Not performing well when duplicate points appear. about fast_gicp

[koide3]

I suppose the cause is the number of points rather than duplicate points. We employed a GaussNewton solver that is fast but sometimes a bit unstable with respect to a large number of points. I applied an approx voxel grid filter (gridsize=0.1m) to your point clouds, then the algorithm worked. We empirically found that the algorithm can be unstable when # of points > 15k.

[koide3]

Hi @Hezip ,
Thank you for reporting the issue. Let me confirm that you replaced kdtree.nearestKSearch with octree.approxNearestSearch, then GICP_MT and VGICP didn't work well, right? Can you share your modified code so that I can test it?

[Hezip]

No code changes.
One of the point cloud I provided was obtained through octree.approxNearestSearch, and there are duplicate points in it.

[Hezip]

Thank you very much for your answer. After I remove the duplicate points, the effect is better, but I think you are right, it should be related to the GaussNewton solver.