koide3
Hi @HSarham , You can include ```gicp/impl/*_impl.hpp``` to use the point types you want to use. You can also add them to ```fast_gicp.cpp```. https://github.com/SMRT-AIST/fast_gicp/blob/master/src/fast_gicp/gicp/fast_gicp.cpp
Hi @aiai84 , You can feed any type of points to FastGICP, but it doesn't use any information other than XYZ coordinates.
Thanks for reporting the issue. I'll check if voxelization has some memory problems soon.
Yep, it must be good to support more NeighborSearchMethod options in the CUDA version. I'm quite busy with writing some papers, but I try to find spare time to update...
I updated VGICP_CUDA so it supports DIRECT7 and DIRECT27. I also added a new neighbor voxel search method DIRECT_RADIUS which uses voxels within a radius so the GPU version can...
Hi @kisung-kim , Maybe the input point cloud was too large, and it ran out of the GPU memory. What GPU are you using and how much memory does it...
Ah, you are right. It can access invalid memory regions when coord == (0, 0, 0). I didn't notice this bug because I usually filter out very close points. I'll...
I just pushed a fix, and it will be merged into the main branch soon. https://github.com/SMRT-AIST/fast_gicp/pull/97
Thanks for letting me know this issue and providing nice benchmark results. When I wrote the code, I noticed that ```pcl::Registration``` creates kdtree inside, but I thought its construction cost...
You can use this package as a ROS-independent library. It would be good to link it with your program as a non-ROS library.