forums.naturalpoint.com

Hello,

I've purchased 4 Flex C120 cameras for using them with the Point Cloud SDK. I disposed the cameras 2.20 meters above the floor, positioned at the 4 corners of a 4x4 square, and all of them pointing to the center. My marks are spherical balls of 45mm of diameter, and they can be seen fine by all cameras (as I checked in the preview of the calibration application).

The problem I have is with the accuracy of the results at the end. It seems that the calibration doesn't get properly the relative position of the cameras, since they look in the viewport not to form a square. Then, the results after that are not so much good, even I tried to override the default settings on the Ground Plane window.

I've tried to calibrate with fluorescent light and with no light at all, with and without synchronization, with different speed movement during the wand step, and with several configurations for exposure/threshold/ilumination. I found that too much exposure is not good for the results, and it also works better when moving the wand slowly. However, there is still to much error in the results.

What accuracy should I be able to get at that distance (let's say 3.5 meter with marks of 45 mm)? There is something I can do to improve the results?

Thanks in advance.

PS. I've tried also with the new 1.1.032 beta 02, but the results are the same.


EDIT:

I've just found that rotating 2 of the cameras 90 degrees the accuracy turns to be really good.

Hello:

Interesting note that rotating 2 of the cameras helped the calibration, thanks for that info.

Could you tell us the error numbers at the end of the calibration routine? What they were when it was "bad" and now that it is improved?

Overall, your application sounds like it should work well.

Sure, I'll post the errors. Just wait a little, since I have to finish soon a task (for next week). When I finish I will recalibrate the cameras as they are now and in the before state, and I'll let you know about the results.

OK, sounds good.

Hello again. I've been testing more about calibrating, and I found that the problem was not in the camera position, but there was another point. Using "Very Fast" or "Fast" accuracy made the calibration results better in most cases, since the "Most accurate" one usually made the calibration to have more error. I made another guy to perform the calibration and he arrived to the same conclussion. Using the most accurate version we needed sometimes several hours to get an acceptable calibration, but the "Very fast" option gave us quickly (just try one or twice) acceptable results. Might be the iterative process running there was not converging.

Another thing we tried, and we didn't got good results (but would improve the capabilities of the calibration tool), is to calibrate cameras that do not share captured volume. We tried with six cameras, and 2 of them almost didn't have overlapping captured volume. The calibration was not possible to be done.
So my question is if the program is built in a way that 'transitive calibration' can be performed (camera A and B overlap, and so B and C, but not A and C, but the system can place all of them properly). That would make feasible a wider range of camera placements.

I hope this gives some good feedback and some improvement for the program for the future.

This merging of capture areas is something that we are looking into, but is currently not possible with the available software.

As for the calibration data, I will pass on the info, to thedevelopment team....

Moises, our calibration algorithm does favor tend to favor all cameras pointing into a single volume. You can point cameras in a non-overlapping arrangement as long as there is significant overlap. As cameras have less overlap the correlation between the wanding data is reduced and our calibration algorithm has a harder time converging on an acceptable solution. We are however working to improve this situation.