Thank you for sharing. Can you give the detailed derivation process of Jacobi?

[BuaaYfl]

          Yes, it is an approximated way; the covariances of source points are fixed at the linearization point and do not change during cost evaluation. I implemented an exact version with covariance drivatives, but it didn't improve accuracy nor stability while making optimization significantly slow. If you want, you can try it at the following link:

https://github.com/SMRT-AIST/fast_gicp/blob/87cd6288d14bd155e8b7a2144f68bb5246aecc52/include/fast_gicp/gicp/gicp_loss.hpp

Originally posted by @koide3 in #80 (comment)