What is Multiple Object Tracking Precision?
Multiple Object Tracking Precision is a CLEAR MOT localization measure that averages the spatial similarity or distance of accepted ground-truth and predicted detection matches.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Separate localization quality from match coverage
A CLEAR MOT evaluator first constructs one-to-one matches subject to a localization threshold and continuity rule. For the resulting matched pairs, MOTP averages their localization quantity. Unmatched Ground Truth and predictions do not enter the numerator directly; they are represented by FN and FP in MOTA and related detection metrics.
The original CLEAR MOT definition expresses MOTP as average distance error. Current visual-tracking leaderboards often use IoU similarity instead, so every reported value needs an explicit direction and unit.
Declare whether MOTP is similarity or distance
For a similarity convention, MOTP = sum(similarity of matched pairs) / TP, and higher is better. With bounding boxes, similarity is often IoU and the result lies between zero and one. For a distance convention, MOTP = sum(distance of matched pairs) / TP, and lower is better; the unit may be pixels, meters, or another state-space measure.
Converting distance = 1 - similarity is valid only when the protocol explicitly defines that normalized relationship. It is not valid for arbitrary Euclidean or Mahalanobis distances.
Interpret MOTP together with the matching protocol
TrackEval reports MOTP as the sum of localization similarities divided by the CLEAR true-positive count. Its matching first preserves eligible previous-frame pairs and then favors greater similarity, which can affect the set being averaged.
A tracker can obtain high MOTP by localizing a small easy subset accurately while missing many objects. Report TP, FN, FP, MOTA, HOTA components, the representation and threshold, and never compare a distance-style MOTP directly with a similarity-style result.
Key Characteristics
- Averages localization only over accepted detection matches
- Uses either a higher-is-better similarity or lower-is-better distance
- Depends on the matching threshold and continuity policy
- Does not penalize unmatched detections inside its average
- Can use boxes, masks, points, or task-specific state distances
- Is unrelated to classification precision despite its name
Common Use Cases
- Reporting bounding-box alignment on CLEAR MOT benchmarks
- Comparing localization under one fixed matching protocol
- Auditing 2D, 3D, mask, or point tracking precision
- Separating localization changes from detection error changes
- Reproducing legacy tracking tables with explicit conventions
Example
Loading code...Frequently Asked Questions
How is MOTP calculated?
After protocol-defined one-to-one matching, add the localization similarity or distance of every true-positive pair and divide by the number of pairs. The report must state which quantity is averaged and which direction is better.
Is higher MOTP always better?
No. Higher is better when MOTP is average similarity, such as IoU. Lower is better when it is average distance or localization error. Read the evaluator definition instead of inferring direction from the metric name.
Why can MOTP be high when tracking is poor?
MOTP averages only accepted matches. A tracker may localize a small easy subset very well while producing many misses, false detections, or identity errors. Those failures require MOTA, HOTA, IDF1, and raw counts.
What is the difference between MOTP and classification precision?
Classification precision is `TP / (TP + FP)` and measures false-positive contamination. MOTP measures localization quality among matched detections; its historical use of the word precision does not imply that classification formula.
Can MOTP values be compared across implementations?
Only when both implementations use the same detection representation, similarity or distance function, threshold, matching policy, ignored-data rules, aggregation, and score direction. Otherwise the same label can describe incompatible quantities.