What is Local Tracking Metrics?
Local Tracking Metrics are horizon-parameterized multi-object tracking measures that recompute optimal track correspondence inside temporal windows, including Local IDF1 and Average Local Tracking Accuracy.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Recompute track correspondence inside each horizon
For every center frame t and radius r, the evaluator restricts tracks to [t-r, t+r] and solves an optimal one-to-one assignment for that interval. LIDF1 accumulates interval IDTP counts and detection-count denominators; ALTA accumulates optimally assigned temporal-IoU mass and average active-track-count denominators. The final result is a ratio of accumulated numerators and denominators, not a simple mean of window scores.
The Local Metrics paper defines both measures and explains why separate accumulation handles empty and edge-truncated windows correctly.
The Go example isolates one always-visible truth track that is split between two predicted identities, making the horizon effect inspectable without hiding it behind detection or localization errors.
Read the horizon as an identity-duration requirement
A radius below one produces single-frame windows, so LIDF1(r) = ALTA(r) = DetF1; identity labels can be reassigned independently every frame. For r >= T-1, every window spans the complete sequence and the metrics equal their strict endpoints: IDF1 and ATA.
Intermediate horizons ask whether identities remain coherent for roughly 2r+1 frames, clipped at sequence boundaries. Express the horizon in seconds as well as frames when datasets use different frame rates, and report several meaningful horizons rather than selecting one after seeing rankings.
Use the curve to diagnose association range
A steep drop from short to medium horizons indicates locally correct detections that cannot preserve identity through occlusion or re-entry. A later drop points to long-range re-identification limitations. ALTA gives tracks equal top-level weight and is more sensitive to fragmentation; LIDF1 gives detections equal weight and can discount short incorrect tracks.
Each horizon requires assignment across all centered windows, costing roughly O(T*K^3) with a Hungarian solver for K tracks. These metrics complement HOTA, MOTA, and raw switch timing; they do not encode business harm, latency, or safety.
Key Characteristics
- Parameterizes identity evaluation by a temporal horizon
- Recomputes optimal one-to-one assignment in every window
- Includes detection F1 and strict global metrics as endpoints
- Separates short-range from long-range association behavior
- Offers detection-weighted LIDF1 and track-weighted ALTA
- Costs one sequence of assignment problems per horizon
Common Use Cases
- Selecting identity horizons for robotics and video analytics
- Diagnosing short-term versus long-term re-identification
- Comparing trackers that share similar detection quality
- Measuring fragmentation across application time scales
- Auditing sensitivity hidden by one full-sequence score
Example
Loading code...Frequently Asked Questions
What are LIDF1 and ALTA?
LIDF1 is a finite-horizon version of IDF1, and ALTA is a finite-horizon version of ATA. Both recompute the optimal truth-prediction track assignment inside windows centered on each frame.
What does the horizon mean in local tracking metrics?
The radius `r` defines a window `[t-r, t+r]` around each frame. It represents how long identity must remain consistent, so it should be reported in frames and converted to seconds for practical interpretation.
What happens at the shortest and longest horizons?
With single-frame windows, both metrics equal detection F1 because identity can be reassigned every frame. When every window covers the full sequence, LIDF1 equals IDF1 and ALTA equals ATA.
How are LIDF1 and ALTA different?
LIDF1 accumulates correctly identified detections and weights detections equally. ALTA accumulates temporal IoU under track assignment and normalizes by track counts, making it more sensitive to short fragments and duplicate tracks.
Should one horizon be used to rank all trackers?
Only when that horizon is justified by the application before evaluation. Otherwise report a curve or a small predeclared set of horizons; selecting the best-looking horizon after observing results invites metric gaming.