let me know what you think may be the issue, and thanks for all the help!

Best,

Elhanan

[Andrew Straw]

That is strange. I do not have any immediate insights except that with such low reprojection errors it is unlikely that there is a problem with the intrinsic calibration unless these low reprojection error numbers are only from the center of your cameras' FOV whereas perhaps the edges have more error?

There are configurations where the MCSC algorithm just fails, but that results in huge repojection errors.

I am working also on a simplified april-tag based calibration system which is more direct that then MCSC approach and will try to clean up my jupyter notebook to be self explanatory and share with the group here. This will allow calibration without the MCSC algorithm. In the meantime, you could think about how to make april tags visible to your cameras and 3D locations that will define your coordinate system.

[Elhanan Ben Yishay]

Hi Andrew,

So I've done some testing and I think the issue may be when using the internal strand-cam calibration procedure vs the ROS calibrator? (I often use the ROS camera calibrator app for the high resolution cameras, because it is much faster; otherwise I usually get errors about channel full with strand-cam). After re calibrating all cameras within braidz, there is still some distortion (this cone-like shape), but it's better than before. I can maybe try to calibrate all the regular cameras with ROS, and then the high-resolution cameras with braid, and see if there's any difference.

Is there any chance there's a difference in the calibration output or procedure between braid and ros?

Thanks

Elhanan

[Andrew Straw]

Clearly the calibrations are different with the newer one being better. I did try to exactly implement how the ROS camera calibration package worked. Your statement about the speed, though, makes me think this must be wrong. Otherwise I would guess the difference is in the actual images used rather than the algorithm. The UI in ROS is better at providing feedback and perhaps you are use that to guide your calibration better.

I hope to have a bit of time in the next few days to work on the April Tag fiducial markers to simplify the calibration procedure.

--
You received this message because you are subscribed to the Google Groups "multi-camera software from the Straw Lab" group.
To unsubscribe from this group and stop receiving emails from it, send an email to multicams+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/multicams/f2e5be1c-9009-43e2-aae4-c5b721438610n%40googlegroups.com.

[Andrew Straw]

Hi,

The initial draft of a tutorial which uses April Tags in known, predefined locations is here: https://github.com/strawlab/dlt-april-cal/blob/main/tutorial.ipynb .

There are a couple of TODOs at the bottom of this page which will likely be important for you but what is here will hopefully be enough for you to see the big picture and start working on this approach if it seems useful for you.

Contributions, clarifications, questions on this tutorial are gratefully accepted. I think making this process smooth and well documented will greatly facilitate 3D tracking and the ideas are not specific to any tracking software.

Best,

Andrew

[Andrew Straw]

I just released Braid 0.10.1 which supports using PyMVG calibration format as inputs to Braid. I actually thought this was already enabled but after writing this tutorial, realized that it wasn't working and I fixed that.