Target-Guided Composed Image Retrieval [ACM MM 2023]

Authors

Haokun Wen¹, Xian Zhang¹, Xuemeng Song²*, Yinwei Wei³, Liqiang Nie¹*

¹ Harbin Institute of Technology (Shenzhen), Shenzhen, China
² Shandong University, Qingdao, China
³ Monash University, Melbourne, Australia
* Corresponding authors

Links

• Paper: ACM DL
• Checkpoints: Download Link

Key dependencies

clip==0.2.0
matplotlib==3.5.1
numpy==1.22.3
pandas==1.4.2
Pillow==9.0.1 / 10.0.0
seaborn==0.12.2
torch==1.7.0
torchvision==0.8.0
tqdm==4.65.0

Dataset Preparation

FashionIQ & Shoes

Spell-corrected data files for FashionIQ and Shoes are provided in correction_dict folder.

CIRR

For CIRR, test results must be submitted through the official CIRR evaluation website. Submission files can be generated using:

python cirr_test_submission.py

---

Usage

Training

FashionIQ

python train.py --dataset 'fashioniq' --model_dir <output_dir> \
  --mu_ 0.1 --nu_ 10 --kappa_ 0.5 --tau_ 0.1 --P 4 --Q 8

Shoes

python train.py --dataset 'shoes' --model_dir <output_dir> \
  --mu_ 0.05 --nu_ 5 --kappa_ 0.5 --tau_ 0.1 --P 3 --Q 6

CIRR

python train.py --dataset 'cirr' --model_dir <output_dir> \
  --mu_ 0.1 --nu_ 1 --kappa_ 0.1 --tau_ 0.05 --P 4 --Q 8

---

Citation

If you find this work useful, please cite:

@inproceedings{wen2023target,
  title={Target-Guided Composed Image Retrieval},
  author={Wen, Haokun and Zhang, Xian and Song, Xuemeng and Wei, Yinwei and Nie, Liqiang},
  booktitle={Proceedings of the ACM International Conference on Multimedia},
  pages={915--923},
  year={2023}
}

---

License

This program is licensed under the GNU General Public License 3.0.
See https://www.gnu.org/licenses/gpl-3.0.html for details.
Any derivative work must also be licensed under the GNU GPL as published by the Free Software Foundation, either Version 3 or (at your option) any later version.