Mars Terrain: a CV Segmentation task

Project for the Artificial Neural Networks and Deep Learning course held at Politecnico di Milano in the academic year 2024-2025 by Professor Boracchi and Professor Matteucci.

🏆 This project received a perfect evaluation of 5.5/5.5.

Table of contents

• Installing
  • Prerequisites
  • Cloning the Repository
  • Installing the Libraries
• Project
• Results
• Structure
• Authors

Installing

Prerequisites

• Python 3.x
• Pip (for installing Python libraries)

Cloning the Repository

1. Clone the repository to your local machine using the following command:

git clone https://github.com/simone-licciardi/anndl-hw2

2. Navigate to the project directory using the following command:

cd anndl-hw2

Installing the Libraries

With Pip, install the required libraries by running the following command:

pip install -r requirements.txt

Project

Objective: segment Mars terrain images into five classes: Background, Soil, Bedrock, Sand, and Big Rock.

Dataset: 2,615 grayscale images (64x128 resolution), filtered to 2,505 images.

Methodology:

• Built an initial encoder-decoder architecture as benchmark.
• Added layer for Egde Detetions, Thesholding and others methods of computer vision
• Implemented a dual UNet architecture (Global and Local perspectives).
• Designed custom loss functions and loss schedules.
• Applied data augmentation and fine-tuned optimization hyperparameters.

Structure

The repository source contains the notebook to train the final model model.ipynb, and a folder to preprocess the dataset preprocessing. The latter contains two notebooks:

• filtration.ipynb: containing the code we used to inspect images and labels and to remove the images containing aliens and their labels.
• augmentation.ipynb: containing the code we used to implement data augmentation (simple geometric transformations and image/mask blending).

The repository archive is an history of our the most relevant notebooks to understand the development of our final solution and its elements are detailed in the report.

Results

Our model achieved a 64.91% test set Mean IoU and perfect evaluation. You can see the ranking here. You can check out the final report.pdf.

Authors

• Noemi Bongiorni (@NoemiBongiorni)
• Simone Licciardi (@simone-licciardi)
• Alessandro Pedone (@alessandropedone)
• Federico Maria Riva (@fede-mat)