This project develops a deep learning-based model to recover thin skeleton representations from thick, noisy road network images generated from OpenStreetMap (OSM) data.
We handle ambiguities inherent in skeletonization while leveraging advanced techniques such as U-Net with residual blocks, iterative thinning, and multi-task learning.

This work is completed for CSE627: Advance Machine Learning, following project guidelines.

road_skeletonization/
├── datasets/
│   └── road_skeleton_dataset.py       # Dataset and collate function
├── engine/
│   ├── trainer.py                      # Trainer class
│   └── tester.py                       # Tester class
├── models/
│   ├── skeleton_unet.py                # U-Net model with residual blocks
│   └── loss.py                         # Weighted Focal Loss
├── transforms/
│   └── transforms.py                   # Data augmentation transforms
├── utils/
│   └── dataloader.py                   # Data splitting and loader functions
├── config.py                           # Configurations (paths, hyperparameters)
├── train.py                            # Training script
├── test.py                             # Testing and evaluation script
├── README.md                           # Project description
└── requirements.txt                    # (optional) Dependencies list

1. Clone the repository and navigate into the project folder.

git clone https://github.com/Nadim-Mahmud/Road-Skeletonization
cd Road-Skeletonization

2. Install required Python packages:

pip install torch torchvision torchmetrics matplotlib pillow

(Optional) Create and activate a virtual environment before installing.

3. Update dataset path: Edit config.py and set:

root_path = 'thinning_data/data'

Run the training script:

python train.py

This will:

• Train the model on the training set
• Validate it on the validation set
• Automatically save the best model as best_model.pth

Training uses early stopping if validation performance does not improve.

After training completes, run:

python test.py

This will:

• Evaluate the best model on the test set
• Print Accuracy, Precision, Recall, and F1 Score
• Save sample visualizations to test_samples.png

• OSM Data Extraction: Selected a subset of road types from OSM and retrieved the corresponding data.
• Rasterization: Rendered roads into 256×256 grayscale images with thickness based on attributes (e.g., lane count, width).
• Distortions: Introduced realistic distortions (blur, noise) to simulate real-world imperfections.
• Ground Truth: Generated 1-pixel thick skeletons from road centerlines.

• Baseline: U-Net architecture to predict thin skeleton from thick, noisy input.
• Advanced Option (choose one):
  • Iterative Thinning: Simulate gradual de-thickening.
  • Multi-Task Learning: Incorporate auxiliary tasks (e.g., distance maps).
  • Training on Misaligned Data: Handle skeleton misalignment with custom losses or post-processing.

• Verification: Conducted unit tests, visual inspection, and controlled experiments with known outcomes.
• Correctness: Ensured every pipeline component (data, network, metrics) works correctly through intermediate inspections.

• Qualitative Results: Visual samples showing input, ground truth, and model predictions.
• Quantitative Metrics:
  • Test Loss
  • Mean Squared Error (MSE) using distance transforms
  • Node Precision and Recall for 1-valent to 4-valent nodes
  • (Optional) IoU, Dice Coefficient

Performed ablation study by varying:

• Learning rate
• Loss functions
• Architectural choices

Reported results in tables and justified the final chosen configuration.

• Project organized as Python scripts runnable from command line or Google Colab via %run script.py.
• GPU support compatible.
• Code must be linked but is not graded — only the slides are graded.
• Slides must display final qualitative/quantitative results, methodology, and ablation studies.

Metrics tracked:

• Accuracy
• Precision
• Recall
• F1 Score
• MSE
• Node-based precision and recall

Visualization:

• Side-by-side comparisons of Input, Ground Truth, and Prediction images saved in test_samples.png.