Absolute Trajectory Error / Relative Pose Error

Two standard metrics for how far a SLAM or odometry system’s estimated trajectory deviates from ground truth.

ATE and RPE are the two most common metrics for evaluating localization algorithms like visual SLAM and visual odometry; the TUM RGB-D benchmark from the Technical University of Munich gave the widely used definitions and evaluation scripts. ATE (absolute trajectory error) first matches the estimated trajectory to the ground-truth trajectory by timestamp, then aligns the two as a whole with a single rigid-body transform (monocular SLAM also needs an extra scale estimate), and computes the root-mean-square error (RMSE, in meters) of the position differences at each moment, reflecting the trajectory’s overall global consistency. RPE, strictly “relative pose error,” compares relative motion over a fixed time or distance interval — for example, translational error in m/s and rotational error in deg/s — reflecting local drift, which suits evaluating odometry that has no loop closure. The popular open-source tool evo supports TUM, KITTI, EuRoC, and other formats, where the metric corresponding to ATE is called APE.

ExampleComparing a SLAM system’s output trajectory to ground truth with evo_ape: an ATE RMSE of 0.02 means the whole trajectory deviates by about 2 centimeters on average after alignment; running evo_rpe with a 1-second interval then shows how much drift accumulates every second.

Also called

ATE, RPE, APE, Absolute Pose Error

Related

Visual SLAM · Visual Odometry · Ground Truth · Trajectory · Loop Closure Detection

Sources

TUM RGB-D Dataset: Useful tools (ATE / RPE evaluation)
evo: Python package for the evaluation of odometry and SLAM