Skip to content

Limitations and FAQ

Measured on one 1080p broadcast clip each, not yet a labelled benchmark:

  • Tennis: court calibrated in 88% of frames (the rest are close-ups and replays), ball found in 71%, both players identified in 86%.
  • Football: pitch calibrated in 50–75% of broadcast frames in our tests; positions are approximate (a few metres).

A benchmark against hand-labelled matches is on the roadmap.

  • Player identities, names and team home/away (no identity model).
  • Ball height in football; forehand/backhand in tennis.
  • Goals, throw-ins, free kicks, fouls and offside.
  • Real time on a laptop GPU for tennis at 25 fps.

Does it need cameras in the stadium? No, one broadcast-style video.

Does it guess when it cannot see? No. Unknown positions are null, and the dashboard says why (close-up, no calibration, no model).

Can I use it commercially? Under the AGPL, or with a commercial licence. See Licensing.