Road Segmentation from Satelite images using a custom Unet model

This repository contains code for training and evaluating a custom Unet model to segment roads from satelite images.

Dataset

The dataset can be downloaded from this kaggle link. Some of the images in training set does not have corresponding masks. The training code filters out those images. All the images are of size 1500x1500. Code for data analysis is in this notebook.

Model Architecture

main.py file contains code for training and testing.

E.g.

For training

python main.py train --train_dir=train_dir --model_save_path=path_to_save_model --batch_size=batch_size --epochs=num_epochs

To check all the parameters for training

python main.py train --help

For testing

python main.py test --test_dir=test_dir --checkpoint_path=trained_model_checkpoint_path --save_preds

To check all the parameters for testing

python main.py test --help