Change Detection

Problem Statement

The objective is to detect changes between two images in a time series. Both input images share the same dimensions and represent the same location. The desired output is a binary image highlighting the areas of change, which could entail the introduction of new objects, the absence of previously present objects, or alterations in object positions.

Deep Learning Approach

Prerequisites

• Python 3
• PyTorch
• torchsummary
• tqdm
• OpenCV
• scikit-learn
• NumPy
• Matplotlib
• torchvision
• Jupyter Notebook

Setup with Conda

conda create -n change-detection python=3.8
conda activate change-detection
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia -y
conda install tqdm numpy matplotlib scikit-learn jupyter -y
pip install opencv-python -y
pip install torchsummary -y

Models

• Basic Unet (takes the difference between the two images -- before and after -- as input)
• Diff UNet (with Diffence Between input before and after)
• Siamese Nested UNet (UNet++)

Dataset

Our dataset consists of 4868 samples captured from Egyptian lands before and after a certain period. We partitioned the dataset into 80% for training and 20% for validation purposes.

Hyper Parameters

• Learning Rate: 0.001
• Learning Rate: StepLR with step size 10 and gamma 0.2
• Batch Size: 16
• Epochs: 50
• Optimizer: Adam
• Loss Function: BCEWithLogitsLoss

Results Comparison

Model	Training Jaccard Score	Validation Jaccard Score
Basic UNet	0.85	0.77
Diff UNet	0.86	0.79
Siamese Nested UNet (UNet++)	0.88	0.81
Siamese Nested UNet (UNet++) with all dataset	0.92	-

Testing

To test any model, use Predict.ipynb notebook and modify the following :

1. Change the model path
2. Import the model from models directory
3. Change the testset path (if needed)
4. Run the notebook

Collaborators 🤝

Bemoi Erian	Peter Atef	Mark Yasser	Karim Mahmoud