PatchVQ-docker

Docker inference support for using Patch-VQ (‘Patching Up’ the Video Quality Problem) to compute mean opinion score (MOS) on your own video files. A build of this container has been pushed to Docker Hub already (https://hub.docker.com/r/pixop/patchvq/) for your convenience.

The original project can be found at: https://github.com/baidut/PatchVQ

Examples

Run CUDA accelerated inference on your own video file

Compute MOS on video.mp4 located in the current directory:

docker run --gpus all -it --ipc=host --rm -v $(pwd):/mnt/host pixop/patchvq:latest /mnt/host/video.mp4

Note: Requires Nvidia Container Toolkit to be installed (https://github.com/NVIDIA/nvidia-docker).

Run CPU inference on your own video file

Compute MOS on video.mp4 located in the current directory:

docker run -it --ipc=host --rm -v $(pwd):/mnt/host pixop/patchvq:latest /mnt/host/video.mp4

Warning: This is going to be very slow compared to GPU inference for most people.

Run inference on a 1280x720 crop out of the first 8 seconds only

Compute MOS on video.mp4 located in the current directory while using FFmpeg pre-processing to crop and stop processing after 8 seconds:

docker run -it --ipc=host --rm -v $(pwd):/mnt/host pixop/patchvq:latest /mnt/host/video.mp4 "-vf crop=1280:720 -t 8"

Note: It is possible to supply any number of arguments to FFmpeg this way.

Notes

1. The implementation is not production quality by any means and is merely designed to be a quick, minimal effort way to compute the MOS. No error checking of any kind is performed!
2. The original implementation was tweaked a bit to output PNGs instead of JPEGs to prevent the MOS drop due to lossy encoding. Saving JPEGs at the highest quality still produces noticeable degradation.
3. The container has only been tested on a 64-bit Linux system.
4. Please contact the PatchVQ authors for any questions about the core methodology and the included pretrained models (release v0.1).