What is HOTA? Higher Order Tracking Accuracy Metric Explained
What is HOTA? Higher Order Tracking Accuracy Metric Explained
HOTA (Higher Order Tracking Accuracy) is a metric for evaluating multi-object tracking (MOT) algorithms. It decomposes tracking performance into three key components: detection accuracy (DetA), association accuracy (AssA), and localization accuracy (LocA). The final HOTA score combines detection and association accuracy using their geometric mean across multiple localization thresholds (α). This provides a balanced evaluation of a tracker's ability to detect objects, maintain their identities over time, and localize them accurately.
HOTA Definition
HOTA is defined as the geometric mean of Detection Accuracy (DetA) and Association Accuracy (AssA), averaged over multiple localization thresholds (α). It provides a balanced evaluation of a tracker's ability to detect objects, maintain their identities over time, and localize them accurately.
HOTA Formula: DetA, AssA, and LocA
The final HOTA score is calculated by combining Detection Accuracy (DetA) and Association Accuracy (AssA) using their geometric mean across different localization thresholds (α). Localization Accuracy (LocA) provides an additional measure of how accurately matched objects are localized.
How HOTA Works: Localization, Detection, Association
Localization Accuracy (LocA)
Localization measures the spatial alignment between a predicted detection and a ground truth detection. It is calculated using the Jaccard index (IoU), which measures the overlap between the predicted and ground truth bounding boxes.
LocA = (1 / |TP|) × Σ Localization_score(c) for c ∈ TP
Detection Accuracy (DetA)
Detection accuracy (DetA) measures the agreement between predicted detections and the set of ground truth detections.
DetA = |TP| / (|TP| + |FN| + |FP|)
where TP = True Positives (matched detections), FN = False Negatives (missed ground truth detections), and FP = False Positives (extra predicted detections).
Association Accuracy (AssA)
Association accuracy (AssA) measures how well a tracker maintains object identity over time.
AssA = (1 / |TP|) × Σ Association_score(c) for c ∈ TP
Association_Score = |TPA| / (|TPA| + |FNA| + |FPA|)
where TPA = True Positive Associations (matched detections in a track), FNA = False Negative Associations (missed detections in a track), and FPA = False Positive Associations (extra detections in a track).
Why Geometric Mean?
The final HOTA score combines detection and association accuracy using the geometric mean. The geometric mean ensures that a tracker cannot compensate for poor association by over-detecting, or vice versa. Any improvement in one dimension (with the other held constant) will always increase the HOTA score. This property is called strict monotonicity.
What HOTA Reveals About Tracking Errors
HOTA provides more than a single overall tracking score. By separating detection, association, and localization performance, it helps identify different types of tracking errors, including missed detections, additional detections, track splits, and incorrect track associations.
This makes HOTA useful not only for comparing different trackers, but also for understanding where a tracking system performs well and where it needs improvement.
HOTA vs MOTA vs IDF1
MOTA (Multiple Object Tracking Accuracy) is defined as MOTA = 1 - (FP + FN + IDSW) / GT, where IDSW is the number of identity switches and GT is the number of ground truth detections. MOTA can range from negative infinity to 1, or from negative infinity to 100% when expressed as a percentage. Because MOTA is strongly influenced by detection errors, it does not separately measure association or localization accuracy.
Metric | Detection | Association | Localization | Range |
|---|---|---|---|---|
MOTA | ✅ | ❌ | ❌ | -∞ to 1 |
IDF1 | ❌ | ✅ | ❌ | 0 to 1 |
HOTA | ✅ | ✅ | ✅ | 0 to 1 |
HOTA FAQ
Q: What is the HOTA metric?
A: HOTA (Higher Order Tracking Accuracy) is a metric for evaluating multi-object tracking algorithms. It combines detection accuracy (DetA) and association accuracy (AssA) via geometric mean, averaged over localization thresholds.
Q: What is the difference between HOTA and MOTA?
A: MOTA primarily reflects detection accuracy and fails to penalize association errors like ID transfers. HOTA separately measures detection and association accuracy for a balanced evaluation.
Q: What is the HOTA formula?
A: HOTA = integral over α of √(DetA_α × AssA_α), where DetA is detection accuracy and AssA is association accuracy at localization threshold α.
Q: What is DetA in HOTA?
A: DetA (Detection Accuracy) = |TP| / (|TP| + |FN| + |FP|), measuring how well the tracker detects objects.