🏈 The Playmakers

Real-Time NFL Player Position Classification

A deep learning framework for identifying player positions from pre-snap formations using contextual object detection

John Michael Slezak, Atif Khan, Chenhao Lu, Ruida Zeng

Department of Computer Science, Brown University

Overview | Features | Installation | Project Structure | Results | Citation

---

Overview

Playmakers is a computer vision framework for classifying NFL player positions from pre-snap formations using YOLOv8. The unique challenge is that all players appear visually similar—the model must learn from spatial context and positioning rather than individual features.

• ✅ Context-Aware Detection — Leverages player positioning and formation patterns
• ✅ Real-Time Performance — Optimized for live game footage analysis
• ✅ Custom Dataset — 503 annotated images from NFL championship games
• ✅ 8 Position Classes — WR, CB, Safety, LB, RB, QB, TE, Fullback

---

Features

🎯 End-to-End Pipeline From data preprocessing to real-time inference, all components are production-ready. 📊 Custom Dataset Processed NFL footage from 2024 Super Bowl + AFC/NFC Championships with 500+ annotated images.

⚡ Real-Time Inference YOLOv8 architecture enables blazing fast predictions suitable for live analysis. 📈 Comprehensive Evaluation Detailed metrics including confusion matrices, F1 curves, and precision-recall analysis.

---

Installation

Prerequisites

• Python 3.8+
• CUDA-capable GPU (recommended for training)

Setup

# Clone the repository
git clone https://github.com/yourusername/the-playmakers.git
cd the-playmakers

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Download Dataset

The dataset can be downloaded from Roboflow or extracted from the zip files in WR/datasets/:

# Option 1: From Roboflow
from roboflow import Roboflow
rf = Roboflow(api_key="YOUR_API_KEY")
project = rf.workspace("cs-1430").project("wr-finder")
version = project.version(3)
dataset = version.download("yolov8")

# Option 2: Extract from included zip files
# Unzip WR/datasets/NewWRData.zip

---

Project Structure

the-playmakers/
│
├── 📄 README.md                    # Project documentation
├── 📄 requirements.txt             # Python dependencies
├── 📄 .gitignore                   # Git ignore rules
├── 📄 Final Report.pdf             # Academic paper (PDF)
├── 📄 Final Poster.pptx            # Project poster presentation
│
├── 📁 WR/                     # Main project data and code
│   │
│   ├── 📁 notebooks/               # Jupyter notebooks
│   │   ├── main.ipynb              # Data loading + color matching + YOLO detection
│   │   ├── WRYOLOCUSTOM.ipynb      # YOLOv8 training via Roboflow (primary)
│   │   ├── player_detection.ipynb  # YOLOv5 training pipeline
│   │   ├── test_vis.ipynb          # Model evaluation & visualization (Colab)
│   │   ├── newwr_test_vis.ipynb    # Model evaluation (local Windows)
│   │   └── field_filter.ipynb      # Field preprocessing experiments
│   │
│   ├── 📁 images/                  # Sample images
│   │   ├── afc_championship_2024.jpg           # Input image (2024 AFC Championship)
│   │   ├── afc_championship_2024_processed.jpg # Processed output after filtering
│   │   ├── superbowl_lviii_detection_1.png     # Detection example (Super Bowl LVIII)
│   │   └── superbowl_lviii_detection_2.png     # Detection example (Super Bowl LVIII)
│   │
│   ├── 📁 datasets/                # Dataset archives
│   │   ├── NewWRData.zip           # Main dataset (train/val/test splits)
│   │   ├── runs.zip                # Training runs (weights, logs, metrics)
│   │   ├── WR Finder.v1i.yolov8.zip      # Roboflow YOLOv8 format
│   │   └── WR Finder.v1i.tensorflow.zip  # Roboflow TensorFlow format
│   │
│   ├── 📁 scripts/                 # Python scripts
│   │   └── filter_field.py         # Standalone field filtering script
│   │
│   └── 📁 docs/                    # Additional documentation
│       └── color matching poster stuff.docx
│
├── 📁 Final Report/                # LaTeX source for academic paper
│   ├── main.tex                    # Main LaTeX document
│   ├── ProjectFinal_ProjectReportTemplate.bib    # Bibliography
│   ├── ProjectFinal_ProjectReportTemplate.pdf    # Compiled PDF
│   ├── architecture.png            # YOLOv8 architecture diagram
│   ├── confusion_matrix.png        # Classification confusion matrix
│   ├── confidence_curve.png        # F1-confidence curve
│   ├── precison_curve.png          # Precision-recall curve
│   ├── color_matching_1.png        # Color matching result example
│   ├── color_matching_2.png        # Color matching result example
│   ├── matches_1.png               # Correct detection example
│   ├── matches_2.png               # Correct detection example
│   ├── mismatch_1.jpg              # Incorrect detection example
│   ├── mismatch_2.jpg              # Incorrect detection example
│   ├── LabelingExample.png         # Roboflow labeling screenshot
│   └── [LaTeX style files]         # cvpr.cls, cvpr.sty, etc.
│
└── 📁 Templates/                   # Course templates (CS 1430)
    ├── README.md                   # Template instructions
    ├── ProjectFinal_ProposalTemplate.tex
    ├── ProjectFinal_ProgressReportTemplate.tex
    ├── ProjectFinal_ProgressReportTemplate.pdf
    ├── ProjectFinal_PosterTemplate.pptx
    ├── ProjectFinal_PresentationTemplate.pptx
    └── ProjectFinal_ProjectReportTemplate/

---

Notebooks Guide

Notebook	Purpose	Platform
main.ipynb	Data exploration, color matching algorithm, base YOLO detection	Colab
WRYOLOCUSTOM.ipynb	Primary training notebook - YOLOv8 with Roboflow	Colab
player_detection.ipynb	Alternative YOLOv5 training (50 epochs)	Colab
test_vis.ipynb	Model evaluation with NMS + visualization	Colab
newwr_test_vis.ipynb	Same as above, configured for local execution	Local
field_filter.ipynb	Field preprocessing to remove logos/end zones	Colab

Quick Start with Notebooks

# For Google Colab notebooks, upload to Colab and run
# For local notebooks:
cd "WR/notebooks"
jupyter notebook newwr_test_vis.ipynb

Note: Notebooks contain hardcoded paths for Google Drive or local directories. Update the base_dir and data_dir variables to match your setup.

---

Dataset Details

Archive	Contents
NewWRData.zip	503 labeled images split into train/val/test in YOLO format
runs.zip	Training outputs: model weights (best.pt), logs, metrics plots
WR Finder.v1i.yolov8.zip	Roboflow export in YOLOv8 format
WR Finder.v1i.tensorflow.zip	Roboflow export in TensorFlow format

Position Classes (8 total)

ID	Position	Description
0	Corner Back	Defensive backs covering receivers
1	Fullback	Offensive blocker/runner
2	Linebacker	Middle defensive players
3	Quarterback	Offensive play caller
4	Running Back	Primary ball carrier
5	Safety	Deep defensive backs
6	Tight End	Offensive lineman/receiver hybrid
7	Wide Receiver	Primary pass catchers

---

Results

Sample Detections

Performance Metrics

Metric	Value
mAP@0.5	0.759
F1 Score	0.62 @ conf 0.229
Best Class (WR)	94.2% precision
Corner Back	88.5% precision
Safety	87.9% precision

📊 View Detailed Evaluation Charts

F1-Confidence Curve

Precision-Recall Curve

Confusion Matrix

---

Methodology

Two Detection Approaches

1. Color Matching (Experimental)

Located in WR/notebooks/main.ipynb:

• Converts images to LAB color space
• Detects players by team jersey colors
• Identifies offense by formation width
• Finds WRs by edge positioning
• Limitation: Sensitive to lighting, logos, end zones

2. YOLOv8 Object Detection (Primary)

Located in WR/notebooks/WRYOLOCUSTOM.ipynb:

• Single-pass detection for real-time performance
• Grid-based localization with non-max suppression
• Trained on 503 custom-labeled images
• 8 position classes

---

Team Contributions

Member	Contribution
John Michael Slezak	Dataset creation, 503 image labels, YOLO training, report writing
Atif Khan	YOLO research & implementation, model training, technical docs
Chenhao Lu	Detection research, pre-trained YOLO testing, color matching
Ruida Zeng	Field filters, new dataset testing, results analysis, visualization

---

Future Directions

• 🔄 Expanded Dataset — More diverse formations and player setups
• 🤖 Alternative Models — Experiment with DETR or Swin Transformer
• 📹 Player Tracking — In-play movement analysis across frames
• 🎮 Live Integration — Real-time broadcast overlay support

---

Citation

@project{playmakers2024,
  title={The Playmakers: NFL Position Classification Through YOLO},
  author={Slezak, John Michael and Khan, Atif and Lu, Chenhao and Zeng, Ruida},
  institution={Brown University},
  course={CS 1430 Computer Vision},
  year={2024}
}

---

TA Mentor: Joel Manasseh | Professor: Srinath Sridhar

Brown University CS 1430 — Computer Vision Final Project

For questions or suggestions, please reach out via GitHub Issues