Official implementation of the paper "I2VC: A Unified Framework for Intra & Inter-frame Video Compression".

:rocket: News

• (May 22, 2024)
  • Code for our implementation is now available.

---

Main Contributions

1. Unified framework for Intra- and Inter-frame video compression: The three types of frames (I-frame, P-frame and B-frame) across different video compression configurations (AI, LD and RA) within a GoP are uniformly solved by one framework.
2. Implicit inter-frame feature alignment: We leverage DDIM inversion to selective denoise motion-rich areas based on decoded features, achieving implicit inter-frame feature alignment without MEMC.
3. Spatio-temporal variable-rate codec: We design a spatio-temporal variable-rate codec to unify intra- and inter-frame correlations into a conditional coding scheme with variable-rate allocation.

Requirements

To install requirements:

pip install -r requirements.txt

Data preparation

Please follow the instructions at DATASETS.md to prepare all datasets.

Training

To train the model in the paper, run this command:

bash scripts/run.sh

Evaluation

To evaluate trained model on test data, run:

bash scripts/test.sh

Model Zoo

Coming soon.

---

Citation

If you use our work, please consider citing:

@inproceedings{2024i2vc,
    title={I2VC: A Unified Framework for Intra & Inter-frame Video Compression},
    author={Meiqin Liq, Chenming Xu, Yukai Gu, Chao Yao, Yao Zhao},
    publisher = {arXiv},
    year={2024}
}

Contact

If you have any questions, please create an issue on this repository or contact at mqliu@bjtu.edu.cn, chenming_xu@bjtu.edu.cn or yukai.gu@bjtu.edu.cn.

Acknowledgements

Our code is based on DCVC repository. We thank the authors for releasing their code. If you use our model and code, please consider citing these works as well.