TTT-UCDR: Test-time Training for Universal Cross-Domain Retrieval

ArXiv preprint

An extension of Universal Cross-Domain Retrieval: Generalizing across Classes and Domains | ICCV 2021.

Requirements and Setup

Python - 3.7.6, PyTorch - 1.1.0, CUDA - 9.0, cuDNN - 7.5.1, NVIDIA Driver Version >= 384.13

conda create --name torch11 --file requirements.txt
conda activate torch11
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch
conda install -c conda-forge tensorboard
pip install --user future

Download datasets

Check downloads folder for scripts. Change path_dataset in download_sketchy.sh and download_tuberlin.sh.

Pretrained Models

Download from here.

Reproducing our results

Check .sh files in src/algos/SnMpNet for Rotnet, Jigsaw, and Barlow Twins losses.

🎓 Cite

If this code was helpful for your research, consider citing:

@article{paul2022ttt,
  title={TTT-UCDR: Test-time Training for Universal Cross-Domain Retrieval},
  author={Paul, Soumava and Dutta, Titir and Saha, Aheli and Samanta, Abhishek and Biswas, Soma},
  journal={arXiv preprint arXiv:2208.09198},
  year={2022}
}