VMAF is a perceptual video quality assessment algorithm developed by Netflix. VMAF Development Kit (VDK) is a software package that contains the VMAF algorithm implementation, as well as a set of tools that allows a user to train and test a custom VMAF model. For an overview, read this tech blog post, or this slide deck.
- (1/20/18) Moved custom subjective models into a submodule named sureal. If you pull the latest changes, you will have to pull the submoddule by
git submodule update --init --recursiveand addsureal/python/srctoPYTHONPATH. - (8/12/17) VMAF is now included as a filter in FFmpeg main branch, and can be configured using:
./configure --enable-libvmaf. - (7/16/17) VMAF is now packaged into a library call
libvmafand can be called from a C/C++ program directly. See this section for details.
Refer to the FAQ page.
The VDK package has its core feature extraction library written in C, and the rest scripting code written in Python. It also has a stand-alone C++ implementation that is Python-independent. To build the C/C++ code, it requires gcc and g++ (>=4.8). To run scripts and tests, it requires Python2 (>= 2.7) installed.
It also requires a number of Python packages:
numpy(>=1.12.0)scipy(>=0.17.1)matplotlib(>=2.0.0)pandas(>=0.19.2)scikit-learn(>=0.18.1)h5py(>=2.6.0)
You will need to install gfortran for compiling scipy, freetype and pkg-config required by matplotlib, and hdf5 required by h5py (C header files needed). These can't be compiled from source here.
Install the dependencies:
sudo apt-get update -qq && \
sudo apt-get install -y \
pkg-config gfortran libhdf5-dev libfreetype6-dev liblapack-dev \
python python-setuptools python-dev python-pip python-tk
Upgrade pip to the newest version:
sudo -H pip install --upgrade pip
Then install the required Python packages:
pip install --user numpy scipy matplotlib notebook pandas sympy nose scikit-learn h5py
Make sure your user install executable directory is on your PATH. Add this to the end of ~/.bashrc and restart your shell:
export PATH="$PATH:$HOME/.local/bin"
First, install Homebrew, then install the dependencies:
brew install gcc freetype pkg-config homebrew/science/hdf5 python
This will install an up-to-date version of Python and pip, but you must add the following to the end of ~/.bash_profile and restart your shell:
export PATH="/usr/local/opt/python/libexec/bin:$PATH"
Now install the required Python packages:
brew install numpy scipy
pip install matplotlib notebook pandas sympy nose scikit-learn h5py
You can verify if these packages are properly installed and its version/location by:
python -c 'import numpy as pkg; print(pkg.__version__); print(pkg.__file__)'
python -c 'import scipy as pkg; print(pkg.__version__); print(pkg.__file__)'
python -c 'import matplotlib as pkg; print(pkg.__version__); print(pkg.__file__)'
python -c 'import pandas as pkg; print(pkg.__version__); print(pkg.__file__)'
python -c 'import sklearn as pkg; print(pkg.__version__); print(pkg.__file__)'
python -c 'import h5py as pkg; print(pkg.__version__); print(pkg.__file__)'
If you see that the printed version number is older than the ones aforementioned, it could suggest that a previously installed package with the same name but older version at a different location may have overshadowed the new one. Make sure that the new one's path appears early in the path list, which can be printed by:
python -c 'import sys; print(sys.path)'
(Or simply delete the older one).
First, pull submodule sureal by running:
git submodule update --init --recursive
After cloning VMAF repository, cd to the repo directory and run:
make
to build the binaries.
Add the python/src and sureal/python/src subdirectories to the environment variable PYTHONPATH:
export PYTHONPATH="$(pwd)/python/src:$(pwd)/sureal/python/src:$PYTHONPATH"
You can also add it to the environment permanently, by appending to ~/.bashrc:
echo export PYTHONPATH="$(pwd)/python/src:$(pwd)/sureal/python/src:$PYTHONPATH" >> ~/.bashrc
source ~/.bashrc
Under macOS, use ~/.bash_profile instead.
The package has thus far been tested on Ubuntu 16.04 LTS and macOS 10.13.
After installation, run:
./unittest
There are two basic execution modes to run VMAF – a single mode and a batch mode.
To run VMAF on a single reference/distorted video pair, run:
./run_vmaf format width height reference_path distorted_path [--out-fmt output_format]
The arguments are the following:
formatcan be one of:yuv420p,yuv422p,yuv444p(8-Bit YUV)yuv420p10le,yuv422p10le,yuv444p10le(10-Bit little-endian YUV)
widthandheightare the width and height of the videos, in pixelsreference_pathanddistorted_pathare the paths to the reference and distorted video filesoutput_formatcan be one of:textxmljson
For example:
./run_vmaf yuv420p 576 324 \
python/test/resource/yuv/src01_hrc00_576x324.yuv \
python/test/resource/yuv/src01_hrc01_576x324.yuv \
--out-fmt json
This will generate JSON output like:
"aggregate": {
"VMAF_feature_adm2_score": 0.92542107502749982,
"VMAF_feature_motion2_score": 4.0498253541666669,
"VMAF_feature_vif_scale0_score": 0.36342048943884936,
"VMAF_feature_vif_scale1_score": 0.76664754213485187,
"VMAF_feature_vif_scale2_score": 0.86285466690193247,
"VMAF_feature_vif_scale3_score": 0.91597177803640772,
"VMAF_score": 65.44885887590759,
"method": "mean"
}
where VMAF_score is the final score and the others are the scores for VMAF's elementary metrics.
adm2,vif_scalexscores range from 0 (worst) to 1 (best)motion2score typically ranges from 0 (static) to 20 (high-motion)
There is also an ffmpeg2vmaf script which can compare any file format supported by ffmpeg (note that you need a recent version of ffmpeg installed):
./ffmpeg2vmaf width height reference_path distorted_path [--out-fmt output_format]
To run VMAF in batch mode, create an input text file, where each corresponds to the following format (check examples in example_batch_input):
format width height reference_path distorted_path
For example:
yuv420p 576 324 python/test/resource/yuv/src01_hrc00_576x324.yuv python/test/resource/yuv/src01_hrc01_576x324.yuv
yuv420p 576 324 python/test/resource/yuv/src01_hrc00_576x324.yuv python/test/resource/yuv/src01_hrc00_576x324.yuv
After that, run:
./run_vmaf_in_batch input_file [--out-fmt out_fmt] [--parallelize]
where enabling --parallelize allows execution on multiple reference-distorted video pairs in parallel.
For example:
./run_vmaf_in_batch example_batch_input --parallelize
VMAF v0.6.1 and later support a custom quality model for cellular phone screen viewing. This model can be invoked by adding --phone-model option in the commands run_vmaf, run_vmaf_in_batch (but also in run_testing and vmafossexec which are introduced the following sections):
./run_vmaf yuv420p 576 324 \
python/test/resource/yuv/src01_hrc00_576x324.yuv \
python/test/resource/yuv/src01_hrc01_576x324.yuv \
--phone-model
./run_vmaf_in_batch example_batch_input --parallelize --phone-model
Invoking the phone model will generate VMAF scores higher than in the regular model, which is more suitable for laptop, TV, etc. viewing conditions. An example VMAF–bitrate relationship for the two models is shown below:
From the figure it can be interpreted that due to the factors of screen size and viewing distance, the same distorted video would be perceived as having a higher quality when viewed on a phone screen than on a laptop/TV screen, and when the quality score reaches its maximum (100), further increasing the encoding bitrate would not result in any perceptual improvement in quality.
VMAF follows a machine-learning based approach to first extract a number of quality-relevant features (or elementary metrics) from a distorted video and its reference full-quality video, followed by fusing them into a final quality score using a non-linear regressor (e.g. an SVM regressor), hence the name “Video Multi-method Assessment Fusion”.
In addition to the basic commands, the VMAF package also provides a framework to allow any user to train his/her own perceptual quality assessment model. For example, directory resource/model contains a number of pre-trained models, which can be loaded by the aforementioned commands:
./run_vmaf format width height reference_path distorted_path [--model model_path]
./run_vmaf_in_batch input_file [--model model_path] --parallelize
For example:
./run_vmaf yuv420p 576 324 \
python/test/resource/yuv/src01_hrc00_576x324.yuv \
python/test/resource/yuv/src01_hrc01_576x324.yuv \
--model model/nflxtrain_vmafv3.pkl
./run_vmaf_in_batch example_batch_input --model model/nflxtrain_vmafv3.pkl --parallelize
A user can customize the model based on:
- The video dataset it is trained on
- The list of features used
- The regressor used (and its hyper-parameters)
Once a model is trained, the VMAF package also provides tools to cross validate it on a different dataset and visualization.
To begin with, create a dataset file following the format in example_dataset.py. A dataset is a collection of distorted videos. Each has a unique asset ID and a corresponding reference video, identified by a unique content ID. Each distorted video is also associated with subjective quality score, typically a MOS (mean opinion score), obtained through subjective study. An example code snippet that defines a dataset is as follows:
dataset_name = 'example'
yuv_fmt = 'yuv420p'
width = 1920
height = 1080
ref_videos = [
{'content_id':0, 'path':'checkerboard.yuv'},
{'content_id':1, 'path':'flat.yuv'},
]
dis_videos = [
{'content_id':0, 'asset_id': 0, 'dmos':100, 'path':'checkerboard.yuv'}, # ref
{'content_id':0, 'asset_id': 1, 'dmos':50, 'path':'checkerboard_dis.yuv'},
{'content_id':1, 'asset_id': 2, 'dmos':100, 'path':'flat.yuv'}, # ref
{'content_id':1, 'asset_id': 3, 'dmos':80, 'path':'flat_dis.yuv'},
]
See the directory resource/dataset for more examples. Also refer to the Datasets section regarding publicly available datasets.
Once a dataset is created, first validate the dataset using existing VMAF or other (PSNR, SSIM or MS-SSIM) metrics. Run:
./run_testing quality_type test_dataset_file [--vmaf-model optional_VMAF_model_path] [--cache-result] [--parallelize]
where quality_type can be VMAF, PSNR, SSIM or MS_SSIM.
Enabling --cache-result allows storing/retrieving extracted features (or elementary quality metrics) in a data store (since feature extraction is the most expensive operations here).
Enabling --parallelize allows execution on multiple reference-distorted video pairs in parallel. Sometimes it is desirable to disable parallelization for debugging purpose (e.g. some error messages can only be displayed when parallel execution is disabled).
For example:
./run_testing VMAF example_dataset.py --cache-result --parallelize
Make sure matplotlib is installed to visualize the MOS-prediction scatter plot and inspect the statistics:
- PCC – Pearson correlation coefficient
- SRCC – Spearman rank order correlation coefficient
- RMSE – root mean squared error
When creating a dataset file, one may make errors (for example, having a typo in a file path) that could go unnoticed but make the execution of run_testing fail. For debugging purposes, it is recommended to disable --parallelize.
If the problem persists, one may need to run the script:
python python/script/run_cleaning_cache.py quality_type test_dataset_file
to clean up corrupted results in the store before retrying. For example:
python python/script/run_cleaning_cache.py VMAF example_dataset.py
Now that we are confident that the dataset is created correctly and we have some benchmark result on existing metrics, we proceed to train a new quality assessment model. Run:
./run_vmaf_training train_dataset_filepath feature_param_file model_param_file output_model_file [--cache-result] [--parallelize]
For example:
./run_vmaf_training example_dataset.py \
resource/feature_param/vmaf_feature_v2.py \
resource/model_param/libsvmnusvr_v2.py \
workspace/model/test_model.pkl \
--cache-result --parallelize
feature_param_file defines the set of features used. For example, both dictionaries below:
feature_dict = {'VMAF_feature':'all', }
and
feature_dict = {'VMAF_feature':['vif', 'adm'], }
are valid specifications of selected features. Here VMAF_feature is an 'aggregate' feature type, and vif, adm are the 'atomic' feature types within the aggregate type. In the first case, all specifies that all atomic features of VMAF_feature are selected. A feature_dict dictionary can also contain more than one aggregate feature types.
model_param_file defines the type and hyper-parameters of the regressor to be used. For details, refer to the self-explanatory examples in directory resource/model_param. One example is:
model_type = "LIBSVMNUSVR"
model_param_dict = {
# ==== preprocess: normalize each feature ==== #
'norm_type':'clip_0to1', # rescale to within [0, 1]
# ==== postprocess: clip final quality score ==== #
'score_clip':[0.0, 100.0], # clip to within [0, 100]
# ==== libsvmnusvr parameters ==== #
'gamma':0.85, # selected
'C':1.0, # default
'nu':0.5, # default
'cache_size':200 # default
}
The trained model is output to output_model_file. Once it is obtained, it can be used by the run_vmaf or run_vmaf_in_batch, or used by run_testing to validate another dataset.
Above are two example scatter plots obtained from running the run_vmaf_training and run_testing commands on a training and a testing dataset, respectively.
The commands ./run_vmaf_training and ./run_testing also support custom subjective models (e.g. DMOS (default), MLE and more), through the submodule repository sureal. Read this paper for some background.
The subjective model option can be specified with option --subj-model subjective_model, for example:
./run_vmaf_training example_raw_dataset.py \
resource/feature_param/vmaf_feature_v2.py \
resource/model_param/libsvmnusvr_v2.py \
workspace/model/test_model.pkl \
--subj-model MLE --cache-result --parallelize
./run_testing VMAF example_raw_dataset.py --subj-model MLE --cache-result --parallelize
Note that for the --subj-model option to have effect, the input dataset file must follow a format similar to example_raw_dataset.py. Specifically, for each dictionary element in dis_videos, instead of having a key named 'dmos' or 'groundtruth' as in example_dataset.py, it must have a key named 'os' (stand for opinion score), and the value must be a list of numbers. This is the 'raw opinion score' collected from subjective experiments, which is used as the input to the custom subjective models.
python/script/run_vmaf_cross_validation.py provides tools for cross-validation of hyper-parameters and models. run_vmaf_cv runs training on a training dataset using hyper-parameters specified in a parameter file, output a trained model file, and then test the trained model on another test dataset and report testing correlation scores.
run_vmaf_kfold_cv takes in a dataset file, a parameter file, and a data structure (list of lists) that specifies the folds based on video content's IDs, and run k-fold cross valiation on the video dataset. This can be useful for manually tuning the model parameters.
You can also customize VMAF by plugging in third-party features or inventing new features, and specify them in a feature_param_file. Essentially, the 'aggregate' feature type (e.g. VMAF_feature) specified in the feature_dict corresponds to the TYPE field of a FeatureExtractor subclass (e.g. VmafFeatureExtractor). All you need to do is to create a new class extending the FeatureExtractor base class.
Similarly, you can plug in a third-party regressor or invent a new regressor and specify them in a model_param_file. The model_type (e.g. LIBSVMNUSVR) corresponds to the TYPE field of a TrainTestModel sublass (e.g. LibsvmnusvrTrainTestModel). All needed is to create a new class extending the TrainTestModel base class.
For instructions on how to extending the FeatureExtractor and TrainTestModel base classes, refer to CONTRIBUTING.md.
We provide a dataset publicly available to the community for training, testing and verification of results purposes. The dataset file is at resource/dataset/NFLX_dataset_public.py, and the videos (in YUV420P format) can be downloaded here.
We also provide an example dataset file containing video file names from VQEG (Video Quality Expert Group) HD3 videos. The dataset file is at resource/dataset/VQEGHD3_dataset.py, and the videos is available for downloading from http://www.cdvl.org/. After login, choose menu 'find videos', and search use keyword 'vqeghd3'. The dataset file includes from src01 to src09 except for src04, which overlaps with the Netflix Public Dataset, and hrc04, hrc07, hrc16, hrc17, hrc18, hrc19, hrc20 and hrc21, which are the most relevant distortion types to adaptive streaming. After downloading the videos, convert them to YUV420P format.
The VDK package combines feature extraction implementation in C and the rest scripting code in Python. The Python layer allows fast prototyping, but sometimes deploying the Python dependency in production is a pain.
Under wrapper, we provide a C++ implementation vmafossexec that has no dependency on Python.
Under root, run vmafossexec as:
wrapper/vmafossexec yuv420p 576 324
python/test/resource/yuv/src01_hrc00_576x324.yuv \
python/test/resource/yuv/src01_hrc01_576x324.yuv \
model/vmaf_v0.6.1.pkl \
--log vmaf_output.xml
For VMAF v0.6.1, the model file is model/vmaf_v0.6.1.pkl. The correspondence is documented here.
Note that vmafossexec depends on a shared library ptools/libptools.so (or on macOS, ptools/libptools.dylib). If you move the executable, make sure to include the shared library in LD_LIBRARY_PATH (or on macOS, DYLD_LIBRARY_PATH).
VMAF is now packaged into a library called libvmaf. You can install the library built using object files under wrapper/obj and ptools.
To install the library libvmaf.a run:
make install
This copies the library header libvmaf.h under usr/local/include, library libvmaf.a under user/local/lib and all the model files under usr/local/share. You can use the header libvmaf.h in your program. It contains an API which can be called from any C/C++ program:
int compute_vmaf(double* vmaf_score, char* fmt, int width, int height, int (*read_frame)(float *ref_data, float *main_data, float *temp_data,
int stride, void *user_data), void *user_data, char *model_path, char *log_path, char *log_fmt, int disable_clip,
int disable_avx, int enable_transform, int phone_model, int do_psnr, int do_ssim, int do_ms_ssim, char *pool_method);
Here, read_frame is a callback function which can be used to pass data from a program to VMAF. user_data is a program specific data that can be used by the callback function. For sample usage of compute_vmaf, refer to wrapper/src/main.cpp.
To uninstall the library run:
make uninstall
Install Docker, then, from the VMAF directory, run:
docker build -t vmaf .
And to use it, just run:
docker run --rm vmaf [CLI]
Where [CLI] is one of [run_vmaf, run_vmaf_training, run_vmaf_in_batch, run_testing, vmafossexec].
For example, if you are under root, to run run_vmaf on a sample reference/distorted video pair under resource/yuv:
docker run --rm -v $(pwd):/files vmaf \
run_vmaf yuv420p 576 324 \
/files/python/test/resource/yuv/src01_hrc00_576x324.yuv \
/files/python/test/resource/yuv/src01_hrc01_576x324.yuv \
--out-fmt json
To run vmafossexec with a specified model file:
docker run --rm -v $(pwd):/files vmaf \
vmafossexec yuv420p 576 324 \
/files/python/test/resource/yuv/src01_hrc00_576x324.yuv \
/files/python/test/resource/yuv/src01_hrc01_576x324.yuv \
/files/model/nflxall_vmafv4.pkl