CBD-pytorch

Official PyTorch implementation of the paper "Counting Beyond Domains: Towards Alignment in Unsupervised Domain Adaptation in Remote Sensing Object Counting".

Pre-trained Models

We provide the checkpoints of models for download via Google Drive:

• Pre-trained Models on Google Drive

Environment Setup

The following environment setup was used to ensure reproducibility:

• Python 3.12
• CUDA Toolkit 12.4
• PyTorch 2.5.1
• NumPy 2.1.3
• Matplotlib 3.9.2
• Pillow 11.0.0

Ensure these dependencies are installed using the following command:

pip install -r requirements.txt

Usage

Testing Pre-trained Models

We provide code to test our models. The provided models PU2CA_parameter.pth and CA2PU_parameter.pth were adapted from PUCPR+ to CARPK and from CARPK to PUCPR+, respectively.

To visualize model performance on sample images, run the following command:

python test_vis_single.py

Visualizations for selected results are included in the images folder. You can modify the script to test on other images.

Testing on Public/Custom Dataset

We also provide a method to test on public/custom datasets. The dataset should be organized in the following structure:

└── datasets
    └── dataset_name
        └── test
            ├── den
            │   ├── 1.npy
            │   ├── 2.npy
            │   └── ...
            └── img
                ├── 1.png
                ├── 2.png
                └── ...

Once your dataset is properly structured, you can run the following command to test:

python test.py

For detailed explanations of the network, please refer to our paper.

Acknowledgements

This codebase builds on and acknowledges the following repositories:

• C-3 Framework
• FIDTM

We thank the authors of these repositories for their contributions to the community.

Contact

If you have any questions or issues, please feel free to reach out to us at: jiayi.wang@whu.edu.cn and guanchen.ding@connect.polyu.hk