veichta / TopoAI

Road segmentation from satellite images using topological approaches.

TopoAI: Topological Approaches for Improved Structural Correctness

Course Project for the Computational Intelligence Lab (CIL) 2023 at ETH Zurich focusing on satelite road segmentation.

Team

Name	Email	Github
Alexander Spiridonov	[e-mail]	aspiridon0v
Alexander Veicht	[e-mail]	veichta
András Strausz	[e-mail]	strausza
Richard Danis	[e-mail]	richdanis

Setup

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
pip install git+https://github.com/bruel-gabrielsson/TopologyLayer.git
pip install -e .

Data

Name	URL	#images
CIL	https://www.kaggle.com/competitions/cil-road-segmentation-2022	144
EPFL	https://www.aicrowd.com/challenges/epfl-ml-road-segmentation	339
RoadTracer	https://paperswithcode.com/dataset/roadtracer	4976

The preprocessed data can be downloaded using the following command:

wget https://polybox.ethz.ch/index.php/s/KhsD19D0iLEmyTH/download -O data.zip

The data is expected to have the following structure:

data
    ├── images
    │   ├── 000000_cil.jpg
    │   ├── 000000_epfl.jpg
    │   ├── 000000_roadtracer.jpg
    │   ├── 000001_cil.jpg
    │   ├── ...
    ├── masks
    │   ├── 000000_cil.png
    │   ├── 000000_epfl.png
    │   ├── 000000_roadtracer.png
    │   ├── 000001_cil.png
    │   ├── ...
    └── weights
        ├── 000000_cil.png
        ├── 000000_epfl.png
        ├── 000000_roadtracer.png
        ├── 000001_cil.png
        ├── ...

This can be achieved by running the following commands:

unzip data.zip

Splitting the Data

In order to split the data into train, val and test sets, run the following command:

python src/preprocessing/split_dataset.py --dataset data

This will create folder for each split containing images, masks and weights folders.

Training

The training can be started by running the following command:

python main.py --datasets <list of dataset> --model <model name> --device <device>

For a full list of arguments, run:

python main.py --help

Reproducing the Baseline

The baseline results can be reproduced by running the following commands:

python main.py --data_path data --datasets cil --epochs 300 --lr 0.001 --model unet++ --patience 40

python main.py --data_path data --datasets cil --epochs 300 --lr 0.001 --model spin --patience 40

python main.py --data_path data --datasets cil --epochs 300 --lr 3e-4 --model upernet-t --patience 40 --miou_weight 1 --focal_weight 1 --mse_weight 1