This repo holds the code for the paper:

DVC: An End-to-end Deep Video Compression Framework, Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, Zhiyong Gao, CVPR 2019 (Oral). [arXiv]

Please check the PyTorch reimplmentation from my colleague.

2020-04-07: I upload the 4 pre-train models with different lambda values and you can use different lambda for rate control.

We provide the test code of our DVC framework. In our implementation, we use the learning based image compression algorithm (Ballé et al) as the intra compression. Specifically, for the video codec model with lambda=k, the image codec model with lambda=4k is used as the intra frames. For the image codec, please refer to GoogleCompression. And the pre-train models of our DVC are located in TestDemo/VideoCodec/model.

Currently, we do not provide the entropy coding module. The generated features from image/video codecs are saved to .pkl files. We give the estimated Bpp for these features. It is straightforward to compress these features by using traditional entropy coding tools, such as CABAC.

cd ./TestDemo/VideoCodec

Video Encoding,

python  Encoder.py --EncoderModel /path/to/encoder/model/frozen_model_E.pb  --input_frame /path/to/currentframe/im002.png --refer_frame  /path/to/previousframe/im001.png  	--output /output/feature/folder/

Video Decoding,

python Decoder.py --DecoderModel /path/to/decoder/model/frozen_model_D.pb  --refer_frame /path/to/previous/im001.png --loadpath  /path/to/feature/folder/

Evaluation results on the UVG dataset and HEVC Class B (1080p) and Class E (720p). Please refer to our paper for more experimental results.

We also provide the scrips for generating all the RD curves of our paper in folder RDCurve.

If you find our paper useful, please cite:

@article{lu2018dvc,
  title={DVC: An End-to-end Deep Video Compression Framework},
  author={Lu, Guo and Ouyang, Wanli and Xu, Dong and Zhang, Xiaoyun and Cai, Chunlei and Gao, Zhiyong},
  journal={arXiv preprint arXiv:1812.00101},
  year={2018}
}

You can contact Guo Lu by sending mail to [email protected]