Siamese U-Net for Building Damage Assessment

This project implements a Siamese U-Net for pixel-wise building damage classification using the xView2 dataset.
It supports training, evaluation, and a FastAPI backend for inference.

        

Architecture used: Siamese-Unet with resent34 encoder

NOTE: In this implementation i had used difference of feature map unlike this image which uses concatenation, due to resource constraints.

Image source: Research Gate

---

Features

• Siamese U-Net with dual encoders (shared weights)
• Pixel-wise segmentation (5 damage classes)
• Supports Dice + CrossEntropy loss
• Class balancing & metrics (mIoU, Dice, Pixel Acc)
• Test-time evaluation script with CSV report
• FastAPI backend
• Docker-ready backend

---

Project Structure

.
├── app/
│ ├── main.py # FastAPI app
│ ├── model.py # Load model + predict helpers
│ ├── utils.py # Color masks, overlay helpers
├── model_architecture.py # SiameseUNet definition
├── train_siamese_unet.py # Training script
├── test_script.py # Test/eval script
├── requirements.txt
├── Dockerfile
└── README.md

---

Requirements

# Create a virtual env (optional but recommended)
conda create -n damage-seg python=3.10
conda activate damage-seg

# Install dependencies
pip install -r requirements.txt

Training

Train your Siamese U-Net:

python train_siamese_unet.py

The model weights will be saved as .pth.

Evaluation

Run on your test dataset:

python test_script.py

Outputs:

Predicted masks: ./test/predicted_masks/

CSV report: ./test/metrics_report.csv

Metrics: mIoU, Dice, per-class IoU, per-class Pixel Accuracy

Example Evaluation Results

Metric	Value
Mean IoU	0.6008
Mean Dice	0.6298
Overall Pixel Accuracy	0.9551

Per-class:

Class	IoU	Dice	Pixel Acc
Class 0	0.9659	0.9813	0.9702
Class 1	0.4664	0.5455	0.9666
Class 2	0.3264	0.3456	0.9816
Class 3	0.5744	0.5863	0.9941
Class 4	0.6707	0.6902	0.9977

These numbers are from the test set. They show strong segmentation of undamaged areas and reasonable damage class detection. There is scope to further improve damage class performance with more data, augmentations, and advanced training strategies. Changes to the architeceture like concatenation of feature maps before decode instead of difference, will greatly help!

---

FastAPI Inference

Run the backend:

uvicorn app.main:app --host 0.0.0.0 --port 8000

Test the /predict endpoint:

Use Swagger UI: http://127.0.0.1:8000/docs

Or Postman / curl with multipart/form-data for pre_disaster and post_disaster images.

License

MIT — free for research and personal projects.

Acknowledgements

xView2 Dataset

segmentation-models-pytorch