how to understand calibration result

[ghost]

Hi, I tried twice for this calibration toolbox. I got two different answers for the same camera-velodyne setup. Our velodyne is located ~5cm above the camera, and the camera is pointing 15 degree downwards. Here is the result we got: ​Second test: After 100 iterations

Average translation is: -0.220729 -0.273677 4.82629 Average rotation is: 0.905049 0.419485 0.0701358 0.40611 -0.901351 0.15047 0.126337 -0.107699 -0.986124 Average transformation is: 0.905049 0.419485 0.0701358 -0.220729 0.40611 -0.901351 0.15047 -0.273677 0.126337 -0.107699 -0.986124 4.82629 0 0 0 1 Average RMSE is: 0.325415 RMSE on average transformation is: 0.331159

First test: After 100 iterations

Average translation is: 0.422593 0.831942 3.90745 Average rotation is: 0.900951 0.36973 -0.227127 0.22701 -0.847701 -0.479447 -0.369801 0.380398 -0.84767 Average transformation is: 0.900951 0.36973 -0.227127 0.422593 0.22701 -0.847701 -0.479447 0.831942 -0.369801 0.380398 -0.84767 3.90745 0 0 0 1 Average RMSE is: 0.346004 RMSE on average transformation is: 0.353328

Is there any documentation on how to read this result? I visualized rotation matrix with ROS tf, but neither of them seems correct. Do you have any suggestions what I should do to get good results? Will this toolbox work for sensor setup like ours? Btw, our camera is 100degree FoV and Velodyne is 16 beams. Really appreciate your help!

[chinitaberrio]

I don't know, sorry

[ankitdhall]

The results can be read as estimates that have been computed over multiple iterations. An average of the translation and rotation are computed over the N iterations. The translation is a 3x1 vector with x, y, z components and the rotation values are in the form of a 3x3 matrix.

The average transformation is just the stacked matrix [R t; 0 0 0 1] (4x4 matrix). The RMSE is the root mean squared error between the 3D points viewed from the camera and laser scanner after applying the transformation.

lidar_camera_calibration was specifically made for a setup that had (multiple) cameras (both stereo and monocular) and a laser scanner with 16 scan-lines. You can have a look at the fusion results of point clouds (available in the README) from stereo cameras after calibrating using this method.

There are a couple of things you should have a look at to make sure your setup is correct (these are also mentioned in the README): [1] Check the values you have entered in the config_file.txt [2] marker_coordinates.txt has the correct measurements of the board on which the ArUco marker is stuck. [3] marker size in find_transform.launch is correctly entered [4] ensure that the intrinsic calibration of the camera is of high quality; better calibration will give more accurate estimates
You could probably re-read the README to make sure that you haven't missed anything. Also you can have a look at #13
For more technical details you can refer to this.