smauq
commented
3 years ago @patripfr so I should have solved now all the issues we talked about (parallelizing removal of tracks and handling -1 track ids). If you could check my code and I think there's a few last points:
- Some simple unit test maybe
- Initialize landmarks and triangulate only for affected vertices/landmarks (this one you also have to check if it makes sense from a speed perspective)
- If you could for curiosity run a trajectory quality evaluation on EuRoC between using and not using the new
elqit would be super cool to know if it improves trajectory quality at all
ethzasl-jenkins
Added a new command
remove_invalid_observations->rio, that solves the issues brought up by check_map_consistency, namely landmarks that due to loop closure are observed multiple times from the same frame (which is impossible). This command splits away the worse tracks and reinitializes them again as new landmarks.