eeeeeeeeeeeeee3/soccer-ball-detection

Soccer Ball Detection with RF-DETR

Automated soccer ball detection pipeline using RF-DETR (Roboflow DETR) optimized for tiny object detection (<15 pixels).

Model Details

Architecture

• Model: RF-DETR Base
• Backbone: ResNet-50
• Classes: Ball (single class detection)
• Input Resolution: 1120x1120 (optimized for memory)
• Precision: Mixed Precision (FP16/FP32) training, FP16 inference

Performance

Based on training evaluation report (Epoch 39):

• mAP@0.5:0.95: 0.682 (68.2%)
• mAP@0.5: 0.990 (99.0%)
• Small Objects mAP: 0.598 (59.8%)
• Training Loss: 3.073
• Validation Loss: 3.658

Quick Start

Installation

pip install -r requirements.txt

Usage

from src.perception.local_detector import LocalDetector
from pathlib import Path

# Initialize detector
detector = LocalDetector(
    model_path="models/checkpoints/latest_checkpoint.pth",
    config_path="configs/default.yaml"
)

# Detect ball in image
results = detector.detect(image_path)

Process Video

python main.py --video path/to/video.mp4 --config configs/default.yaml --output data/output

Training

Train from Scratch

python scripts/train_ball.py \
    --config configs/training.yaml \
    --dataset-dir datasets/combined \
    --output-dir models \
    --epochs 50

Resume Training

python scripts/train_ball.py \
    --config configs/resume_20_epochs.yaml \
    --dataset-dir datasets/combined \
    --output-dir models \
    --resume models/checkpoints/latest_checkpoint.pth \
    --epochs 50

Project Structure

soccer_cv_ball/
├── main.py                 # Main orchestrator
├── configs/               # Configuration files
├── src/
│   ├── perception/        # Detection, tracking
│   ├── analysis/         # Event detection
│   ├── visualization/    # Dashboard
│   └── training/         # Training utilities
├── scripts/               # Training and utility scripts
├── models/               # Model checkpoints
└── data/                 # Dataset (not included)

Configuration

Key configuration files:

• configs/training.yaml - Main training configuration
• configs/default.yaml - Inference configuration
• configs/resume_*.yaml - Resume training configurations

Datasets

This model was trained on:

• SoccerSynth-Detection (synthetic data)
• Open Soccer Ball Dataset
• Custom validation sets

Precision Strategy

• Training: Mixed Precision (FP16/FP32) - RF-DETR default
• Inference: FP16 (half precision) for ~3x speedup
• Future: INT8 via QAT (Quantization-Aware Training) for edge devices

Citation

If you use this model, please cite:

@software{soccer_ball_detection,
  title={Soccer Ball Detection with RF-DETR},
  author={Your Name},
  year={2026},
  url={https://huggingface.co/eeeeeeeeeeeeee3/soccer-ball-detection}
}

License

MIT License - See LICENSE file for details.

Acknowledgments

• RF-DETR by Roboflow
• SoccerSynth-Detection dataset
• Open Soccer Ball Dataset