Why use OptimizerExt?

[LauZanMo]

Thanks for your works! I noticed that you rewrite gtsam optimizers in XXOptimizerExt, and the factors seems to not work with the native gtsam optimizer. Can you tell me the reasons why you rewrite these optimizers, and what are the benefits to use OptimizerExt?

[koide3]

There are two reasons:

1. The LM optimizer in GTSAM (gtsam::LevenbergMarquardtOptimizer) terminates the optimization iteration when the nonlinear error gets increased. But, data association changes can increase the error in point cloud registration, and the default LM optimizer wrongly terminates the optimization in that case. The LM"Ext" optimizer prevents the wrong termination with slightly modified termination criteria.
2. The "Ext" optimizers (LM and iSAM2) implements a custom batched linearization routine for GPU-based factors.

Note that you can still use factors in gtsam_points with the optimizers in GTSAM, but you may face the above early termination issue and degraded processing speed of GPU-based factors.