This is an official PyTorch implementation of Bridging Vision and Language Encoders: Parameter-Efficient Tuning for Referring Image Segmentation.

1. Environment

   • PyTorch (e.g. 1.8.1+cu111)
   • Other dependencies in requirements.txt

     pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
     pip install -r requirements.txt

2. Datasets

   • The detailed instruction is in prepare_datasets.md

3. Pretrained weights

   • Download the pretrained weights of ResNet-50/101 and ViT-B to pretrain

     mkdir pretrain && cd pretrain
     # ResNet-50
     wget https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt
     # ResNet-101
     wget https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt
     # ViT-B
     wget https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt

To do training of ETRIS, modify the script according to your requirement and run:

bash run_scripts/train.sh

To do evaluation of ETRIS, modify the script according to your requirement and run:

bash run_scripts/test.sh

The code is based on CRIS. We thank the authors for their open-sourced code and encourage users to cite their works when applicable.

If ETRIS is useful for your research, please consider citing:

@article{xu2023bridging,
  title={Bridging Vision and Language Encoders: Parameter-Efficient Tuning for Referring Image Segmentation},
  author={Xu, Zunnan and Chen, Zhihong and Zhang, Yong and Song, Yibing and Wan, Xiang and Li, Guanbin},
  journal={arXiv preprint arXiv:2307.11545},
  year={2023}
}