Ask Your Question

inferiority of circular marker for stereo-calibration

asked 2013-09-17 06:05:24 -0500

Arne gravatar image

Hello, i was evaluating different sets of parameters during the stereo-calibration. It turned out that only 25 calibrations with circular markers were under the top 1000 of about 6720 calibrations ordered by rms. Does this have anything to do with a lack of functions like cornersubpix for circular markers. Because according to some literature i read about photogrammetry, the circular markers should be much better to detect while having a better precision. I hope you can explain this issue to me, and tell me if i am right with my guess.

kind regards

edit retag flag offensive close merge delete


I doubt anyone will explain this, it's almost a research question. Yes circular markers should be better especially if combined with other methods, which unfortunately are not implemented in opencv. See e.g. paper by Datta, Kim, Kanade.

Alex gravatar imageAlex ( 2013-09-17 06:46:22 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2013-09-17 06:54:40 -0500

jensenb gravatar image

The OpenCV circle grid detector implementation uses subpixel precision, so I doubt that would be the reason for any quality difference when compared to the checkerboard detection implementation. Without knowing anything further about your experiment setup, its impossible to say why you got such results. In my experience I get the exact opposite result, circular board patterns give me a good 50% lower RMS value.

As an aside not, I would be careful about using the RMS score to judge the quality of your calibration. Camera calibration and especially stereo calibration is highly sensitive to the quality of your input and because of this it is critical that you take calibration images with high variety of poses, and for stereo that image pairs are correctly synchronized. The camera calibration implementation in ROS has some good metrics for evaluating the quality of the input calibration images, check it out here.

edit flag offensive delete link more


First i want to thank you for your quick answer. For testing i am using an extended version of the sample for stereo-calibration. I just recorded some images as advised in the ROS Wiki. And even if its not perfect i was able to make both sets look equal. (chessboard and circular marker) The strange thing is now, that even if not in all of the pictures the chessboard is found i will get a calibration. But when using the circular marker, the pattern in found in all images, i get NAN as rms and the rectified Image is gray. How is that possible while all patterns are found in almost the same poses as the chessboard, which is not found in all images. I am really confused.

Arne gravatar imageArne ( 2013-09-17 08:47:57 -0500 )edit

Either they haven't introduced this functionality yet or you need to switch the row and column lengths for your pattern.

blueVision gravatar imageblueVision ( 2013-09-18 01:38:28 -0500 )edit

I did change the row and column, otherwise the findCirclesGrid function won't find anything.

Arne gravatar imageArne ( 2013-09-18 03:09:53 -0500 )edit

Question Tools

1 follower


Asked: 2013-09-17 06:05:24 -0500

Seen: 752 times

Last updated: Sep 17 '13