ENVC

Offical implementation of Learning Cross-Scale Weighted Prediction for Efficient Neural Video Compression

Installation

We recommend using conda environment for installation.

conda create --name $ENV_NAME
conda activate $ENV_NAME
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip3 install -r requirements.txt
cd third_party/msda
bash make.sh

Evaluation

Currently, we only provide the evaluation code without practical entropy coding, which will be completed in the future.

Preparing test sequences

Test sequences are required to be in sRGB format. The folder structure of test sequences is like this:

|-- testset
	|-- sequence1
		|-- 000001.png
		|-- 000002.png
		|-- 000003.png
		...
	|-- sequence2
	|-- sequence3
	...

Downloading pre-trained models

The pretained model can be downloaded here . We currently only provide the models optimized with MSE metric.

Model	Name	config_file
ENVCwAR	mse512	./cfg/low_rate_model.yaml
ENVCwAR	mse1024	./cfg/low_rate_model.yaml
ENVCwAR	mse2048	./cfg/low_rate_model.yaml
ENVCwAR	mse3072	./cfg/high_rate_model.yaml
ENVCwAR	mse4096	./cfg/high_rate_model.yaml

Evaluating a pre-trained model

Run the evaluation script evaluate.py:

python evaluate.py -h

This will give you a list of options. To evaluate a single sequence in the testset mentioned above, you can run:

python evaluate.py -i /testset/sequence1 -c ./cfg/low_rate_model.yaml --ckpt_path /ckpt/ENVCwAR/mse512.pth

This will give the logout like:

[Frame sequence1 000001] type I bpp 0.175401 psnr 37.8629 msssim 0.980757
[Frame sequence1 000002] type P bpp 0.042709 psnr 36.9020 msssim 0.978645
[Frame sequence1 000003] type P bpp 0.052453 psnr 36.8629 msssim 0.978108
...
[Sequence '/testset/sequence1'] bpp 0.054796 psnr 36.3074 msssim 0.973306
[Dataset '/testset'] bpp 0.054796 psnr 36.3074 msssim 0.973306

You can also evaluate multiple test sequences at once using the path in glob pattern:

python evaluate.py -i "/testset/sequence*" -c ./cfg/low_rate_model.yaml --ckpt_path $CKPT_PATH

Citation

If you use this library for research purposes, please cite:

@article{guo2023learning,
  title={Learning Cross-Scale Weighted Prediction for Efficient Neural Video Compression},
  author={Guo, Zongyu and Feng, Runsen and Zhang, Zhizheng and Jin, Xin and Chen, Zhibo},
  journal={IEEE Transactions on Image Processing},
  year={2023},
  publisher={IEEE}
}