Improving data association

[akshay-krishnan]

Data association is the association of features into tracks before bundle adjustment. Currently this is done with Union-find.

We currently observe the following problems with data association:

1. Too many small tracks
2. Wrong correspondences may cause two different tracks to be merged into one track.
3. Once a track is triangulated, if the mean reprojection error is high, all observations are discarded.
4. The reprojection error threshold used to make this work is too high.
5. We don't refine tracking with better poses after bundle adjustment.

Creating a high level issue to address these problems. Feel free to add to the list any discrepancies you observe and add yourself to assignees when contributing. :)

[akshay-krishnan]

Initial directions to investigate:

• Force tracks to have minimum length 3
• Remove only outliers from the track after triangulation.
• Try creating more tracks from outliers and merging them after bundle adjustment.

[travisdriver]

#281 to address 3.

[travisdriver]

• [ ] Use RANSAC by default in triangulation
• [ ] Reduce reprojection error threshold <100 in triangulation