Mean Per-Joint Position Error

The average distance between predicted and ground-truth joint positions, measuring how accurate a pose estimate or motion tracking is.

MPJPE was originally the standard metric for 3D human pose estimation, widely used on datasets such as Human3.6M: for every frame, the Euclidean distance between each joint's predicted 3D position and its ground truth is computed, then averaged across all joints and frames, usually in millimeters, with lower being better. There are three common variants: root-relative MPJPE, which first aligns the root joint (the pelvis) and so measures only local pose; global MPJPE (g-MPJPE), computed with no alignment at all, which also captures overall positional drift; and PA-MPJPE, computed after a rigid-body (Procrustes) alignment. In humanoid robotics, motion-tracking work such as PHC and ASAP uses it to measure how precisely a robot reproduces a reference motion in simulation or on the real robot, usually reported together with success rate, velocity error, and acceleration error.

ExamplePHC reports both root-relative MPJPE and global MPJPE (in mm) on the AMASS test set, and counts tracking as failed whenever the average joint deviation from the reference motion exceeds 0.5 meters.

Also called

MPJPE

Related

Motion Tracking · Human Pose Estimation · Motion Retargeting · Perpetual Humanoid Control · ASAP · AMASS (Archive of Motion Capture as Surface Shapes)

Sources

Perpetual Humanoid Control for Real-time Simulated Avatars (PHC, arXiv 2305.06456)
A simple yet effective baseline for 3d human pose estimation (arXiv 1705.03098)
Human3.6M Dataset