YOLOv8x (stock COCO checkpoint, mirror)
Not trained by me, and not modified. This is Ultralytics' unmodified stock YOLOv8x COCO checkpoint, mirrored so my tennis pipeline can fetch all its weights from one place. AGPL-3.0 permits redistribution; all credit belongs to Ultralytics.
Why it is here
The person class detects tennis players well - they are large, high-contrast
and slow relative to the ball, so a per-frame detector needs no temporal help.
The catch worth stating
A person detector finds people, not players. On a professional clip it returns ball kids, line judges and the chair umpire alongside the two players. Filtering by court position and "exactly one player per side of the net" is what turns person detection into player detection - the model cannot do it for you.
Note it is slow on CPU: ~14 s/frame at 1080p in my testing. Use a GPU, or a smaller variant, if throughput matters.
License
AGPL-3.0, as published by Ultralytics.