fast_gicp icon indicating copy to clipboard operation
fast_gicp copied to clipboard

Published 20 hours ago • verified publisher icon SMRT-AIST

Applicability to indoor scenes

Open aditya2592 opened this issue 5 years ago • 5 comments

Hi

I wanted to ask if you think the voxel based GICP proposed in this work can also work with indoor scenes where the task is to estimate object poses? I was trying it out for a couple of indoor scenes and while the GICP algorithm in the repo works fine, the VGICP is unable to converge to the same solution that GICP does in several cases.

aditya2592 avatar Mar 14 '20 20:03 aditya2592

Hi @aditya2592 , It would be possible, but you need to change the voxel resolution parameter. The default resolution is 1m, it is too large for indoor scenes. For indoor scenes, you can try 0.05 ~ 0.5m resolution depending on the situation.

koide3 avatar Mar 15 '20 01:03 koide3

Thanks for the tip. Any other parameters do you think that need to be changed? I also noticed that corr_dist_threshold_ parameter is not being used in fast_vgicp_impl.hpp. Without the threshold, how do we determine valid correspondences?

aditya2592 avatar Mar 15 '20 17:03 aditya2592

Because VGICP takes the voxelization approach that naturally cut off distant correspondences, it doesn't have corr_dist_threshold param. You can consider that the voxel resolution is something equivalent to that param.

koide3 avatar Mar 16 '20 04:03 koide3

Thanks, I noticed significant improvement after changing the search method to DIRECT7 in addition to reducing the voxel resolution, and also changing the regularization method to FROBENIUS

aditya2592 avatar Mar 17 '20 20:03 aditya2592

Hi Is your task to use it to realize location?

LM getting stuck between two poses，scan matching has not converged!

git-chenjiqing avatar Dec 21 '21 09:12 git-chenjiqing