Question about Factor Graph Subgraph Implementation

[JokerJohn]

Hello, Dr.Shan, Thanks for your work!

I noticed in your code for the factor graph subgraph, you optimize the IMU bias using 100 LIDAR poses, and then use the new bias for prediction. When implementing this feature, every time you collect 100 LIDAR poses, you marginalize the latest LIDAR pose and use it as a prior, resetting the factor graph. Resetting the factor graph also means you need to collect another 100 LIDAR poses to meet the sliding window size of 100 again. Repeating this process each time seems a bit inconsistent with the traditional sliding window approach. May I ask the reason for this design choice?

Why not directly use the FixlegIncrementalSmoother class provided in the GTSAM unstable module? I'm guessing that while the latter can use ISAM for automatic marginalization, it's challenging to manually remove old state variables, leading to an ever-growing graph size. Is that the case? Is my assumption correct? If not, please correct me. I'd like to know your initial thoughts on this issue.

Thank you for your time and clarification.