YOLOv11n Object Detection Model Fine-tuning

This repository demonstrates how to fine-tune YOLOv11n on multiple fire detection datasets. It provides a complete pipeline for combining multiple datasets from Roboflow, training a unified model, and evaluating its performance.

Features

• Multiple dataset combination and preprocessing
• YOLOv11n model fine-tuning
• Comprehensive evaluation metrics
• Performance visualization tools
• Support for various image formats

Prerequisites

1. Clone the Repository

git clone https://github.com/prakash-aryan/yolov11n-fire-detection.git
cd yolov11n-fire-detection

2. Set Up Python Environment

# Create virtual environment
python -m venv venv

# Activate virtual environment
# On Windows:
venv\Scripts\activate
# On Linux/macOS:
source venv/bin/activate

3. Install Required Packages

# Upgrade pip
python -m pip install --upgrade pip

# Install packages
pip install ultralytics pandas matplotlib seaborn tqdm
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118  # For CUDA 11.8
# OR for CPU only
# pip install torch torchvision

4. Verify Installation

python -c "import torch; print('CUDA Available:', torch.cuda.is_available())"

Note: Always activate the virtual environment before working on the project:

# Windows: venv\Scripts\activate
# Linux/macOS: source venv/bin/activate

Dataset Collection

1. Getting Roboflow Datasets

1. Create an account on Roboflow
2. Create or find fire detection datasets
3. For each dataset:
   • Go to export page
   • Select "YOLOv11" format (this format is compatible with YOLOv11n)
   • Get the download URL:

   https://universe.roboflow.com/ds/XXXXX?key=YYYYY
   

2. Configure Dataset Downloads

1. Edit prepare_datasets.sh
2. Replace placeholder URLs with your Roboflow URLs:

download_and_extract "YOUR_ROBOFLOW_URL_1" "roboflow1.zip" 1
download_and_extract "YOUR_ROBOFLOW_URL_2" "roboflow2.zip" 2

Project Structure

├── prepare_datasets.sh    # Dataset download and organization
├── prepare_dataset.py     # Combines dataset configurations
├── train_model.py         # YOLOv11n training script
├── evaluate_model.py      # Model evaluation
├── visualize_results.py   # Results visualization
└── datasets/             # Downloaded and processed datasets
    ├── train/
    ├── valid/
    ├── test/
    └── combined_data.yaml

Usage Pipeline

1. Download and Prepare Datasets

# Make script executable
chmod +x prepare_datasets.sh

# Run dataset preparation (after adding your URLs)
./prepare_datasets.sh

2. Combine Dataset Configurations

python prepare_dataset.py

This creates a unified combined_data.yaml with:

• Combined class names
• Updated paths
• Merged dataset configurations

3. Train YOLOv11n Model

python train_model.py

Key training parameters:

parameters = {
    'model': 'yolo11n.pt',  # Base YOLOv11n model
    'epochs': 100,
    'batch_size': 64,
    'imgsz': 640,
    'lr0': 0.001,
    'lrf': 0.01,
    'weight_decay': 0.0005,
    'warmup_epochs': 3,
    'device': 0  # GPU device id
}

4. Evaluate Model

python evaluate_model.py

This generates:

• Precision, recall metrics
• Confusion matrix
• mAP50 and mAP50-95 scores
• Training curves

5. Visualize Results

python visualize_results.py

Model Architecture

We use YOLOv11n as our base model, which offers:

• Efficient architecture for edge devices
• Good balance of speed and accuracy
• Suitable for real-time fire detection

Training Details

The training process includes:

1. Loading pretrained YOLOv11n weights
2. Fine-tuning on combined fire detection datasets
3. Automatic learning rate scheduling
4. Multi-scale training
5. Regular checkpointing

Performance Metrics

The evaluation provides:

• Precision and Recall curves
• Confusion matrix
• mAP (mean Average Precision)
• Inference speed metrics
• Per-class performance analysis