This project was conducted by :
- Nicolas Francio
- Louis Leger
- Daphné de Quatrebarbes
This project focuses on segmenting roads from aerial images using various algorithms and optimization techniques. This README provides essential information on file organization and usage to ensure a smooth experience.
The data sets required to train and test the prediction models we have implemented can be found on [AIcrowd](https://www.aicrowd.com/challenges/epfl-machine-learning-project-1) Functions for submission are provided and located in the file import. Note To run our code, please download these files and put them in the same folder as our code files.
The scripts file is composed of python files defining our models, training methods, preprocessing, evaluation metrics and methods, data augmentation and imports.
Our best model was proven to be dine utuning a pre trained DeepLab model with AI crowd submission of 0.907. It is located in the notebook experiments_deeplab.ipynb.
A 4 page scientific report describes the most relevant feature engineering techniques and implementations that we worked on, explains how and why these techniques improved our predictions and includes an ethical consideration associated to this machine learining problem.
To run the run.py and obtain our best prediction the user would need to have the following necessary libraries, as our model is too heavy to upload to github so we made an accessible link through google drive and necessitates this library and 160MB to store the model: import gdown
import torch import torchvision from torchvision import models from torch import nn import torch.optim as optim from torch.utils.data import DataLoader, Dataset import os from PIL import Image
from data import mask_to_submission import cv2 as cv import matplotlib.pyplot as plt import numpy as np from skimage import color import os import seaborn as sns import pandas as pd
import os import numpy as np import matplotlib.image as mpimg import re