Structure-Preserving Motion Estimation for Learned Video Compression

This is the official implementation and appendix of the paper:

Structure-Preserving Motion Estimation for Learned Video Compression. Han Gao, Jinzhong Cui, Mao Ye, Shuai Li, Yu Zhao, Xiatian Zhu. ACM Multimedia 2022. [pdf]

TODO

• Upload appendix.pdf (Done);
• Upload codes (Done);
• Upload pretrained models (Done);
• Update README.md (Continuous maintenance).

Overview

Requirements

• Python==3.8
• Pytorch==1.9

Data Preparation

Testing dataset

• Download HEVC dataset, UVG dataset(1080p/8bit/YUV/RAW) and MCL-JCV dataset, and convert them from YUV format to PNG format.

Test

• Change the configs in class named HEVC_dataset of the file dataset.py to the path of the data to be tested, e.g. :

  root="/xxx/HEVC_dataset/Class_B", filelist="./Tools/filelists/B.txt"
  

• Run test.py for testing, in which the config named --model_path is the pretrained model path, and --lambda_weight is the lambda value of the prerained model, e.g. :

  python -u test.py --model_path="./Checkpoints/2048.pth" --lambda_weight=2048
  

Acknowledgement

During implementation, we drawed on the experience of CompressAI, PyTorchVideoCompression and DCVC. The model weights of intra coding are from CompressAI.

Citation

If you find this paper useful, kindly cite:

@inproceedings{gao2022structure,
  title={Structure-Preserving Motion Estimation for Learned Video Compression},
  author={Gao, Han and Cui, Jinzhong and Ye, Mao and Li, Shuai and Zhao, Yu and Zhu, Xiatian},
  booktitle={Proceedings of the 30th ACM International Conference on Multimedia (MM’22)},
  year={2022}
}

Contact

If any questions, kindly contact with Han Gao via e-mail: [email protected] / [email protected].