⛳ SwingTrace: golf club detector (YOLO11m)

Detects the club shaft, club head and hands in golf swing videos. Used by SwingTrace to rebuild the club-head path through the whole swing, including the blurry frames at the top and at impact.

Try it: Colab notebook · Code: github.com/nishsm/swingtrace

Use it

pip install git+https://github.com/nishsm/swingtrace.git
swingtrace my_swing.mp4            # boxes + club path

Or just the detector, with Ultralytics:

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

model = YOLO(hf_hub_download("nishsm/swingtrace", "swingtrace-yolo11m.pt"))
results = model.predict("my_swing.mp4", conf=0.25)

Classes

id	label	meaning
0	0	shaft
1	1	club head
2	3	hands / grip

The raw label strings come from the source dataset.

Results

Unseen phone videos (7 clips, 3,018 frames, conf 0.5): club head detected in 71.6% of frames; head or shaft (so the head can be recovered from the shaft) in 93.4%. These clips have no hand labels, so the counts include occasional false positives.

Validation set: precision 0.930 · recall 0.849 · mAP@50 0.918 · mAP@50-95 0.674.

Treat the validation mAP as an upper bound. The dataset's frames come from videos and were split at random, and 91% of validation frames have a training frame within 2 frames of them.

Training

• Base: YOLO11m (Ultralytics 8.3)
• Data: golf-club-tracking by club-head-tracking on Roboflow Universe (CC BY 4.0)
• 640 px, batch 2, mixed precision, early stopping (patience 100): best epoch 456 of 556, about 48 h on an RTX 4060 laptop GPU (8 GB)

Limitations

Trained and tested on single-golfer clips from a static phone camera. Expect worse results with moving cameras, several people in frame, or unusual angles. It sometimes boxes thin bright objects (for example a ceiling light) as a shaft.

License

AGPL-3.0, following Ultralytics YOLO, which these weights are fine-tuned from.