ICP taking long time to converge

asked 2011-11-22 10:50:22 -0500

2831 ●55 ●68 ●96

updated 2014-01-28 17:10:50 -0500

ngrennan
1 ●1 ●1

Hi, I took two point clouds of 640x480 size separated by very small distance and ran ICP. The code is running for long time and i had to exit it manually. I guess its like a gradient decent method of finding the transformation. But is there a way to set some parameters to it. How does the RGBD-slam's ICP run so fast? (Is there a package for Generalized ICP available in pcl?) Should i downsample the points before i run ICP? Pls share the information that you know.

Thanks, Karthik

edit retag flag offensive close merge delete

add a comment

0

answered 2011-11-27 23:33:06 -0500

Felix Endres

6468 ●28 ●57 ●93 http://www.informatik....

While the functionality is still there, (G)ICP is actually not used anymore in RGBDSLAM as a colleague and me couldn't get it to improve the results from feature correspondences with reasonable effort (i would appreciate any help here, it should be possible somehow).

But when we used it, we downsampled the point clouds considerably to speed it up. Curiously icp was faster writing the points to a file, running the standalone gicp executable and parsing the result, than calling the compiled-in library. We have found no explanation for this.

We took the implementation from Stanford