This project presents the Neural architecture to compress videos (sequence of image frames) along with the pre-trained models. Our work is inspired by DVC and we use tensorflow-compression for bitrate estimation and entropy compression. Compression is realized in terms of actual file size.

If you find our paper useful, please cite:

@inproceedings{dhungel2020efficient,
  title={An Efficient Video Compression Network},
  author={Dhungel, Prasanga and Tandan, Prashant and Bhusal, Sandesh and Neupane, Sobit and Shakya, Subama},
  booktitle={2020 2nd International Conference on Advances in Computing, Communication Control and Networking (ICACCCN)},
  pages={1028--1034},
  year={2020},
  organization={IEEE}

For installation, simply run the following command:

pip install -r requirements.txt

for GPU support, replace the tensorflow==1.15.0 line in requirements.txt with tensorflow-gpu==1.15.0 .

Note: Precompiled packages for tensorflow-compression are currently only provided for Linux (Python 2.7, 3.3-3.6) and Darwin/Mac OS (Python 2.7, 3.7). For windows please refer this.

Pre-trained models are available at checkpoints. The models suffixed with "msssim" are the ones that are optimized with MS-SSIM while the rest are optimized with PSNR. The Integer in the filename denotes the lambda (weight assigned to distortion compared to the bitrate). Higher the value of lambda lower will be the distortion and higher will be the bitrate.

Run the following command and follow the instructions:

python compress.py -h 

For example,

python compress.py -i demo/input/ -o demo/compressed/ -m checkpoints/videocompressor1024.pkl -f 101

The execution compresses the frames in demo/input/ to compressed files in demo/compressed/.

Note: Right now, our work is only applicable to the RGB frames of height and width, that are multiple of 16. Needless to say, higher resolution images require more time to train, compress and decompress

Run the following command and follow the instructions:

python decompress.py -h 

For example,

python decompress.py -i demo/compressed/ -o demo/reconstructed -m checkpoints/videocompressor1024.pkl -f 101

The execution will reconstruct the original frames in demo/reconstructed/ with some compression artifacts.

We trained the network with vimeo-septuplet dataset.To download the dataset, run the script download_dataset.sh as:

sh download_dataset.sh

Here, we provide the small portion of the large dataset, to present the dataset outlook. You can train your own model by simply executing the following command and following the instructions:

python train.py -h 

For training the dataset structure should be same as vimeo-septuplet structure, otherwise you should write your own data-parser to the network.

We perform compression and reconstruction in a single file test.py for evaluation. To evaluate the compression and distortion, execute:

python test.py -h

and follow the instructions. For example,

python test.py -m checkpoints/videocompressor256-msssim.pkl

Experimental results are available at the evaluation.

Note: The compression and reconstruction without GPU will be slower than the above demonstration.

The images are first frame, second frame, optical flow, reconstructed optical flow, motion compensated frame, residue, reconstructed residue and reconstructed frame respectively.

Prasanga Dhungel
Prashant Tandan
Sandesh Bhusal
Sobit Neupane