ARC-SORT hits 81.82 HOTA on SportsMOT and runs at 50 fps

The ARC-SORT team has presented a tracker for multi-object tracking in sports broadcasts. On the SportsMOT test set it scores 81.82 HOTA and has the fewest identity switches among the 69 leaderboard entries as of August 2026. MOTA is 97.09, a tie for third place. On DanceTrack it reaches 62.1 HOTA with the same association settings. It runs at about 50 frames per second on a laptop RTX 3060. The tracker itself trains nothing, and the authors promise to publish the code.

A tracker that knows how uncertain it is

Anyone who has worked with football or basketball footage knows the problem. Players accelerate and turn within fractions of a second, while the camera pans and changes zoom. Classic tracking-by-detection predicts each track's position with a Kalman filter, then matches the prediction to detections. When motion is unpredictable, the prediction drifts and the matching breaks down.

ARC-SORT uses a Kalman filter with several interacting motion models (IMM). For each track, it produces a prediction covariance. The authors turn this covariance into a per-track search region: the less reliable the forecast, the wider the zone in which a detection can be matched to the track. The method is called uncertainty-adaptive Expansion IoU (EIoU).

An important detail from the ablations: what works is the continuous coupling. The same covariance used as a hard Mahalanobis gate made results worse at every threshold tested. Richer covariance statistics added nothing measurable.

Crossings and leaving the frame

When two tracks cross, an identity error is most likely. In ARC-SORT, the assignment cost itself shows whether the decision was a close call. Such pairs stay under observation. Once appearance data accumulates after the crossing, the tracker can perform at most one corrective ID swap if the first decision turns out to be wrong. This adds +1.0 HOTA on SportsMOT and DanceTrack tests, at a cost of less than a millisecond per frame.

A separate case is a player who leaves the frame for several seconds. Extrapolating motion over that period is useless. So offline re-identification drops the usual motion gate and stitches tracklets together by appearance alone.

The numbers

• SportsMOT test: 81.82 HOTA, the fewest ID switches among 69 entries (as of August 2026), MOTA 97.09 (tied for third place).
• DanceTrack test: 62.1 HOTA with the same association parameters.
• On Deep-EIoU's public detections and embeddings: 80.31 HOTA, which is +3.1 over Deep-EIoU.
• ID swap correction: +1.0 HOTA on both datasets, less than 1 ms per frame.
• Speed: about 50 frames per second on a laptop RTX 3060.

Where this helps in practice

Comparing on identical Deep-EIoU inputs removes the main question hanging over trackers: whether the credit belongs to the detector or to the association logic. Here the +3.1 HOTA gain comes from the logic alone.

For those building sports analytics, this means they can swap out the association module without touching the detector or embeddings. A speed of 50 fps on a laptop GPU leaves headroom for real-time processing, for example live statistics. At the same time, part of the result (re-identification) runs offline, that is, after a clip has been processed.

The limitations are also worth keeping in mind: until the code is published, the numbers cannot be reproduced, and DanceTrack results are noticeably lower than SportsMOT. The practical takeaway for engineers: filter uncertainty is more useful for widening the search region than for hard-pruning candidates.