(Oct. 17th, 2022) We release part of our docker file below issue 3 for reimplementation.

Our paper is accepted by ECCV2022. Here is the link of our paper.

This is the PyTorch code of the TS2-NET. The code has been tested on PyTorch 1.7.1.

TS2-Net is a text-video retrieval model based on CLIP. With token shift transformer and token selection transformer, our model achieve SOTA performance on MSR-VTT, VATEX, LSMDC, ActivityNet, and DiDeMo. Our model achieves 54.0 R@1 on MSR-VTT test set.

The token shift operation shift all channels of a token back-and-forth, to preserve the complete token representation and enhance the interaction of adjacent frames.

The token selection module selects tokens contributes most to the local spatial semantics, to better model video representation. (Note: In our released code, we manually regard [CLS] token as the most informative, which we find it can have better performance. It means that we select [CLS] token and other K-1 tokens in token selection stage.)

# From CLIP
conda install --yes -c pytorch pytorch=1.7.1 torchvision cudatoolkit=11.0
pip install ftfy regex tqdm
pip install opencv-python boto3 requests pandas

For TS2-Net (ViT-B/32), and without inverted softmax

For TS2-Net (ViT-B/16) and without inverted softmax

With inverted softmax

Please refer to ArrowLuo/CLIP4Clip to get data annotation. For vatex, please refer to https://eric-xw.github.io/vatex-website/download.html and dataloader_vatex_retrieval.py to prepare VATEX dataset, and get the split file from https://github.com/cshizhe/hgr_v2t.

• For msrvtt,

sh scripts/run_msrvtt.sh

• For vatex,

sh scripts/run_vatex.sh

• Change DATA_PATH to your own data path
• You can change --pretrained_clip_name to ViT-B/16

If you find this code to be useful for your research, please consider citing.

@inproceedings{liu2022ts2net,
      title={TS2-Net: Token Shift and Selection Transformer for Text-Video Retrieval}, 
      author={Yuqi Liu and Pengfei Xiong and Luhui Xu and Shengming Cao and Qin Jin},
      year={2022},
      booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
}

@article{liu2022ts2net,
  title={TS2-Net: Token Shift and Selection Transformer for Text-Video Retrieval},
  author={Yuqi Liu and Pengfei Xiong and Luhui Xu and Shengming Cao and Qin Jin},
  year={2022},
  journal={arXiv preprint arxiv:2207.07852},
}

• Our code is based on ArrowLuo/CLIP4Clip