2024-24679-fencing-touch-predictor

Project: CMU Fencing Classification Project Author: Ethan Kessler (Carnegie Mellon University) License: MIT Date: 2025

Description:

This model predicts which frame a touch is scored in a fencing bout from red/green intensity extracted from video frames. It was trained using AutoGluon on a curated dataset with frame information extracted using CV2 from 10 short fencing videos retrieved from https://actions.quarte-riposte.com/.

Framework:

• Model: WeightedEnsemble_L3
• Ensemble Weights: 'NeuralNetFastAI_BAG_L2': 0.55, 'LightGBMXT_BAG_L2': 0.15, 'ExtraTreesMSE_BAG_L2': 0.15, 'ExtraTreesMSE_BAG_L1': 0.1, 'LightGBM_r131_BAG_L2': 0.0

Performance:

• Validation Score ≈ -0.0539 (-root_mean_squared_error)
• Training runtime ≈ 0.02s
• Validation runtime ≈ 0.0s

Example Usage:

from autogluon.tabular import TabularPredictor model = AG_PRED("emkessle//2024-24679-fencing-touch-predictor") feat_cols = ["red_ratio", "green_ratio", "red_diff", "green_diff", "z_red", "z_green"] features = df[feat_cols].copy() result = ag_predictor().predict(features) annotated = result[0].plot() Notes:

• Intended for referee-assistive scoring and highlight extraction.
• Trained on clean data for Olympic-level scenarios. Limitations:

---

• Reduced accuracy under poor lighting or glare.
• May not generalize to local/college amateur venues.
• Does not identify fencers or track motion. Ethical Use:

---

For research, education, and sports analytics only. All data sourced from public fencing footage. Citation:

Kessler, E. (2025). "2024-24679-fencing-touch-predictor" Hugging Face: https://huggingface.co/emkessle/2024-24679-fencing-touch-predictor