Closed stefangachter closed 8 months ago
It doesn't matter because the final value of the metrics always involves a norm (see the part below the formula you referenced).
That's also the reason why you can swap the two input trajectories and still get the same result.
I agree with you and I am aware of this fact. However, if you would like to go a step beyond and compare the error with the estimated uncertainties, then it is highly relevant how the error is computed. This is not (yet) part of evo but would be a desirable feature. In my view, any researcher doing estimation should go beyond and have a look at the uncertainties as well. Thus, the documentation and code should be consistent such that anybody using the outcome of evo and would like to go "beyond" is guided "correctly".
Changed the order in the notebook now: https://github.com/MichaelGrupp/evo/commit/47c96ae68fab14bdaea059776fff31a86ccf0305
Thanks!
Hi Michael, I am a little bit puzzled but came across the following: The documentation here https://github.com/MichaelGrupp/evo/blob/master/notebooks/metrics.py_API_Documentation.ipynb states:
That is, the difference between the estimated and reference pose is the inverse reference pose times the estimated pose. This definition results in an error in the local frame. However, if I dig into the code, then the error in the global frame is computed:
If I am not mistaken, then
x_t := x_est_t
andx_t_star := x_ref_t
Again, if I am not mistaken, then
p1 := x_est_t
andp2 := x_ref_t
Thus
x_est_t^{-1} * x_ref_t
but the documentation states the opposite. Hopefully, I didn't overlook something.