sohex / easyvmaf

Python script to easily compute VMAF using FFmpeg. It allows to deinterlace, scale and sync Ref and Distorted video automatically

easyVmaf

Python tool based on ffmpeg and ffprobe to deal with the video preprocesing required for VMAF inputs:

• Deinterlacing
• Upscaling/downscaling
• Frame-to-Frame Syncing
• Frame rate adaptation

Details about How it Works can be found here.

Updates

• With libvmaf v2.0.0, the vmaf models files were refactored. Basically, the models use now fixed-point data types instead of float-point data types. So depending on the libvmaf version you are using, you could need to modify the easyVmaf/config.py file:

  • libvmaf < v2.0.0 : Use the old vmaf_*.pkl model files with float-point data types: vmaf_4k_v0.6.1.pkl, vmaf_v0.6.1.pkl, etc.

  • libvmaf >= v2.0.0 : Use the new vmaf_*.json model files with fixed-point data types: vmaf_4k_v0.6.1.json, vmaf_v0.6.1.json, vmaf_v0.6.1neg.json

• New neg model is supported since libvmaf v2.0.0 only. So if you want to use it, be sure to update libvmaf first.

Requirements

• Linux/OSX

• Python >= v3.0

• FFmpeg build with libvmaf. More details here

Installation

• Just clone the repo and run it from the source folder.

$ git clone https://github.com/gdavila/easyVmaf.git
$ cd easyVmaf

• Run from docker image. More info at the end of this document.

Usage

$ python3 easyVmaf.py -h
usage: eVmaf [-h] -d D -r R [-sw SW] [-ss SS] [-subsample N] [-reverse]
             [-model MODEL] [-phone] [-verbose] [-output_fmt TYPE]

Script to easy compute VMAF using FFmpeg. It allows to deinterlace, scale and sync Ref and Distorted video samples automatically:             

 	 Autodeinterlace: If the Reference or Distorted samples are interlaced, deinterlacing is applied            

 	 Autoscale: Reference and Distorted samples are scaled automatically to 1920x1080 or 3840x2160 depending on the VMAF model to use            

 	 Autosync: The first frames of the distorted video are used as reference to a sync look up with the Reference video.             
 	 	 The sync is doing by a frame-by-frame look up of the best PSNR            
 	 	 See [-reverse] for more options of syncing            

 As output, a json or XML file with VMAF score is created

optional arguments:
  -h, --help    show this help message and exit
  -sw SW        Sync Window: window size in seconds to get a subsample of the Reference video. The sync look up will be done between the first frames of the Distorted input and this Subsample. (default=0. No sync).
  -ss SS        Sync Start Time. Time in seconds from the beginning of the Reference video from which the Sync Window will be applied. (default=0).
  -subsample N  Specifies the subsampling of frames to speed up calculation. (default=1, None).
  -reverse      If enable, it Changes the default Autosync behaviour: The first frames of the Reference video are used as reference to sync with the Distorted one. (Default = Disable).
  -model MODEL  Vmaf Model. Options: HD, HDneg, 4K. (Default: HD).
  -phone        It enables vmaf phone model (HD only). (Default=disable).
  -verbose      Activate verbose loglevel. (Default: info).
  -output_fmt TYPE   Output vmaf file format. Options: json or xml (Default: json)

required arguments:
  -d D          Distorted video
  -r R          Reference video 

NOTE: HDneg is not supported yet.

Examples

Syncing: Reference Video delayed in regard with the first frame of Distorted one.

VMAF computation for two video samples, reference.ts and distorted-A.ts. Both videos are not synced: reference.ts is delayed in comparition with distorted-A.ts, i.e., the first frame of distorted-A.ts matchs with the frame located at 0.7007 seconds since the begining of reference.ts (blue arrow on the figure). To sync both videos automatically using easyVmaf, the next command line is used:

```bash
$ python3 easyVmaf.py -d distorted-A.ts -r reference.ts -sw 2


...
...
[Ignored outputs]
...
...

Sync Info:
offset:  0.7007000000000001 psnr:  48.863779
VMAF score:  89.37913542219542
VMAF json File Path:  distorted-A_vmaf.json
```

The previus command line takes a synchronisation window sw of 2 seconds , this means that the sync lookup will be done between the first frame of distorted-A.ts (actually, in practise it takes into account several frames) and a subsample of reference.ts of 2 seconds lenght since its begin.

```bash
$ python3 easyVmaf.py -d distorted.ts -r reference.ts -sw 2
...
...
[Ignored FFmpeg outputs]
...
...
Sync Info:
offset:  0.7007000000000001 psnr:  48.863779
VMAF score:  89.37913542219542
VMAF json File Path:  distorted.json
```

Syncing: Distorted Video delayed in regard with the first frame of Reference one.

This time, distorted-B.ts is delayed in comparition with reference.ts, i.e., The first frame of reference.ts matchs with the frame located at 8.3003 seconds since the begining of distorted-B.ts. To sync the videos automatically, the next command line is used:

```bash
$ python3 easyVmaf.py -d distorted-B.ts -r reference.ts -sw 3 -ss 6 -reverse


...
...
[Ignored FFmpeg outputs]
...
...

Sync Info:
offset:  8.300300000000000 psnr:  34.897866
VMAF score:  92.34452778643345
VMAF json File Path:  distorted-B_vmaf.json
```

The previous command line applies a syncronization window sw of 3 seconds, a sync start time ss of 6 seconds and the reverse flag.

Note the use of the reverse flag (that was not used on the first example). This flag allows to interchange to which video the syncWindow will be applied (reference or distorted).

Docker Image usage

A docker image is available on docker hub to run easyVmaf in a straightforward way.

The Docker Image is basically an ubuntu image with ffmpeg and libvmaf already installed. You can check the Dockerfile for more details.

The easiest way to run easyVmaf through Docker is mounting a shared volume between your host machine and the container. This volume should have inside it all the video files you want to analyze. The outputs (vmaf information files) will be putting in this shared folder also.

Example

docker run --rm -v <local-path-to-your-video-files>:/<custom-name-folder> gfdavila/easyvmaf -r /<custom-name-folder>/video-1.mp4 -d /<custom-name-folder>/video-2.mp4

Some video samples located on the docker image:

NAME                        TIME

                           t=0
                            |
BBB_reference_10s.mp4       */-----------------------------*/
BBB_sampleA_distorted.mp4           */---------------------*/
BBB_sampleB_distorted.mp4       */-------------------------*/

Run docker container to get VMAF between BBB_reference_10s.mp4 and BBB_sampleA_distorted.mp4:

:~$ docker run --rm  gfdavila/easyvmaf -r video_samples/BBB_reference_10s.mp4 -d video_samples/BBB_sampleA_distorted.mp4 -sw 1 -ss 1

Run docker container to get VMAF between BBB_sampleA_distorted.mp4 and BBB_sampleB_distorted.mp4:

:~$ docker run --rm  gfdavila/easyvmaf -r video_samples/BBB_sampleA_distorted.mp4 -d video_samples/BBB_sampleB_distorted.mp4 -sw 2 -ss 0 -reverse