Road segmentation

Description

This project is part of the ML course at EPFL. The goal is to develop a classifierthat successfully classifies roads and background segments in satellite images. We performd data preprocessing and developed two models, a Convolutional NeuralNetworks (CNN) model and a U-Net model then compared them. An F1-score of 0.85 was achieved using the U-net model on the train set and a F1-score of 0.886 on the test set.

Instructions

We provide the requirements.txt file which allows to install dependencies as follows:

$ pip install -r requirements.txt

You can run the models as follows:

git clone <repo_url>  #clone the repo
cd src
python run.py #for Unet architecture (The one used for the submission)
#python run.py -cnn  # for cnn architecutre

#use the --help to see the arguments that you can enter

Dataset

The dataset is available from the https://www.aicrowd.com/challenges/epfl-ml-road-segmentation page It contains:

• training set: 100 RGB images of dimension (400,400,3) with 100 ground truth images (black and white) of dimensions (400,400,1)
• test set: 50 RGB images of dimension (608,608,3)

include the dataset in the "data" folder

Project structure

├── data                            # data directory
|   ├── training
|       |__ groundtruth             # Directory containing the train groundtruth
|       |__ images                  # Directory containing the train images
|   ├── test
|       |__ images                  # Directory containing the test images
├── src                			        # Source code directory
|   ├── utils
|       |__ data_augmentation.py    # file where we perform data augmentation
|       |__ load_helpers.py         # Helper functions used in our project
|       |__ metrics.py              # file where we define the metrics (F1 score, precision, recall)
|       |__ preprocess_data.py      # helper functions for preprocessing data 
│   ├── run.py                      # script to run on terminal and run the models or train new ones
│   ├── make_submission.py          # file that contains one function called to make predictions
│   ├── model.py                    # abstract class for the models
│   ├── Unet.py                     # file where we code our Unet model
│   ├── cnn.py                      # file where we code our CNN model
├── plots              	            # Directory containing plots of metrics after running models
├── history              	          # Directory giving th evolution of the metrics of models
├── submissions          	          # Directory csv documents of main submisisons we did on Aicrowd
├── trained models                  # Directory containing some already build and trained models
├── requirements.txt			          # Contains requirements to be installed using pip
├── project2_description.pdf			  # pdf file of project description
├── requirements.txt			          # Contains requirements to be installed before running the project
├── LICENSE 
├── report.pdf			                # Contains the report of our project
└── README.md

Note: We didn't include the data folder but we kept it in the architecture to show how it should be when you add the data.

Trained models:

We included only 3 of our trained models as examples:

• 'cnn_new_aug_200_16_leaky': CNN model, Leaky-Relu, 200 epochs,
• 'unet_new_augg_400_batch4_300ep': Unet model, Leaky-Relu , 300 epochs, batch size: 4, image size: 400
• 'unet_new_augg_608_batch4_300ep.csv': Unet model, Leaky-Relu , 300 epochs, batch size: 4, image size: 608

Note: The 'unet_new_augg_608_batch4_300ep.csv' is the the one that gave the best results and with whom the submission was made.

Note We included the history, submission and plots for these models

Authors

Zeineb Ayadi
Selim Fekih
Nizar Ghandri