ML_project_EPFL

• Semantic road segmentation project for CS-433 Machine Learning at EPFL
• Prediction model for semantic road segmentation of satellite images created by Team "mediterranean_OS"

Table of Contents

1. Introduction
2. Datasets
3. Setup
4. Project Structure
5. Usage
6. Authors

Introduction

This project implements a UNet-based deep learning model to address the problem of semantic road segmentation. Semantic segmentation is critical for applications such as autonomous driving and urban planning, enabling precise identification of road regions in images. Apart from the basic UNet architecture, UNet3+ and a RESNet34 backbone have been added to compare the performance of predictions against the base implementation.
For this project, we used an already implemented UNet3+: Link to original repository

Datasets

The dataset used for training is made up of 100 different satellite images (400x400px) of american style suburbs, along with their respective ground truth images highlighting the roads that are used during the supervised learning. For testing, the dataset is comprised by 50 satellite images (608x608px) of similar american style suburbs. Additionally, a selected portion of images from the Massachussets dataset has been added to 'training_extra', to improve fairness.

Setup

• Clone the repository with git clone https://github.com/CS-433/ml-project-2-mediterranean_os.git
• This project uses Pytorch with CUDA, in case you installed Pytorch without CUDA enabled, you will need to reinstall Pytorch following this short guide

Project Structure

The train and test images have been uploaded to the repository due to only being 100 for training (plus respective ground truth images), and 50 for testing. Keeping the images provides an easier setup as the repository already comes with a suited way to train the models since the get go.

├───resources
├───src
│   └───models
├───test_set_images
├───training
│   ├───groundtruth
│   └───images
└───training_extra
    ├───groundtruth
    └───images

Usage

Hyperparameter tuning: Multiple parameters in train.py can be adjusted to fine tune the performance of the model. These parameters are:

• BATCH_SIZE: the size of the mini-batches used by the model during training, smaller mini-batches reduce computational cost but can impact performance.
• NUM_EPOCHS: number of training cycles the model will go through.
• LEARNING_RATE: controls the step size at which the optimization algorithm updates model weights during training.
• MASSACHUSSETS: It is a boolean. When true, the model is trained with an added 46 pictures taken randomly from the Massachussets dataset, which are located in the training_extra directory.

Choosing different models: Different models can be trained and later used to make predictions, the available models can be found in /models. Changing the model can be done by importing the model in train.py and using it at the top of the main() function in train.py.:

• UNet: UNet description
• UNet3+: description
• RESNet34 backbone: both models can be run with a RESNet34 backbone to improve performance at the cost of more computation required.

Running the program:

• Run train.py and wait until training finishes, the current epoch is shown as well as bars to track progress. When it finishes it will save the model model_state_dict.pt.
• Run predict.py to test the model, it will generate test.csv with the predicted pixels and the folder /submissions with the predicted images.

Authors

Student's name	SCIPER
Leonardo Denovellis	386074
Luis Bustamante	366705
Antonio Silvestri	376875