Some questions about the evaluation indicators

[SakuraMemoryKnight]

When I run evo_ape, I can get:

Aligning using Umeyama's method... (with scale correction)
Rotation of alignment:
[[ 0.99966481  0.01007943  0.02384699]
 [-0.0095949   0.99974696 -0.02034599]
 [-0.02404603  0.02011036  0.99950856]]
Translation of alignment:
[-0.48982277  0.97985428  4.66481522]
Scale correction: 16.0559813338
--------------------------------------------------------------------------------
Compared 1511 absolute pose pairs.
Calculating APE for translation part pose relation...
--------------------------------------------------------------------------------
APE w.r.t. translation part (m)
(with Sim(3) Umeyama alignment)

       max  11.813678
      mean  5.250972
    median  4.376774
       min  0.109121
      rmse  6.126788
       sse  56719.204065
       std  3.156710

I want to know what do the Rotation of alignment and the Translation of alignment mean? Besides,there is only ape translation part ,how do I know the ape of rotation?

[MichaelGrupp]

what do the Rotation of alignment and the Translation of alignment mean?

They represent the 3D transformation that aligns the trajectory to the reference trajectory. (calculated with the Umeyama algorithm)

how do I know the ape of rotation?

There is a --pose_relation option. For rotation angle you can use --pose_relation angle_rad or --pose_relation angle_deg for degrees.

[SakuraMemoryKnight]

@MichaelGrupp Thank you very much! I get it. i want to ask you another question.

This is one of the outcomes of evo_res and what does it mean?Thank you!

[MichaelGrupp]

It is a box plot of the error values of all poses: https://en.wikipedia.org/wiki/Box_plot

[SakuraMemoryKnight]

@MichaelGrupp Thanks for your reply! Please forgive me for having a new problem. I read some of the papers on slam, and I found that in these papers the performance of slam was evaluated with the average relative translation （%） and rotation （deg/m）, It's just like the kitti list shows： How do evo get this result？Thank you！

[SakuraMemoryKnight]

@MichaelGrupp I'm sorry to bother you again. According to my understanding, both ape and rpe are comparing the errors between frames, and their units are m or deg. However, how did the average pose errors per meter or per 100 meters mentioned in the paper and kitti's official website be obtained?

[MichaelGrupp]

I'm pretty sure there is code & documentation for that from the KITTI dataset. You could configure evo to do something similar, but keep in mind that it is not the official KITTI tool.