UGSDA-PyTorch

Official PyTorch implementation of the paper "Domain-Agnostic Crowd Counting via Uncertainty-Guided Style Diversity Augmentation" accepted at ACM Multimedia 2024.

Pre-trained Models

We provide the pre-trained models for download via Google Drive:

• Pre-trained Models on Google Drive

Environment Setup

The following environment setup was used to ensure reproducibility:

• Python 3.8
• CUDA Toolkit 11.3.1
• PyTorch 1.11.0
• NumPy 1.23.0
• Matplotlib 3.6.2
• Pandas 2.0.3
• Pillow 9.4.0

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 SHHA_parameter.pth and SHHB_parameter.pth were trained on the ShanghaiTech Part A (SHHA) and Part B (SHHB) datasets, 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.csv
            │   ├── 2.csv
            │   └── ...
            └── img
                ├── 1.jpg
                ├── 2.jpg
                └── ...

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.

Citation

If you use this code in your research, please cite our paper:

@inproceedings{ding2024domain,
  title={Domain-Agnostic Crowd Counting via Uncertainty-Guided Style Diversity Augmentation},
  author={Ding, Guanchen and Liu, Lingbo and Chen, Zhenzhong and Chen, Chang Wen},
  booktitle={ACM Multimedia 2024}
}

Acknowledgements

This codebase builds on and acknowledges the following repositories:

• C-3 Framework

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 me at: [email protected]