seeding multicamselfcal with initial guess?

[Floris van Breugel]

Hi Andrew,

We are really struggling to calibrate out 14-camera array. I suspect part of the problem may be that the working volume is large enough that most cameras do not see most of the volume, but all points in the volume are seen by at least 3 and often more of the cameras. 

We have tried many configurations, and were once able to get a convergence, but it seems quite tricky, so I'm wondering if there is anything we can do to improve the convergence.

Main question:

I vaguely recall there may be a way to seed multicamselfcal with an initial guess of the camera positions and orientations. Is there indeed an option for that? Do you know where and how one would do that?

Other things:

1. We have done the checkerboard calibrations, the lenses do not introduce a lot of distortion

2. We ensure that all cameras have ~1000 surviving points, and coverage is relatively consistent across cameras

3. We checked the braidz file, and the data looks to be high quality

Thanks for any suggestions!

- Floris

[Andrew Straw]

Hi Floris,

Unfortunately, I don’t know of a “seeding” approach built in.

One issue we never solved with MCSC was that having cameras in a central position pointing outwards was a geometry that we could never get to work. I think the issues here https://doi.org/10.1109/TPAMI.2007.1132 may be relevant. (MCSC is far from state of the art in multicamera calibration these days. It would be great to move Braid etc to something more modern and robust.) We were able to change our approach but never fully understood the underlying issue.

I understand that the first step of MCSC is to build a large matrix which is then decomposed using SVD to the best linear model. RANSAC is run to eliminate outlier points and find the best model. This is then refined through an iterative optimizer ("bundle adjustment”). It would be interesting to do a sanity check on the initial linear model. I’m not sure how to do that. Also, the MCSC code has a lot of plotting which is disabled — I can’t remember permanently or automatically on detecting that it is being run in octave. You could re-enable the plots and see if that helps gain insight. If you delve into the MCSC code enough to find the linear model stage, perhaps you can swap in your initial seed parameters at that stage. In other words, skip the SVD on the new data, but insert the parameters from an initial guess and still use the bundle adjustment from MCSC. (Or alternatively, use an alternative implementation.)

As an alternative to MCSC for use with Braid, we have had success using April Tags at known 3D locations to solve camera calibrations using DLT: https://github.com/strawlab/dlt-april-cal/blob/main/tutorial.ipynb Presumably one could additionally run bundle adjustment on the results. I like that this approach is conceptually simpler and gives calibrations which do not have to be “aligned”.

I’d be curious to hear updates about your progress with this.

Best,

Andrew

--
You received this message because you are subscribed to the Google Groups "MultiCamSelfCal" group.
To unsubscribe from this group and stop receiving emails from it, send an email to multicamselfc...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/multicamselfcal/50c42797-b07f-4c0c-83ff-55b181fce76bn%40googlegroups.com.

[Floris van Breugel]

Thanks Andrew for the quick and detailed response!

We did end up getting our array to calibrate, and the key seems to be to have at least 3 cameras that see almost the entire tracking volume. I suspect that this helps constrain the "fill missing points" steps in a way that allows the algorithm to converge. I wonder if the same issue is what prevented arrangements with outward facing cameras from converging? 

If we end up trying any new approaches in the future I'll revisit my response..

- Floris

To view this discussion on the web visit https://groups.google.com/d/msgid/multicamselfcal/8B8148EE-4122-4A45-9298-A860A35287A0%40gmail.com.

--

Floris van Breugel   |   http://www.florisvanbreugel.com
Assistant Professor of Mechanical Engineering 

& Graduate Program for Neuroscience

University of Nevada, Reno

Wildlife and Landscape Photography

Galleries: http://www.ArtInNaturePhotography.com/
Blog: http://www.ArtInNaturePhotography.com/wordpress/