What is OSPA(2) Tracking Metric?
OSPA(2) Tracking Metric is a distance between finite sets of tracks that uses a time-averaged OSPA-derived distance between individual tracks as the base distance for an outer OSPA assignment.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Apply OSPA once to tracks and again to track sets
Each track can be represented as a state sequence over its support. An inner distance averages capped state disagreement over a declared time domain, treating one-sided existence as a maximum singleton-versus-empty error. That track distance becomes the base metric for an outer OSPA assignment between the truth-track set and estimate-track set.
The original OSPA(2) paper introduced this construction because independent framewise set distances cannot describe whole-track disagreement. The Go example illustrates p = 1 with two tracks per set and support-union averaging; production code must also handle unequal outer cardinality and the exact benchmark convention.
Make the evaluation window and lifecycle rules explicit
A track pair can disagree through spatial error, non-overlapping lifetimes, or both. The cut-off limits state error and sets the cost of one-sided existence; the order controls sensitivity to larger errors. The outer OSPA normalization then makes the result a per-track distance.
Window length changes the question. A short sliding window emphasizes recent association quality; a full-sequence window penalizes long-range fragmentation and identity mismatch. Declare whether tracks are truncated at window boundaries, how gaps are represented, which timestamps enter the inner average, and how sequence-level results are aggregated.
Distinguish track-set distance from benchmark identity scores
A recent trajectory-metric review places OSPA(2) among metrics on sets of tracks and highlights the importance of temporal alignment and false or missed track segments. It does not make one metric universally best; the application determines which errors should dominate.
T-GOSPA explicitly decomposes localization, missed, false, and switching costs through time-varying assignments. HOTA balances detection and association over localization thresholds, while IDF1 applies a sequence-level identity match. Publish OSPA(2) parameters and window rules, and report component, latency, and calibration evidence separately.
Key Characteristics
- Defines a distance between finite sets of complete tracks
- Uses a time-averaged track distance inside an outer OSPA assignment
- Penalizes localization error and non-overlapping track support
- Captures fragmentation that independent framewise OSPA can miss
- Depends on cut-off, order, time window, and lifecycle conventions
- Keeps one whole-track assignment over the evaluated window
Common Use Cases
- Comparing multi-target trackers that output complete trajectories
- Evaluating long-range continuity through occlusion and reappearance
- Detecting fragmented estimates despite accurate per-frame positions
- Assessing large-scale labeled Random Finite Set trackers
- Studying how evaluation-window length changes tracker rankings
Example
Loading code...Frequently Asked Questions
Why is the metric called OSPA(2)?
It is read as OSPA-on-OSPA. An OSPA-derived distance compares individual tracks over time, and that distance becomes the base cost inside a second OSPA assignment between finite sets of tracks.
How is OSPA(2) different from framewise OSPA?
Framewise OSPA independently reassigns current states at every timestamp and can miss identity swaps. OSPA(2) compares whole tracks over a window and keeps the outer assignment consistent, so fragmentation and temporal mismatch affect the result.
How does window length affect OSPA(2)?
A short window emphasizes recent position and association quality, while a long or full-sequence window exposes longer fragmentation and identity errors. Window boundaries, gaps, and aggregation rules must be identical for fair comparisons.
Is OSPA(2) the same as T-GOSPA?
No. OSPA(2) applies an OSPA construction to a whole-track base distance and a track-set assignment. T-GOSPA uses time-varying assignments plus an explicit switching cost and offers a different localization, missed, false, and switch decomposition.
Can OSPA(2) replace HOTA or IDF1?
Not universally. OSPA(2) is a metric-space construction for track sets, HOTA balances detection and association under a benchmark matching protocol, and IDF1 emphasizes sequence-level identity agreement. Report the metric that matches the operational question.