Use Huber loss for Shonan

[travisdriver]

[johnwlambert]

Hi Travis, thanks for the PR. Just wanted to make sure you've seen this other closed PR: https://github.com/borglab/gtsfm/pull/344

[dellaert]

Don't understand actually why we closed #344 without merging. I think we could have just merged it with a config flag set to OFF? Shonan is really bad at handling outliers so in theory it should be ON by default, but maybe tuned such that it does not negatively affect datasets without outlier pairwise edges.

[johnwlambert]

Don't understand actually why we closed #344 without merging. I think we could have just merged it with a config flag set to OFF? Shonan is really bad at handling outliers so in theory it should be ON by default, but maybe tuned such that it does not negatively affect datasets without outlier pairwise edges.

@dellaert The main reason I closed #344 was that robust Shonan didn't appear to give any improvement on any dataset over standard Shonan. In my experiences, the results were always inferior or identical

[ayushbaid]

Do we still want to merge this?

[travisdriver]

Do we still want to merge this?

I think having the option of using the robust optimization is useful. Even though it does not currently show any improvements, it may just be that the hyperparameters are not tuned properly.