Object Tracking Metrics: MOTA, MOTP, IDF1, and HOTA

A tracker can detect every object perfectly and still be useless if it keeps swapping their IDs. That’s why tracking needs metrics that score detection and association together — and why no single number did it well until HOTA. We’ll watch one door camera make exactly one identity mistake and see how differently each metric reacts.

Multi-object tracking (MOT) outputs boxes with consistent IDs across frames. Per frame it’s detection — match hypotheses to ground truth by IoU — but the new axis is whether an object keeps the same ID over its whole trajectory.

The running example: DoorCam, two people, one slip

DoorCam films a doorway for 5 frames. Two people, A and B, are present the whole time — so there are $2 \times 5 = 10$ ground-truth boxes. DoorCam detects every box perfectly (no misses, no ghosts) with one flaw: halfway through, it loses person A for a beat and re-labels them with a new ID. One identity slip, otherwise flawless.

MOTA — the classic detection-centric score

Intuition: add up everything that went wrong per frame — misses, false alarms, ID switches — and divide by how many objects there were.

From the CLEAR MOT metrics [1]:

$$\text{MOTA} = 1 - \frac{\sum_t \big(FN_t + FP_t + \text{IDSW}_t\big)}{\sum_t GT_t}.$$

MOTP — localization quality

Intuition: of the boxes you did match, how tightly do they fit? — average overlap, nothing about identity.

$$\text{MOTP} = \frac{\sum_{t,i} \text{IoU}(p_{t,i}, g_{t,i})}{\sum_t c_t},$$

where $c_t$ is the number of matches in frame $t$. For DoorCam, if the 10 matched boxes average IoU 0.82, then MOTP = 0.82 — a quality-of-fit number orthogonal to everything else, and silent about the ID slip.

IDF1 — the identity-centric score

Intuition: forget per-frame counting — ask “for how much of each person’s life did we keep one consistent ID on them?”

IDF1 [2] solves one global assignment between predicted and true trajectories, then takes an F1 over identity matches:

$$\text{IDF1} = \frac{2\,\text{IDTP}}{2\,\text{IDTP} + \text{IDFP} + \text{IDFN}}.$$

HOTA — the modern, balanced standard

Intuition: score detection and association separately, then take their geometric mean — so a tracker must be good at both; being great at one can’t paper over the other.

HOTA [3]:

$$\text{HOTA} = \sqrt{\text{DetA} \cdot \text{AssA}},$$

where DetA is detection accuracy (Jaccard over matched detections) and AssA is association accuracy averaged over correctly-matched pairs (HOTA is itself averaged over localization thresholds $\alpha$).

Now you try

Three calculators, one per metric. Push IDSW up in the MOTA panel and watch how little it moves — then drop AssA in the HOTA panel and watch HOTA fall hard. That contrast is the whole reason the field moved to HOTA.

In code

For MOTA / MOTP / IDF1, the motmetrics library is the standard:

import motmetrics as mm

acc = mm.MOTAccumulator(auto_id=True)

# For each frame, give GT ids, hypothesis ids, and a cost matrix (e.g. 1 - IoU).
acc.update(
    [1, 2],                      # ground-truth object ids present this frame
    ["a", "b"],                  # tracker hypothesis ids
    [[0.1, 0.9],                 # cost (1 - IoU), shape: n_gt x n_hyp
     [0.8, 0.2]],
)

mh = mm.metrics.create()
summary = mh.compute(
    acc,
    metrics=["mota", "motp", "idf1", "num_switches", "mostly_tracked", "mostly_lost"],
    name="seq1",
)
print(summary)

For HOTA specifically, use the official TrackEval toolkit — it’s the reference implementation the benchmarks score against.

Caveats & common pitfalls

• No single classic number is enough. MOTA is detection-dominated (DoorCam: 0.90) and IDF1 association-dominated (0.80); reporting only one hides half the story. HOTA exists exactly because of this [3] — quote it, but keep DetA/AssA visible.
• MOTA can go negative and isn’t a percentage. Treat it as an error budget, not an accuracy you can compare loosely across datasets [1].
• Everything rides on the matching threshold and the detector. Tracking metrics inherit the IoU-matching choices from detection; a different threshold reshuffles ID switches and FPs.
• Use the reference implementations. MOTA/MOTP/IDF1 from motmetrics, HOTA from TrackEval — hand-rolled global assignment gets the edge cases (gaps, re-entries, crowd regions) subtly wrong.

Takeaways

• Tracking scores two things: detection and identity. No single classic metric captured both — DoorCam’s one ID switch scored 0.90, 0.80, or 0.87 depending on who’s asking.
• MOTA = detection-centric (FN + FP + ID switches), can be negative; MOTP = localization quality only.
• IDF1 = identity-centric via global trajectory matching; MT/ML summarize trajectory coverage.
• HOTA = √(DetA · AssA) balances the two and is the current standard — report it, and read DetA vs AssA to see which half failed.

Next, the hardest to pin down — scoring images that have no ground truth → Image generation metrics: FID, IS, KID, LPIPS.

Further reading

• See the metrics in the wild: the MOTChallenge benchmark — leaderboards and datasets where HOTA/MOTA/IDF1 are reported.
• Reference implementation: TrackEval — HOTA, CLEAR-MOT, and Identity metrics in one toolkit.
• Related on CondadosAI: object-detection metrics (the per-frame matching tracking builds on); segmentation metrics.

References

[1] Bernardin, K., & Stiefelhagen, R. (2008). Evaluating Multiple Object Tracking Performance: The CLEAR MOT Metrics. EURASIP Journal on Image and Video Processing, 2008, 246309. DOI:10.1155/2008/246309.

[2] Ristani, E., Solera, F., Zou, R., Cucchiara, R., & Tomasi, C. (2016). Performance Measures and a Data Set for Multi-Target, Multi-Camera Tracking. ECCV Workshops. arXiv:1609.01775. (Defines IDF1.)

[3] Luiten, J., Os̆ep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixé, L., & Leibe, B. (2021). HOTA: A Higher Order Metric for Evaluating Multi-Object Tracking. International Journal of Computer Vision, 129, 548–578. DOI:10.1007/s11263-020-01375-2.

[4] Dendorfer, P., et al. (2021). MOTChallenge: A Benchmark for Single-Camera Multiple Target Tracking. International Journal of Computer Vision, 129, 845–881. Tooling: motmetrics, TrackEval.

[5] Forsyth, D. A., & Ponce, J. (2012). Computer Vision: A Modern Approach (2nd ed.), ch. 11 (tracking). Pearson.