Add Essential Matrix Optimization

[travisdriver]

Add essential matrix optimization similar to what's presented in this paper: https://openaccess.thecvf.com/content_iccv_2015/papers/Sweeney_Optimizing_the_Viewing_ICCV_2015_paper.pdf.

This is what's also used by GLOMAP. Some things to think about:

• GLOMAP uses a Cauchy loss instead of Huber: https://github.com/colmap/glomap/blob/18a70eed0e36eae7c42e1c5ef0ae6a6f8cdc6943/glomap/estimators/view_graph_calibration.h#L25
• Image pairs are filtered based on essential matrix residuals: https://github.com/colmap/glomap/blob/18a70eed0e36eae7c42e1c5ef0ae6a6f8cdc6943/glomap/estimators/view_graph_calibration.cc#L172

[travisdriver]

Very promising results on South Building

Without E Optimization

With E Optimization

[travisdriver]

Better results across the board

visual_comparison_dashboard.zip

[akshay-krishnan]

among non-rc3 medium sized datasets, it seems clearly better for south building, but worse for gerrard hall and palace of fine arts. we should focus on these medium sized and other non-CI large datasets.