Every sports-analytics product quotes an accuracy number. The trouble is that player tracking isn't one task, so a single number can hide a lot. Here's what the common metrics measure and which ones matter for coaches and analysts.
Detection: is there a player here?
Detection finds people in each frame and draws a box around them. It's usually measured with precision (how many boxes are real players) and recall (how many players got a box). Modern models are very good at this on clear broadcast footage, and much less reliable with low resolution, motion blur, crowds of players in the box or players partly hidden behind each other.
Tracking: is it the same player as last frame?
Tracking links detections over time. The common summary metric is MOTA (multi-object tracking accuracy), which combines missed players, false detections and identity switches. A high MOTA can still hide frequent switches, because misses and false positives often dominate the score.
Identity: is it player number 7?
For analysis, what matters most is whether each track stays attached to the right person for the whole match. Metrics like IDF1 and HOTA reward keeping identities correct over time. If a system swaps two midfielders after a corner, every distance, heatmap and sprint count after that is attributed to the wrong player.
Questions to ask any vendor
- Which metric is the number, and on what footage was it measured?
- Was the test footage from your league, cameras and conditions, or a public benchmark?
- How are identity switches handled and corrected?
- How does accuracy change for wide shots, close-ups and replays?
- Can you see the raw tracks for a sample match, not just the dashboard?
Why benchmarks need real footage
Public datasets are useful for comparing models, but every league has its own camera angles, kits, pitches and broadcast style. The only accuracy number that matters is the one measured on footage like yours, with the outputs your analysts will actually use.
That's how we approach sports computer vision at Kaizenext: start from a client's real matches, agree the metrics that matter for their analysis, and report against those.