What is Average Tracking Accuracy?
Average Tracking Accuracy is a VACE metric that globally matches ground-truth and predicted tracks by temporal intersection over union, then normalizes the optimal overlap by their average track count.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Build temporal IoU from per-frame overlap
For truth track i and predicted track j, the relaxed VACE protocol first marks each frame where their detections satisfy a spatial-overlap threshold. It then computes Q_ij = matchingFrames / unionPresenceFrames, where the denominator counts frames in which either track exists. This temporal IoU penalizes missing support, extra support, and incompatible localization.
The original VACE evaluation paper introduced comprehensive detection and tracking measures. The exact spatial criterion and relaxed-versus-original variant must be stated when reproducing ATA.
Optimize one assignment across the whole video
A bipartite assignment chooses at most one predicted track for each truth track and vice versa, maximizing TrackTP = sum(Q_ij) over selected pairs. With K truth tracks and Khat predicted tracks, ATA = TrackTP / (0.5 * (K + Khat)). The score is also the harmonic mean of ATR = TrackTP / K and ATP = TrackTP / Khat.
Because every matched track contributes at most one, short and long trajectories have equal top-level weight. Splitting one truth identity into multiple predictions increases Khat and leaves only one fragment available for the global match.
Use ATA for strict association evidence
TrackEval's VACE implementation computes the relaxed ATA variant with a 0.5 per-frame overlap threshold and reports SFDA as a complementary frame-detection measure. Evaluator preprocessing, empty tracks, class handling, and sequence combination remain part of the contract.
ATA is more association-sensitive than detection-weighted MOTA, while IDF1 gives equal weight to detections rather than tracks. Local Tracking Metrics turn ATA into ALTA at finite horizons; HOTA provides a different partial-credit decomposition. Report several views rather than selecting the score that favors one tracker.
Key Characteristics
- Uses temporal IoU between complete trajectories
- Finds one global one-to-one track assignment
- Weights tracks rather than individual detections
- Penalizes fragmentation, merges, misses, and extra tracks
- Has a normalized range from zero to one
- Depends on the spatial-overlap and preprocessing protocol
Common Use Cases
- Evaluating strict full-sequence identity preservation
- Comparing trackers with different fragmentation behavior
- Giving short and long tracks equal top-level importance
- Auditing identity merges and duplicate trajectories
- Providing an association-sensitive complement to MOTA
Example
Loading code...Frequently Asked Questions
How is ATA calculated?
Compute temporal IoU for every truth-prediction track pair, choose the one-to-one assignment with maximum total IoU, and divide that sum by half the total number of truth and predicted tracks.
What is temporal IoU in ATA?
In the relaxed VACE form, it is the number of frames where a track pair passes the spatial-overlap threshold divided by the number of frames where either track exists. It measures overlap in temporal support after localization eligibility.
How is ATA different from IDF1?
Both use a global track assignment. IDF1 maximizes correctly identified detections and therefore weights detections equally; ATA maximizes temporal IoU and normalizes by track counts, giving each track equal top-level weight.
How does fragmentation affect ATA?
If one true trajectory is split into several predictions, only one fragment can be selected by the global one-to-one assignment. The other fragments also increase the predicted-track denominator, so ATA penalizes fragmentation strongly.
When should ALTA be used instead of ATA?
Use ALTA when identity only needs to remain correct over an application-specific time horizon or when you need to see how association degrades as that horizon grows. ATA remains the strict full-sequence endpoint.