Acceleration of ICP via extrapolation strategies

[VC86]

Hello,

I have noticed since quite some time that in the seminal paper [1] describing ICP, there is a section related to accelerating convergence using a line search extrapolation that essentially increases the updates' length if certain criteria are met.

In the ICP implementation of Open3D [2] this is also not done, to the best of my understanding. This is also the case for many implementations of ICP on GitHub [3] and elsewhere, too.

I was wondering if anyone here had any experience or recommendation on this feature; I have tried implementing it on a toy ICP implementation, but I presently did not measure the rather big improvements that other authors seem to have measured [4] (Simon claims that coupled acceleration in rotation - translation domain already reduces to 25% the number of iterations of vanilla ICP...).

[1]: P. J. Besl and N. D. McKay, “A method for registration of 3-D shapes,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 14, no. 2, pp. 239–256, Feb. 1992, doi: 10.1109/34.121791.
[2]: https://github.com/isl-org/Open3D/blob/master/cpp/open3d/pipelines/registration/Registration.cpp
[3]: https://github.com/symao/libicp
[4]: D. A. Simon, “Fast and Accurate Shape-Based Registration,” 1996.