I am new to metrics used to assess pose estimations. I am trying to implement evaluations for this paper
The paper provides some details of how trajectories are normalized and error is calculated.
" Metrics used are Absolute Translation Error(ATE), Relative Translation Error(RTE) and Relative Rotation Error(RRE). Since camera tracks in the
Sintel dataset have very different lengths (from less than 1 meter to larger than
100 meters), simply averaging over all sequences introduces bias towards long trajectories. Therefore, before calculating the metrics, ground truth trajectories
are normalized to unit length. For all methods, we align the predicted results to
the normalized ground truth tracks using Umeyama alignment with scale
calibration."
It sounds like they normalized the trajectories such that the length of it becomes 1. However, I am not sure how to implement this. Does anyone know how do I go about normalizing the ground truth trajectories?
I am new to metrics used to assess pose estimations. I am trying to implement evaluations for this paper
The paper provides some details of how trajectories are normalized and error is calculated.
" Metrics used are Absolute Translation Error(ATE), Relative Translation Error(RTE) and Relative Rotation Error(RRE). Since camera tracks in the Sintel dataset have very different lengths (from less than 1 meter to larger than 100 meters), simply averaging over all sequences introduces bias towards long trajectories. Therefore, before calculating the metrics, ground truth trajectories are normalized to unit length. For all methods, we align the predicted results to the normalized ground truth tracks using Umeyama alignment with scale calibration."
It sounds like they normalized the trajectories such that the length of it becomes 1. However, I am not sure how to implement this. Does anyone know how do I go about normalizing the ground truth trajectories?