Playing Card Detection with YOLOv8

Educational computer vision notebook exploring real-time object detection for Hi-Lo blackjack counting practice. Demonstrates PyTorch + YOLOv8 training, OpenCV inference pipelines, and CUDA-accelerated ML fundamentals.

📓 Open card-detection.ipynb to get started

Results

• mAP@0.5: 99.5% on validation set
• Inference: 30+ FPS on RTX 4070
• Dataset: 21,203 training images, 52 classes (13 ranks × 4 suits)
• Model: YOLOv8n (3.2M params, 6MB)

What's Inside

card-detection.ipynb - Comprehensive notebook covering:

1. Dataset Exploration - YOLO annotation format, class distribution, bbox analysis
2. ML Fundamentals - Loss functions (CIoU + BCE + DFL), training loop concepts
3. Model Training - YOLOv8n fine-tuning with data augmentation
4. Inference Pipeline - Manual NMS implementation, OpenCV visualization
5. Hi-Lo Counting - Blackjack counting strategy with state management

Tech Stack

• PyTorch 2.5.1+ (CUDA 12.1)
• YOLOv8 (Ultralytics)
• OpenCV 4.8+
• matplotlib, seaborn (visualization)
• NumPy, pandas (data analysis)

Quick Start

Setup

pip install ultralytics opencv-python torch matplotlib seaborn pandas
jupyter notebook card-detection.ipynb

Requirements: Python 3.10+, CUDA 11.7+ (for GPU acceleration)

Project Structure

card-detection/
├── card-detection.ipynb              # Main educational notebook
├── datasets/
│   ├── train/                        # 21,203 training images
│   ├── valid/                        # 2,020 validation images
│   └── data.yaml                     # Dataset config (52 classes)
└── backend/runs/card_detection/
    └── weights/best.pt               # Fine-tuned model weights

Key Concepts Covered

Dataset & Annotations

• YOLO annotation format (normalized bounding boxes)
• Class distribution analysis and balancing
• Bbox aspect ratio and area statistics
• Data augmentation (mosaic, mixup, HSV jitter)

Training Pipeline

• Transfer learning from COCO pretrained weights
• Multi-component loss function: 7.5×CIoU + 0.5×BCE + 1.5×DFL
• AdamW optimizer with learning rate scheduling
• Training metrics visualization and interpretation

Inference & Post-Processing

• Manual NMS (Non-Maximum Suppression) implementation
• Confidence threshold tuning (0.5-0.7 recommended)
• OpenCV visualization pipeline
• Real-time performance optimization

Hi-Lo Card Counting

• Blackjack counting strategy implementation
• State management (running count, seen cards)
• Duplicate detection prevention

Card Range	Value	Strategy
2-6	+1	Favors player
7-9	0	Neutral
10-A	-1	Favors dealer

Performance

Metrics (after 5 epochs):

• Precision: 99.36%
• Recall: 99.51%
• mAP@0.5: 99.5%
• mAP@0.5:0.95: 70.77%

Inference Speed (RTX 4070):

• 30ms/frame (33 FPS)
• ~2GB GPU memory

Dataset

Roboflow Playing Cards Dataset

• 21,203 training images
• 2,020 validation images
• 52 classes (13 ranks × 4 suits)
• Well-balanced: ~400 instances per class

Learning Resources

• YOLOv8 Documentation
• OpenCV Python Tutorials
• Object Detection Metrics Explained
• Hi-Lo Card Counting Strategy

Author

Raghunath Prabhakar

Acknowledgments

• Ultralytics - YOLOv8 framework
• Roboflow - Playing Cards dataset
• OpenCV Community

---

Educational project for hands-on computer vision learning MIT License