Skip to content

We built and trained a CNN U-Net in Pytorch to perform semantic segmentation of roads on satellite images.

License

Notifications You must be signed in to change notification settings

LucasBrazCappelo/ML_EPFL_Project_2

Repository files navigation

Machine Learning CS-433 - Class Project 2 - Road Segmentation - EPFL

We built and trained a Convolutional Neural Network in Pytorch to perform semantic segmentation of roads on satellite images. We implemented a U-Net with padding inspired by Ronneberger et al. (2015). We did Data augmentation by rotating (45, 90, 135, 180, 225, 270, 315) or flip image.

Our submission files are saved in the folder submissions.

The train and evaluation contains 100 train images of size 400x400 and 50 evaluation images of size 608x608.

You can consult our report for more details.

Contributors

Our U-Net graph:

See some of our results:

Setup Environment

We run on Windows 10 with Conda 4.11 (Python 3.8) installed and Pytorch 1.10. It should work with any 3.6+ Python version with pip.

pip3 install torch==1.10.1+cu113 torchvision==0.11.2+cu113 torchaudio===0.10.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
pip3 install numpy
pip3 install sklearn
pip3 install scipy
pip3 install skimage
pip3 install matplotlib
pip3 install tqdm
pip3 install glob
pip3 install Pillow

Run

To generate our final AIcrowd submission execute:

./run.py

It should take around 1 hour and half maximum with a recent GPU.

About

We built and trained a CNN U-Net in Pytorch to perform semantic segmentation of roads on satellite images.

Resources

License

Stars

Watchers

Forks