Understanding the sustenance of slam points

[tejalbarnwal]

Hi,

First of all, thank you for such a clean and well-structured implementation of the filter-based VIO system.

am relatively new to the VIO community and have been trying to better understand some aspects of the frontend. I noticed that my runs work really well as long as I have SLAM points available. However, once these points drop out, performance degrades. This suggests that I need to improve the persistence of SLAM points. I don’t mind trading off a bit of computational efficiency if it helps sustain the SLAM point set for longer.

I have a few questions regarding this:

What’s a good way to measure whether SLAM points persist longer across frames?

How do different feature detectors (e.g., Shi-Tomasi, FAST, ORB, etc.) affect SLAM point persistence?

What metrics would you recommend to evaluate and compare different frontend implementations?

I was initially thinking about using the size of the SLAM point cloud as a metric, but I’m not sure if that’s the most informative measure.

Would love to hear your thoughts or suggestions on this! Am aiming for more robustness.
Would appreciate any insights, especially from @goldbattle if you have time.

Thanks!

[tejalbarnwal]

Hi @goldbattle,

I’m also experimenting with different frontend configurations. Is there a recommended way to decouple or isolate the backend testing from the frontend so that I can better diagnose where the issues are coming from?