a problem about scale initialization

[hkust-aerial-robotics]

a problem about scale initialization

[qintonguav]

In this implementation code, we relax the initialization threshold for the better experience.

It's just an initial value to bootstrap the nonlinear system. We don't use it to calibrate the real scale all the time. The scale will be solved precisely in the nonlinear procedure.

In the paper, we use the strict threshold. Such as strict g limitation, IMU variance, and SfM cost...

[vonzy]

@qintony so if the threshold is relaxed, the scale could get wrong possibly (even negative? I wonder if my code get wrong?). could you tell me the strict threshold you used? thanks!!!

[vonzy]

@qintony moreover, the condition number of H is very huge, so the equation is easily influenced by any errors in b?

[qintonguav]

The equation is right.
If alpha and beta don't contain much excitation (such as near stable, zero acceleration), the equation is fully unobservable, the H will be ill-conditioned, and any result will be solved. Sometimes negatives happen. The preconditions are that, alpha, beta and SfM structure should be as accurate as possible.

Following are thresholds

accurate sfm

good gravity

enough IMU excitation

As for a general application, nobody wants to wait a long time to get a much accurate initial value. So we relax these thresholds, get a rough initial guess (much better than no initial guess).

[vonzy]

all right! thanks!