Calibration precision

[Victor Lamoine]

Hello,

I watched Chris Lewis conference (really great video!) and I have several questions:

• At 22:00, how do you tell your calibration is good?, is there any news on that topic?
• What kind of accuracy did you get when calibrating the Asus camera, sub-milimeter, 5 millimeters, 20? (a rough idea is still better than no info at all)
• Would there be an advantage of calibrating the device using the generated point cloud and not the images? Constructing the target frame over the corner of a cube for example.

I will trying to use this package with the David SLS-2 sensor that I recently integrated into PCL. I'll make a tutorial about that if I succeed.

Bye

[Chris Lewis]

@VictorLamoine @shaun-edwards
The accuracy of calibration is determined by several things. The best I can do for you is this:

1. Make sure your residual error is low. In Ceres-solver terminology the is the SSE or sum of squared error. I usually take the square root and divide by the number of observations to get a per-pixel error.
2. If #1 is true, then the co-variance provides an order of magnitude measure of accuracy.

As for the Asus camera calibration accuracy, and assuming you mean extrinsic calibration, my order of magnitude estimate is 5mm. This accuracy depends heavily on the observation geometry. for example, a larger target provides more accuracy.

Now for the bad news. The intrinsic calibration you are probably using is in my opinion flawed. Matlab and Opencv use the exact same cost function and distortion model, so it does not matter which you use. They are flawed because they did not check #2 from above. If they had, they would have observed that distortion parameters are correlated to focal length and center point. This explains why the results from intrinsic calibration have such wide variance. For example, the center point (cx,cy) is directly related to residual error. A small change in center point location is equivalent to shifting all the pixels over, and therefore shows up as a corresponding increase in residual error of about the same size. However residual error is typically .25 pixels, while the variance on cx,cy is about 6-10 pixels when using the standard algorithm. I think I will have an alternate intrinsic calibration procedure soon that should fix this issue. For now, place a cal-target 1 meter from your Asus, calculate its location using the target_finder node, and make sure the distance returned using your intrinsic values is 1meter. Note, the focal point of your camera is someplace behind the lens, and not accurately known, so +-1/4 or 6mm.

Lastly, it is a bad idea to use the point cloud to calibrate. This point cloud is a product of intrinsic and stereo calibration. One of the pull requests also has the beginnings of a automated point cloud calibration process. We should have a video in the coming months explaining both intrinsic and point cloud calibration.

[Chris Lewis]

@xiqiaosang
What problems are you having?

[G.A. vd. Hoorn]

@xiqiaosang: could you plase open a new issue for your problems/questions? You can just copy/paste what you wrote in your comment here. Afterwards, remove your comments from this issue.

This thread is really only meant for @VictorLamoine's issue.

[xiqiaosang]

@gavanderhoorn Sure, it is a good suggestion. :)