EPFL Machine Learning - Road Segmentation 2019

About the Project

In this project, we implemented and trained using satellite images, a residual Unet for classifying pixels on an aerial image as either a road or background.

Contributors

• Lucien Barret @lubacien
• Dorian Popvic@DorianPopovic
• Théophile Bouiller@tbouiller
• AiCrowd team name : theophileetlucienetdorian
• AiCrowd best submission id : 31607 theophile_bouiller (890)

Setup environment

The necessary can be installed using pip and the provided requirements.txt

   pip install -r requirements.txt

In order to run tf_aerial_images tensorflow must be downgraded. In order to get the correct version for tensorflow execute

   pip install --upgrade tensorflow==1.13

Code architecture

In this section we explain how our project folder is organised and where to find the needed files.

Data Folder

This folder contains the train and test data in zip archives. In order to run the code the dataset consisting of training.zip and test_set_images.zip must be placed in the data/ folder, due to limits to the submission size. Executing run.py will unzip these files and make the the following directories containing the files:

1. training : consists of the images and associated groundtruths.
2. test_set_images : Consists of the images we use to test our model.

Scripts folder

All our project implementations can be found inside the /scripts folder.$

Python executables .py

1. run.py: Generates the predicted ground truth using the weights that gave us our best prediction on Aicrowd.
2. train_unet.py: Trains the unet and creates new weights.
3. preprocessing.py: Functions that help us preprocess the data.
4. image_processing.py: Applies modifiers to the the output images.

Run

To generate our best prediction submitted on aicrowd, execute:

   python3 run.py

from the root directory of the project. This will generate the predictions using the given weights. The predictions are saved in the folder /submissionimages, a csv submission file will be created under the name submission.csv in the data/ directory.

To train the model and build the weights, execute:

   python3 train_unet.py

Process the output images

In order to execute a floodfill on the prediction execute:

   python3 image_processing.py

The floodfilled images will be saved in a new folder named sub_mod/. a new folder named mont_dir/ will contain the grouped images from the test set, predictions and floodfilled images.