fencing-scoreboard-yolov8

FENCING SCOREBOARD YOLOv8 MODEL

Project: CMU Fencing Classification Project Author: Michael Stefanov (Carnegie Mellon University) License: MIT Date: 2025

Description:

This model detects illuminated fencing scoreboards (touches scored) from bout video frames. It was trained using YOLOv8 on a curated dataset of ~2000 images (lit/unlit scoreboards, plus negatives).

Framework:

• Model: YOLOv8n (Ultralytics)
• Epochs: 80
• Batch Size: 16
• Image Size: 640x640
• Learning Rate Schedule: Cosine Annealing
• Dataset: mastefan/fencing-scoreboard-yolov8 (Hugging Face)
• Classes: 1 ("scoreboard")

Performance:

• [email protected] ≈ 0.91
• Precision ≈ 0.89
• Recall ≈ 0.87

Example Usage:

from ultralytics import YOLO model = YOLO("mastefan/fencing-scoreboard-yolov8/best.pt") result = model.predict("sample_frame.jpg", conf=0.25) annotated = result[0].plot()

Notes:

• Best results at conf=0.25; restore bounding regions with conf>0.85 for highlight masking.
• Intended for referee-assistive scoring and highlight extraction.
• Designed for single-scoreboard scenarios.

Limitations:

• Reduced accuracy under poor lighting or glare.
• May not generalize to multi-scoreboard venues.
• Does not identify fencers or track motion.

Ethical Use:

For research, education, and sports analytics only. No facial or biometric recognition. All data sourced from public fencing footage.

Citation:

Stefanov, M. (2025). "Fencing Scoreboard YOLOv8 Model." Hugging Face: https://huggingface.co/mastefan/fencing-scoreboard-yolov8