This repository contains our series of works on Deep Hierarchical Video Compression.
- DHVC 1.0: The first hierarchical predictive coding method moves away from the hybrid coding framework, achieving best-in-class performance. Paper is available at Deep Hierarchical Video Compression (AAAI 2024).
- DHVC 2.0: The enhanced hierarchical predictive coding method, which integrates variable-rate intra- and inter-coding into a single model, delivering not only superior compression performance to representative methods but also real-time processing with a significantly smaller memory footprint on standard GPUs. Paper is available at High-Efficiency Neural Video Compression via Hierarchical Predictive Learning (arxiv 2024).
[2025.2.12] We have reconstructed the code and uploaded the pretrained models of DHVC 1.0.
- Python 3.8+
- CUDA 11.0
- pytorch 1.11.0
- For others, please refer to requirements.txt
The pretrained models of DHVC 1.0 can be downloaded from NJU Box.
- Train dataset: Vimeo90k
- Test dataset: UVG、MCL-JCV、HEVC Class B
Please download the pretrained models and configure the environment properly first.
Follow the command below to run testing in the dhvc-1.0 folder:
python test.py -d test_dataset_name -c checkpoint_path -p test_dataset_path -g 32 -f 96 -d represents the name of the test dataset used in log file. -c, -p represent the path of the pretrained models and test dataset. -g, -f represent the GOP size and total frame numbers for evaluation. By default, the pretrained models will be placed in ./pretrained, the test dataset will be placed in ./dataset. The test results can be found in ./runs.
If you find this work helpful to your research, please cite:
@inproceedings{lu2024deep,
title={Deep Hierarchical Video Compression},
author={Lu, Ming and Duan, Zhihao and Zhu, Fengqing and Ma, Zhan},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={8},
pages={8859--8867},
year={2024}
}
@article{lu2024high,
title={High-Efficiency Neural Video Compression via Hierarchical Predictive Learning},
author={Lu, Ming and Duan, Zhihao and Cong, Wuyang and Ding, Dandan and Zhu, Fengqing and Ma, Zhan},
journal={arXiv preprint arXiv:2410.02598},
year={2024}
}
If you have any question, feel free to contact us via [email protected] or [email protected].