Team: OGP_team - Project 2 - Road Segmentation

Project2 - CS-433 - Machine Learning, EPFL (Fall 2021)

Team membbers of "OGP_team":

• Ottander Gustav Olle Christofilus (gustav.ottander@epfl.ch)
• Karlbom Karl Johan Gustav (karl.karlbom@epfl.ch)
• Celada Paolo (paolo.celada@epfl.ch)

Abstract

Image semantic segmentation is the process of assigning a label to each pixel in an image. It has applications in various different fields, from medical images to object detection and task recognition. Using a set of aerial images from urban areas, the aim of this project is to build a classifier able to perform such semantic segmentation, by assigning a label (road=1, background=0) to each pixel. After initial research, where multiple models were considered, a CNN network architecture called U-net was chosen, since it was the best suitable choice for the task. The U-net model is supported by preprocessing, specifically image augmentation, normalization and post-processing. The training phases considered multiple parameters and implementation choices. Once a valid model is setup, predictions on a test set are calculated and submitted to the online platform \textit{AIcrowd}~\cite{AIcrowd-RoadSegmentation}, which calculates both F1 score and accuracy.

Setup and run

To get a hold of the dataset download the files test_set_images.zip and training.zip HERE. Unzip the folders, and place them in a folder 'data'.

Before running the file, make sure to install the packages below.

• os
• matplotlib
• cv2
• os
• re
• skimage
• tqdm
• torch
• torchvision
• tensorflow as tf
• keras
• keras.backend as K
• sys

We had last minute problems with the requirements.txt. If possible, we will provide it later.

To obtain the best result at AIcrowd, run run.py as:

python3 run.py arg

where arg can be one of the following:

• msl4 --> for mean squared loss with 4 features
• msl8 --> for mean squared loss with 8 features
• msl16 --> for mean squared loss with 16 features
• msl32 --> for mean squared loss with 32 features
• focal --> for focal loss with 16 features

File structure:

The main folder should have the following structure:

.
├── data/
│   ├── test_set_images/
│   └── training/
│
├── weights/
│
├── README.md
├── data_postprocessing.py
├── save_output.py
├── image_augmentation.py
├── model.py
├── loss_functions.py
├── data_loading.py
├── metrics.py
├── data_prep.py
├── data_handling.py
├── requirements.txt
├── run.py 
├── main_notebook.ipynb

File names are self-explanatory, but a detailed explanation is found inside