Machine Learning Project 2

The overall goal of this project is perform an image segmentation on aerial/satellite image to find the roads.

Project overview

This project aims to develop a machine learning model to predict the likelihood of a pixel in an image to be a road or not. Based on that prediction we will attribute a label to each pixel of a satellite image. With road beeing set to white and non-road set to black (road=1, background=0).

Example :

How to Run the Code

You may want to run the code with UNet_pred.ipynb.

1. Clone the repository
   Open your terminal and run the following command to clone this repository:

   cd YOUR_CLONING_DIRECTORY
   git clone https://github.com/CS-433/ml-project-2-notaname_p2
   cd ml-project-2-notaname_p2

2. Get the environment and activate it

   conda env create -f environment.yml
   conda activate ML_project2
   

3. Get pretrained Models

   You will need to redownload the UNet_model.pth and RC_params_opti.pth files and place them in the project folder. If you don't you might enconter an error stating the files are corrupted. By using a pretrained model the prediction time would be around 10 minutes, if you want to train the model your self it might take several hours.

4. Get the best prediction

   open UNet_pred.ipynb on any editor, selected the ML_project2 enviromnent and run all cells

Best Prediction and Getting other Predictions

To get best prediction that we obtained, run UNet_pred.ipynb with all parameters set to False (already done by default).

If you want to predict best parameters (best number of layers/bases,...), set UNET_SEARCH to True.

If you want in addition to save them, set both SAVE_UNET and UNET_TRAIN to True. For the hyperparameter search for postprocessing part, set ROAD_CORRECTION_SEARCH to True (to save them set SAVE_RC_PARAMS to True).

The last parameter THRESHOLD_SEARCH is for searching optimal threshold to minimize F1 loss.

Authors: Joana Pires, Leonardo Tredici, Antonin HUDRY