Detects the tennis ball in broadcast footage. Fine-tuned from YOLOv8m.

Same held-out validation set, before and after fine-tuning, at imgsz=960:

A tennis ball is ~15 px in a 1080p frame. YOLO's default imgsz=640 letterboxes it to ~5 px. Raising inference resolution to 960 took detection from 47% to 95.6% with no retraining at all. This 26M-parameter model then beat an 86M-parameter baseline, largely by being trained at the resolution it is served at.

The optimum tracks apparent ball size, not image size: 960 suits a tight amateur clip, while a wide broadcast shot wants 1600. Sweep it for your footage.

Use a low confidence floor and resolve which candidate is the ball from the whole trajectory rather than per frame - a per-frame argmax picks line markings and shoes.

AGPL-3.0, inherited from Ultralytics YOLOv8.