UAV Building Detection

Dataset Access

You can access the data on following link:

https://drive.google.com/file/d/1emLAe7002_syWNxsTO0MgVg4knokFVlQ/view

Research Paper

For detailed methodology and results, please refer to:

https://www.sciencedirect.com/science/article/pii/S1569843222001595

Documentation

Detailed project documentation is available here:

documentation

Tools and Technologies Used

Deep Learning Frameworks

• YOLOv8 (Nano, Small, and X variants)
• YOLOv10 (Nano, Small, and Medium variants)
• Ultralytics Framework

Development Tools

• Python 3.x
• Gradio (for web interface deployment)
• PIL (Python Imaging Library)
• XML parsing libraries (for dataset preprocessing)

Dataset Processing

• Custom preprocessing scripts for:
  • PASCAL VOC to YOLO format conversion
  • Dataset splitting (train/val/test)
• Support for multiple datasets:
  • RescueNet dataset
  • UAVOD-10 dataset

Model Training Infrastructure

• Comet ML (for experiment tracking)
• Custom training configurations:
  • Batch size: 16
  • Epochs: 30
  • Multiple model variants comparison

Deployment

• HuggingFace Spaces (planned)
• Gradio web interface for model inference

Project Structure

.
├── documentation/     # Project documentation
├── preprocess/       # Data preprocessing scripts
├── weights/          # Trained model weights
├── yolos/           # YOLO model configurations
└── gradio_uav_test.py  # Web interface deployment

TODOs:

• Create requirements.txt
• Host the model on HuggingFace Spaces

Model Performance

The project includes comprehensive comparisons between:

• YOLOv8 variants (Nano, Small, X)
• YOLOv10 variants (Nano, Small, Medium)
• Training vs Validation loss analysis
• Output quality comparisons

For detailed performance metrics and comparisons, please refer to the documentation.