HALO

This is the official PyTorch implementation of the ICML 2024 paper Hyperbolic Active Learning for Semantic Segmentation under Domain Shift.

Luca Franco ^{† 1}, Paolo Mandica ^{† 2}, Konstantinos Kallidromitis ³, Devin Guillory⁴, Yu-Teng Li⁴, Trevor Darrell⁴, Fabio Galasso^1,2
1 ItalAI,² Sapienza University of Rome, ³ Panasonic North America, ⁴ UC Berkeley
^† Equal contribution

[arXiv][BibTeX][OpenReview]

---

Usage

You can choose to run the code in a conda environment or in a docker container. Follow the instructions below to set up your preferred environment.

Conda env setup

conda create --name halo -y python=3.9
conda activate halo
pip install -r requirements.txt

Docker env setup

Modify the docker-compose.yml file to set the correct number of available GPUs. Then set your correct path for the DATA_VOLUME and run the following commands:

export UID=$(id -u)
export GID=$(id -g)
export DATA_VOLUME="YOUR_PATH_TO_DATA_FOLDER"

docker compose build
docker compose up -d

Data Preparation

1. Download the Cityscapes Dataset, GTAV Dataset, SYNTHIA Dataset, ACDC Dataset and extract them.

2. If you downloaded the datasets in a different folder, symlink them to the datasets directory:

   ln -s /path_to_cityscapes_dataset datasets/cityscapes
   ln -s /path_to_gtav_dataset datasets/gtav
   ln -s /path_to_synthia_dataset datasets/synthia
   ln -s /path_to_acdc_dataset datasets/acdc

3. Generate the label static files for GTAV/SYNTHIA Datasets by running

   python datasets/generate_gtav_label_info.py -d datasets/gtav -o datasets/gtav/
   python datasets/generate_synthia_label_info.py -d datasets/synthia -o datasets/synthia/

The datasets directory tree should be structured as follows:

├── datasets/
│   ├── cityscapes/     
|   |   ├── gtFine/
|   |   ├── leftImg8bit/
│   ├── gtav/
|   |   ├── images/
|   |   ├── labels/
|   |   ├── gtav_label_info.p
│   └──	synthia
|   |   ├── RAND_CITYSCAPES/
|   |   ├── synthia_label_info.p
│   └──	acdc
|   |   ├── images/
|   |   ├── gt/

Training

The config files for the different training protocols can be found in the configs directory.

Before running the training/testing scripts, make sure to set the correct paths in the config files.

python train.py -cfg CONFIG_PATH

Testing

python test.py -cfg CONFIG_PATH

Training with ACDC Dataset or SegFormer architecture

To train with the ACDC dataset or the SegFormer architecture, you need to switch to the corresponding branch:

git checkout acdc

or

git checkout segformer

The two branches will be merged into the main branch soon.

Acknowledgements

This project is based on the RIPU open-source project, with our re-implementation in PyTorch Lightning for multi-gpu active learning. We thank the authors for making the source code publically available.

Licence

This project is licensed under the terms of the MIT license.

Citation

If you find this repository useful, please consider giving a star ⭐ and citation:

@inproceedings{
    franco2024hyperbolic,
    title={Hyperbolic Active Learning for Semantic Segmentation under Domain Shift},
    author={Luca Franco and Paolo Mandica and Konstantinos Kallidromitis and Devin Guillory and Yu-Teng Li and Trevor Darrell and Fabio Galasso},
    booktitle={Forty-first International Conference on Machine Learning},
    year={2024},
    url={https://openreview.net/forum?id=hKdJPMQvew}
}