Seeking Proxy Point via Stable Feature Space for Noisy Correspondence Learning (IJCAI-2025)

This repository contains the official implementation of the IJCAI-2025 paper: "Seeking Proxy Point via Stable Feature Space for Noisy Correspondence Learning".

Authors: Yucheng Xie, Songyue Cai, Tao Tong, Ping Hu, Xiaofeng Zhu.

🛠 Requirements

We recommend using Anaconda to manage the environment.

• Python 3.7
• PyTorch ~1.7.1
• numpy
• scikit-learn
• Punkt Sentence Tokenizer (nltk)

# Example installation commands
conda create -n proxy python=3.7
conda activate proxy
pip install torch==1.7.1 numpy scikit-learn
python -c "import nltk; nltk.download('punkt')"

📂 Datasets

We follow NCR (NeurIPS 2021) to obtain image features and vocabularies.

After downloading the data, please organize the folders as follows:

|-- data
    |-- data
    |   |-- cc152k_precomp
    |   |-- coco_precomp
    |   |-- f30k_precomp
    |-- vocab
        |-- cc152k_precomp_vocab.json
        |-- coco_precomp_vocab.json
        |-- f30k_precomp_vocab.json

🚀 Training

1. Flickr30K (Synthetic Noise)

Run the following command to train on Flickr30K. You can modify --noise_ratio to 0.2, 0.4, 0.6, or 0.8 to conduct experiments with different noise levels.

python run.py --data_name=f30k_precomp --noise_ratio=0.2 --num_epochs=40

2. MS-COCO (Synthetic Noise)

Similar to Flickr30K, you can adjust the --noise_ratio (0.2 | 0.4 | 0.6 | 0.8).

python run.py --data_name=coco_precomp --noise_ratio=0.2 --num_epochs=20

3. CC152K (Real-world Noise)

Since CC152K is a real-world noisy dataset, no noise_ratio argument is needed.

python run.py --data_name=cc152k_precomp --num_epochs=40

📊 Evaluating

To evaluate the models, run:

python evaluation.py

Note:

• By default, this script evaluates all models located in ./model_ckpt/cream_models/.
• To evaluate a specific model, please modify the model_path variable in evaluation.py.

Pre-trained Models

We provide the pre-trained models used for the paper experiments. You can download them from the following link:

• Google Drive: Download Link

📝 Citation

If you find this work useful or interesting for your research, please consider citing:

@inproceedings{xie2025seeking,
  title={Seeking proxy point via stable feature space for noisy correspondence learning},
  author={Xie, Yucheng and Cai, Songyue and Tong, Tao and Hu, Ping and Zhu, Xiaofeng},
  booktitle={Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence},
  pages={2072--2080},
  year={2025}
}

📧 Contact

If you have any questions, please feel free to create an issue on this repository or contact us at xyemrsnon@gmail.com.