2Player: A General Framework for Self-Supervised Change Detection via Cooperative Learning

This repository contains the pytorch implementation of the 2Player model from the paper:

2Player: A General Framework for Self-Supervised Change Detection via Cooperative Learning
📄 ISPRS Journal of Photogrammetry and Remote Sensing
🔗 Read the paper 🌐 Project page

⚙️ Requirements

This project is entirely written in Python, primarily using the PyTorch library. All required dependencies can be installed using:

conda env create -f environment.yaml

💾 Datasets

This repository supports four datasets used in the paper: two new or improved datasets introduced in this work (HRSCD-Clean and ValaisCD), and two standard benchmarks (LEVIR-CD and WHU-CD).

🆕 HRSCD-Clean (introduced in this paper)

We provide a cleaned and refined version of the original HRSCD dataset.

• Download: HRSCD-Clean
• Setup: Simply download the dataset and place it under datasets/HRSCD_clean

Pruned subsets: We provide multiple pruned versions of HRSCD to mitigate label noise. Each pruned split is defined by a .pkl file listing the retained samples and located in the corresponding folder. When creating the dataset, you can choose the pruned version by specifying the desired .pkl file with the given_list parameter. The 10,000-sample pruned version is the one used in all experiments reported in the paper.

🆕 ValaisCD (introduced in this paper)

ValaisCD is a newly introduced high-resolution change detection dataset built from SwissTopo SWISSIMAGE aerial imagery and SwissTLM3D vector maps.

• Download: ValaisCD
• Setup: Simply download the dataset and place it under datasets/ValaisCD

Pruned subsets: As with HRSCD, we provide several pruned versions of ValaisCD, each defined by a .pkl file listing the retained samples. When calling the dataset, you can choose the pruned version by specifying the desired .pkl file with the given_list parameter. The 5,000-sample pruned version is the one used in the paper’s experiments.

⭐ LEVIR-CD (public dataset)

• Download: LEVIR-CD
• Setup: Place the dataset under: datasets/LEVIRCD

⭐ WHU-CD (public dataset)

• Download: WHU-CD
• Additional preprocessing required: WHU-CD must be split into non-overlapping 256×256 tiles before use:
  • Download the dataset from the link above.
  • Tile images into 256×256 patches with no overlap.
  • Place the resulting tiles under datasets/WHUCD

🏋️ Training 2Player

Training the full 2Player framework involves four main steps:

• Compute dissimilarity maps
• Pretrain Player 2 (reconstruction network)
• Pretrain Player 1 (change detector)
• Joint training of the 2 Players

Below are the detailed commands and expected outputs for each step.

1️⃣ Generating Dissimilarity Maps

Before training, you must generate the dissimilarity maps from the Geographical Correspondence Module by running the following command:

python compute_probability_of_change_maps.py --dataset <HRSCD or Valais or LEVIR or WHUCD>

This script creates dissimilarity maps for every dataset split and stores them under:

datasets/<DATASET_NAME>/<split>/probability/

2️⃣ Pretrain Player 2 (Reconstruction Network)

Player 2 is first trained independently for a few epochs to reconstruct the image at $t_1$ from the image at $t_0$. Run:

python pretrain_Player2.py --dataset <HRSCD or Valais or LEVIR or WHUCD>

This will save a pretrained checkpoint under:

checkpoints/player2/

3️⃣ Pretrain Player 1 (Change Detection Network)

Player 1 is then trained for a few epochs with Player 2 frozen. Run:

python train_2players.py --dataset <HRSCD or Valais or LEVIR or WHUCD> \
     --pretrained_P2 <path-to-pretrained-player2> --alpha <alpha> --beta <beta> \
     --freeze True --epochs 5 --ckpt 0 --P1 <SiamUnet or SNUNet or BIT> --lr <lr>

This generates a second checkpoint under:

checkpoints/2player/

4️⃣ Joint Training of the Full 2Player Framework

Finally, train the 2 players jointly using

python train_2players.py --dataset <HRSCD or Valais or LEVIR or WHUCD> \
     --pretrained_P2 <path-to-pretrained-player2> --alpha <alpha> --beta <beta> \
     --freeze False --epochs <epochs> --ckpt <path-to-pretrained-player1> \
     --P1 <SiamUnet or SNUNet or BIT> --lr <lr>

The final model is also saved under: This generates a second checkpoint under:

checkpoints/2player/

🤖 Trained Models

You can download the pretrained model here. Place them under checkpoints/2player and checkpoints/player2.

Qualitative examples using these pretrained models are presented in examples.ipynb.

📚 Citation

If you use 2Player or if you find this work helpful, please cite:

@article{bechaz_2player_2026,
	title = {{2Player}: {A} general framework for self-supervised change detection via cooperative learning},
	volume = {232},
	issn = {0924-2716},
	url = {https://www.sciencedirect.com/science/article/pii/S0924271625004630},
	doi = {https://doi.org/10.1016/j.isprsjprs.2025.11.024},
	journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
	author = {Béchaz, Manon and Dalsasso, Emanuele and Tomoiagă, Ciprian and Detyniecki, Marcin and Tuia, Devis},
	year = {2026},
	keywords = {Change detection, Cooperative learning, Self-supervised learning, Very high-resolution imagery},
	pages = {34--47},
}

@inproceedings{bechaz2025self,
  title={Self-supervised Change Detection via Cooperative Learning: A Two-Player Model},
  author={B{\'e}chaz, Manon and Dalsasso, Emanuele and Tomoiag{\u{a}}, Ciprian and Detyniecki, Marcin and Tuia, Devis},
  booktitle={2025 Joint Urban Remote Sensing Event (JURSE)},
  pages={1--4},
  year={2025},
  organization={IEEE},
  doi={10.1109/JURSE60372.2025.11076078}
}