RPE (rotation) results of object motions are computed as RMSE?

[Javier Tirado Garín]

Thanks a lot for open-sourcing this great work :)

I've come up with a doubt regarding the computation used in the paper for the Relative Pose Error of an object motion:

By reading the paper, at first, I thought that this RPE it's computed as the RMSE of the relative angles between the estimations and the ground-truth.
But know I'm doubting if my understanding is right, because looking at this part of the code:

VDO_SLAM/src/Tracking.cc

Lines 989 to 997 in 2b13517

 float trace_rpe = 0; 
 for (int i = 0; i < 3; ++i) 
 { 
     if (RePoEr.at<float>(i,i)>1.0) 
          trace_rpe = trace_rpe + 1.0-(RePoEr.at<float>(i,i)-1.0); 
     else 
         trace_rpe = trace_rpe + RePoEr.at<float>(i,i); 
 } 
 float r_rpe = acos( ( trace_rpe -1.0 )/2.0 )*180.0/3.1415926; 

I think it's being computed directly as the relative angle.

So, is this last approach being used for the results reported in the paper?

Thanks in advance!

[Javier Tirado Garín]

My bad, I've just realized this computation is used for one pose only,
my apologies.