Gnc

[lucacarlone]

Added Graduated Non-Convexity (GNC) as a nonlinear optimizer in GTSAM. This is an implementation of the GNC paper (https://arxiv.org/pdf/1909.08605.pdf), but it uses iterative solvers in the inner loop instead of non-minimal global solvers.

Outstanding tasks:

• reintroduce “using BaseOptimizer = BaseOptimizerParameters::OptimizerType;” to let the user choose the iterative solver for the inner loop.
• move GncOptimizer and its params to a proper class/file when ready.
• populate the TLS option (now the code only implements GM = Geman McClure).
• double-check that the 0.5 in front of the factors in gtsam does not require a modification to the weight update rule.

[dellaert]

I left some comments. What is the main templating issue you wanted me to look at ?

[lucacarlone]

@jingnanshi @dellaert @pantonante : I fixed the outstanding issue with optimizerType. I also added a couple of unit tests (and comments) to make sure the inlier threshold is set correctly. @jingnanshi : the only thing I'm not convinced about is the stopping condition for TLS - please check my comment. Besides that, this PR is ready for review!

[dellaert]

Awesome! First round, will look again after comments resolved :-)

[dellaert]

Very cool!
Just a few more code nits, feel free to merge after addressed ;-)

Now we need to apply it to Shonan and compare with robust rotation averaging, right @jingwuOUO :-) ?

[lucacarlone]

all set! I'll merge after @pantonante and @jingnanshi approve!

[jingnanshi]

👍