Enhancing Sample Utilization in Noise-Robust Deep Metric Learning With Subgroup-Based Positive-Pair Selection

---

Official implementation of IEEE Transactions on Image Processing paper Enhancing Sample Utilization in Noise-Robust Deep Metric Learning With Subgroup-Based Positive-Pair Selection.

This is the initial version which contains the core code. Other details are still under development, stay tuned.

⭐️ Key Features

• An asynchronous server that stores all sample features, enabling it to continuously select positive training samples for each training session in real-time.
• A noise-robust deep metric learning framework to enhance information retrieval in computer vision tasks.

Preparing Dataset

Please refer to PRISM to download the CARS_98N dataset. For other datasets, we highly recommend downloading them from OpenDataLab. You can acquire data list from Google Drive.

Training

CUDA_VISIBLE_DEVICES=0 sh run.sh

Acknowledgements

• We would like to thank @Jiaheng Liu for helpful discussion.
• Our PCS module is modified from PRISM
• FeatureServer is inspired from HFSoftmax

Citation

If you find this work helpful, please consider citing our paper:

@article{yu2024enhancing,
  title={Enhancing Sample Utilization in Noise-Robust Deep Metric Learning With Subgroup-Based Positive-Pair Selection},
  author={Yu, Zhipeng and Xu, Qianqian and Jiang, Yangbangyan and Sun, Yingfei and Huang, Qingming},
  journal={IEEE Transactions on Image Processing},
  volume={33},
  pages={6083--6097},
  year={2024},
  publisher={IEEE}
}