TALENT: Target-aware Efficient Tuning for Referring Image Segmentation

TALENT is a framework for Referring Image Segmentation (RIS) designed to address the "non-target activation" (NTA) issue in parameter-efficient tuning. It introduces a Rectified Cost Aggregator (RCA) to aggregate text-referred features and a Target-aware Learning Mechanism (TLM) to calibrate activation into accurate target localization.

Resources

• Paper: TALENT: Target-aware Efficient Tuning for Referring Image Segmentation
• Repository: GitHub - Kimsure/TALENT

Usage

To evaluate the model, follow the installation instructions in the GitHub repository and run the following script:

bash run_scripts/test.sh

To visualize the results, you can set the visualize flag to True in the configuration file.

Acknowledgements

The code for TALENT is based on CRIS, ETRIS, and previous TALENT implementations. We thank the authors for their open-sourced code.

Citation

If you find this work useful, please cite:

@article{talent2026,
  title={TALENT: Target-aware Efficient Tuning for Referring Image Segmentation},
  author={Shuo Jin, Siyue Yu, Bingfeng Zhang, Chao Yao, Meiqin Liu, Jimin Xiao},
  journal={arXiv preprint arXiv:2604.00609},
  year={2026}
}