Average Distance of Model Points

Poses an object’s 3D model under both the true and predicted pose and averages the point-to-point distance, to score 6D pose estimation.

ADD is the most common error metric for 6D object pose estimation, introduced by Hinterstoisser and colleagues at ACCV 2012 alongside the LINEMOD dataset. It works by taking points on an object’s 3D model, transforming them by the ground-truth pose and by the predicted pose separately, and averaging the distance between corresponding points; a common criterion counts a prediction correct if this average distance is under 10% of the model’s diameter, and reports the resulting accuracy. For symmetric objects (bowls, cans) that look the same after a rotation, matching points one to one unfairly penalizes reasonable predictions, so ADD-S is used instead: for each point, it finds the nearest point in the other point set before averaging — the BOP benchmark calls this ADI. PoseCNN (2018) swept this threshold from 0 to 10 centimeters on YCB-Video and reported the area under the accuracy-threshold curve (AUC), a format that later became common. The BOP benchmark notes that ADI can assign a low error to poses that are visibly misaligned, so it switched to three other metrics instead: VSD, MSSD, and MSPD.

ExampleFor a pot 20 centimeters in diameter, the 10%-of-diameter rule sets the threshold at 2 centimeters: a predicted pose whose model points deviate 1.5 centimeters on average from the ground truth counts as correct, while a 3-centimeter deviation counts as wrong.

Also called

ADD, ADD-S, ADI

Related

6D Object Pose Estimation · BOP (Benchmark for 6D Object Pose Estimation) · FoundationPose · YCB Object and Model Set · Chamfer Distance · Pose

Sources

PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes (arXiv:1711.00199)
BOP Challenge 2019：pose-error functions