Convergence comparison between NDT_OMP, GICP and VGICP about fast_gicp HOT 2 CLOSED

Hi, @koide3 , Thanks for sharing your great work to community, I have used these methods in LiDAR odometry and LiDAR loop closure detection modules. These methods are pretty well in LiDAR odometry especially for Fast-VGICP, but when I used these methods in loop closure detection to calculate a initial transform form current LiDAR frame to candidate LiDAR map, GICP and VGICP output wrong results, and even not converged sometimes while NDT-omp can work normally. Have you compared the convergences between these methods, and do you have any ideas to improve convergences?

• Init source(red) and target(green) pointcloud
  
• gicp alignment results
  
• NDT alignment results
  
• FAST-VGICP is not converged.

I'm interested in it. Do you know why ?

from fast_gicp.

ICP-related methods are sensitive to initial parameters, I solved the problem by employing a feature-based registration method (TEASER++) to provide the initial transform, but it may take more time.

from fast_gicp.