[GLIM] NTU VIRAL (sbs_03)

(Coming soon) github.com/koide3/glim
/ k_koide3

[@Noahitis]

this is sick

[@shivavarunadicherla]

Great work. I'm guessing you are fusing the results of IMU's estimate of position and SFM data from camera. How much drift are you getting as of now?

[@kenjikoide6076]

It fuses all Visual-LiDAR-IMU constraints on a unified factor graph. Because it was an easy setup for the LiDAR, it got almost no drift at all.

[@shivavarunadicherla]

​@@kenjikoide6076 If LiDAR's also participating in Motion estimation, it sure would be very accurate. I would bet that if you used only the data from LiDAR you would probably get the same estimated motion. I was thinking that LiDAR was used as ground truth.

[@kenjikoide6076]

@@shivavarunadicherla Yes, visual constraints brought only minor accuracy gain in this dataset indeed. This is just a demonstration. What we are truly aiming for is overcoming situations where point clouds become completely degenerate (e.g., tunnels), and we've confirmed that visual constraints greatly improves the reliability in such situations.