Object Pose 6D Estimation 中文

Metrics atlas

Object Pose Evaluation Metrics

Distinguish the error function, success threshold, and final aggregate score before interpreting results. Each entry includes a quick recap, formulas, limitations, and links to papers in this library.

01 · Error

Where is the error measured?

Model surfaces, visible depth, image projections, and rotation angles measure different aspects of error.

02 · Protocol

How are symmetry and occlusion handled?

ADD-S, VSD, and BOP's symmetry-aware rules handle different issues and are not interchangeable.

03 · Report

Are the thresholds and aggregation consistent?

0.1d, 5px, 5°5cm, AUC, AR, and AP are all part of the evaluation protocol.

Pose components

Rotation, translation, and joint thresholds

Inspect the two pose components separately, then use n° mcm to determine whether both meet the required precision.

3 metrics

Rotation Error

Geodesic Rotation Error

error
Uses the shortest rotation angle between predicted and ground-truth rotations as the error.

Translation Error

3D Translation Error

error
Directly measures the 3D distance between predicted and ground-truth object centers.

n° mcm

Rotation–Translation Threshold Accuracy

threshold
Counts a pose as correct only when both rotation and translation errors fall below their thresholds.

Model-point metrics

Model-point distances and threshold aggregation

The ADD family depends on CAD point sets; ADD-S and ADD(S) address reporting for symmetric objects.

5 metrics

ADD

Average Distance of Model Points

error
Compares each CAD point under predicted and ground-truth poses, averaging the 3D distances.

ADD-S

Average Closest Point Distance

error
Relaxes point identity: each ground-truth point is matched to the nearest point in the predicted model.

ADD(S) / ADD(-S)

Symmetry-dependent ADD Reporting Protocol

protocol
Selects ADD or ADD-S according to object symmetry, then aggregates the results into one score.

ADD / ADD-S AUC

Area Under the ADD Recall Curve

aggregate
Averages recall over a range of error thresholds instead of relying on a single threshold.

2D Projection

2D Projection Error

error
Projects CAD points into the image and measures the average pixel distance between predicted and ground-truth projections.

BOP protocol

BOP: error functions and task scores

VSD, MSSD, and MSPD are errors; AR and AP are aggregation protocols for different tasks.

5 metrics

VSD

Visible Surface Discrepancy

error
Compares only object surfaces that should be visible in the image, treating indistinguishable poses as equivalent.

MSSD

Maximum Symmetry-Aware Surface Distance

error
Minimizes the maximum 3D surface deviation over all allowed global symmetry transformations.

MSPD

Maximum Symmetry-Aware Projection Distance

error
Measures the worst model-point displacement in image pixels after accounting for symmetry.

BOP AR

BOP Average Recall

aggregate
Computes VSD, MSSD, and MSPD recall at multiple thresholds, then aggregates them into BOP's localization score.

BOP AP

BOP Average Precision for 6D Detection

aggregate
Uses precision–recall curves to evaluate detection and pose jointly in 6D detection tasks where the objects present in the image are unknown.

9DoF & shape

Category-level size and shape quality

These complementary metrics are needed when methods also predict dimensions, 3D boxes, or complete shapes.

2 metrics

3D IoU

3D Bounding-box Intersection over Union

error
Compares the overlap of predicted and ground-truth 3D bounding boxes, reflecting position, rotation, and size together.

Chamfer Distance

Bidirectional Chamfer Distance

shape
Uses bidirectional nearest-point matching between predicted and ground-truth shape point clouds to measure reconstruction similarity.