Question for the expansion to graph-based VIO system (like VINS-Fusion)

[1216621137]

Thank you for your excellent work. I have read your related papers published in IROS and RAL, and I have a question:

Can this method be applied to a VIO system (like VINS-Fusion) where an additional edge, derived from the IMU preintegration constraint, connects two successive keyframe nodes? This would mean that there is more than one edge between two nodes.

I look forward to your response.

[JulioPlaced]

Hello,

The inclusion of more edges into the formulation is something to be investigated --as is the case with landmarks. One could condense the edges into one, compute a joint uncertainty, or simply consider them separately. The latter may be the less efficient, but quicker to test.

Keep me updated if you try it!

[1216621137]

Thank you for your response.
I am confused regarding the time of computing D-opt(L_w) and D-opt(FIM), which is O(n^3)+O(l^3) and O(n^3*l^3), respectively, presented in "Fast Autonomous Robotic Exploration Using the Underlying Graph
Structure". If it’s not too much trouble, could you kindly help clarify this point? I would greatly appreciate your guidance.

[JulioPlaced]

Sure. Computing optimality criteria over the full Hessian matrix involves operations over a nl * nl matrix, while leveraging the graph topology allows to decouple computations into a n * n and a l * l matrix. Also, the determinant is the most demanding criteria, which is typically O(m^3), with m the matrix dimension.

Happy to further discuss this kind of issues via mail.