Find a moment where two players carry each other's labels. Tap the two players to swap, choose how far the swap reaches, then press Swap.
Demo clip The calibration is preset. Change it, then tap Analyze to re-run.
Tap a name to rename. Metrics count only time in play (inside rallies). Orange marks the best value of the four.
A rally runs from first contact until the ball is dead for more than 1.6 s. Tap a rally to replay it.
Priorities for each player and pair, ranked by impact. Each item shows the evidence from this clip, a target and a drill for the next session.
Name the four players for each video, and choose the AI that calibrates the court and picks the players out of the crowd.
Names are saved for each video separately. Leave a box empty to keep A1, A2, B1 or B2.
Choose how players are found in your videos. It runs in the browser of whoever presses Analyze, so it costs nothing per video. The choice is remembered in this browser.
Everyone can open a saved video with its full analysis. Only the admin can save or delete. Player progress is tracked from match to match.
Loading…
A rule-based summary of the tracking, plus an optional written report from Claude.
No report yet. The demo rallies work too.
How it works: five frames per second run a CenterNet detector for players and ball and MoveNet pose estimation on each player, all in your browser with WebGL; the video never leaves your device, and only the resulting metrics are saved if you press Save. Court calibration maps feet to real metres, players are tracked across frames (a clothing colour fingerprint re-identifies anyone who leaves the picture and returns), rallies are segmented from ball activity, and hits are inferred from ball direction changes near a player. Ball-derived numbers (rallies, shots, lobs) are the least reliable on wide broadcast footage.