One camera’s recording of a match becomes a record you can review: the ball’s path, the court, every rally and the score.
Court, ball, players, actions, rallies and the review app work today, including on a fixed camera behind the end line; 3D flights work but are limited by the 2D ball track; automatic scoring is next.
An illustration of the match review app with synthetic data — press play, click a rally, drag the timeline.
Real output, not a drawing: a fixed camera behind the end line (GCU vs UCLA) analysed end to end, and the review app.
Only numbers with a source. Proxies are named as proxies; a metric without labelled data is not shown.
median court-position error on held-out broadcast clips (96% within 0.3 m); 0.49 m on held-out gym images
fixed end-line clips whose court, fitted to the painted lines, is right to 2-6 cm (median, hand-labelled floor points); the sixth is stretched in depth, 1.06 m off. Both keypoint models fail on this view
IDF1 of player tracking on SportsMOT volleyball, with a detector fine-tuned to box players only (precision 0.970: no referees or spectators)
ball detection F1 (10 px) of WASB fine-tuned on gym footage, on held-out gym matches; VballNet V4c, in the pipeline today, scores 0.859
macro F1 of block / receive / set / spike per player on 16 unseen matches of the Volleyball dataset (the capstone detector: 0.303)
median 3D error of an attack from one camera with the touch anchored to the player who made it (0.76 m without), on synthetic rallies with realistic detection errors; with several players along the ball’s ray the anchor is withheld, since a wrong one costs metres
rally detection F1 on two held-out broadcast matches; 0.76 on six fixed end-line clips with a model trained on labels read from the score bug
shirt digit mAP@0.5 on teams the model never saw; numbers are voted per player over time
A staged pipeline: a new model reruns only the stages after it.
Camera cuts and frame-to-frame camera motion, so the court follows pans and zooms.
14 keypoints (net band included) per camera shot, the better of two models; a still end-line camera gets its court from the painted lines of an empty-court background plate. Doubtful courts are flagged and not used.
VballNet V4c on every frame; misses stay misses. A gym-tuned WASB replaces it after its broadcast test.
YOLO26s fine-tuned to box players only, BoT-SORT tracking, positions in court metres; referees, line judges and spectators are left out.
Block, receive, set and spike per player from a classifier weighted by where the ball is, serves from a detector, shirt numbers voted over time; they become suggested tags the coach accepts.
One camera, calibrated from the court and the net; each flight fitted to physics with the touches anchored to the players, its error shown and impossible fits dropped. Real flights are often low quality: the 2D ball track is the limit.
Play or no play from the system’s own cues, with one model for broadcasts and one for still end-line cameras.
Landing, in or out and who won the point, with unsure calls flagged for a one-key correction. Until then the coach sets each rally’s winner.
Drop a match in; each stage reports as it runs.
Video with the court, players and ball trail, a rally inspector, and lanes for rallies, ball, landings, actions, tags and review.
J / K between rallies, 1 / 2 to set the winner. Corrections become labels for measuring accuracy.
Each rally’s ball flights on a 2D court and in 3D, with low-quality and unseen parts drawn as such.
Attack and serve tags become attack efficiency, kill rate, aces and serve errors per player, exported as CSV.
Every stage says whether it ran, is unavailable, or is not built yet. Demo data is always labelled.
The court keypoints include the net band, which calibrates the camera fully. Between touches the ball follows a ballistic arc, so one camera’s rays plus gravity fix where the ball is in 3D — with an error per flight, and flights the camera cannot resolve marked as such.
Tennis and professional volleyball rely on several synchronised cameras. With one, the plan combines the 3D fit, learned event spotting and the sound of the bounce, and reports uncertainty near the lines.
Match video carries its own answer key: the score bug changes after every point. Read with OCR and filtered by the rules of the game (one point at a time), it gives each rally’s end and winner, which trained the fixed-camera rally model and will measure automatic scoring.
Each component is scored with standard metrics (court error in metres; HOTA / IDF1 for players; mAP for actions; F1 for the ball). Results and the scripts that produce them live next to the code, and corrections are published when an earlier number turns out wrong.
Product requirements, behaviour specifications and results are in the repository.
Started as a senior capstone at National Taiwan Ocean University, Department of Computer Science and Engineering: Liang Yu-Jia (lead), Tsai Pei-Ying, Chung Chia-Hsin; advisor Professor Ting Pei-Yi. The current rebuild is by Liang Yu-Jia.