For building, you need the following:
- Python3 (3.6 or higher)
- Meson (0.56.1 or higher)
- Ninja (1.7.1 or higher)
- NASM (for x86 builds only, 2.13.02 or higher)
xxd
Follow the steps below to set up a clean virtual environment and install the tools:
python3 -m pip install virtualenv
python3 -m virtualenv .venv
source .venv/bin/activate
pip install meson
sudo [package-manager] install nasm ninja-build doxygen xxd
You need to invoke [package-manager]
depending on which system you are on: apt-get
for Ubuntu and Debian, yum
for CentOS and RHEL, dnf
for Fedora, zypper
for openSUSE, brew
for MacOS (no sudo
).
Ninja package name might be ninja
or ninja-build
, depending on your system. Check the pre-built ninja packages before running the install command.
Run:
meson setup build --buildtype release
Special cases:
- add
-Denable_float=true
flag in the rare case if you want to use the floating-point feature extractors. - add
-Denable_avx512=true
to support wider SIMD instructions to achieve the fastest processing on supported CPUs - add
-Denable_cuda=true
to build with CUDA support, which requiresnvcc
for compilation (tested with CUDA >= 11) - add
-Denable_nvtx=true
to build with NVTX marker support, which enables easy profiling using Nsight Systems
Build with:
ninja -vC build
Build and run tests with:
ninja -vC build test
Install the library, headers, and the vmaf
command line tool:
ninja -vC build install
This will install the following files:
├── bin
│ └── vmaf
├── include
│ └── libvmaf
│ ├── feature.h
│ ├── libvmaf.h
│ ├── model.h
│ ├── picture.h
│ └── version.h
└── lib
├── libvmaf.1.dylib
├── libvmaf.a
├── libvmaf.dylib -> libvmaf.1.dylib
└── pkgconfig
└── libvmaf.pc
Generate HTML documentation with:
ninja -vC build doc/html
libvmaf
now has a number of VMAF models built-in. This means that no external VMAF model files are required, since the models are compiled into and read directly from the library. If you do not wish to compile the built-in models into your build, you may disable them with -Dbuilt_in_models=false
. Previous versions of this library required a .pkl
model file. Since libvmaf v2.0.0, these .pkl
model files have been deprecated in favor of .json
model files. If you have a previously trained .pkl
model you would like to convert to .json
, this Python conversion script is available.
A command line tool called vmaf
is included as part of the build/installation. See the vmaf
README.md for details.
Create a VmafContext
with vmaf_init()
. VmafContext
is an opaque type, and VmafConfiguration
is a options struct used to initialize the context. Be sure to clean up the VmafContext
with vmaf_close()
when you are done with it.
int vmaf_init(VmafContext **vmaf, VmafConfiguration cfg);
int vmaf_close(VmafContext *vmaf);
Calculating a VMAF score requires a VMAF model. The next step is to create a VmafModel
. There are a few ways to get a VmafModel
. Use vmaf_model_load()
when you would like to load one of the default built-in models. Use vmaf_model_load_from_path()
when you would like to read a model file from a filesystem. After you are done using the VmafModel
, clean it up with vmaf_model_destroy()
.
int vmaf_model_load(VmafModel **model, VmafModelConfig *cfg,
const char *version);
int vmaf_model_load_from_path(VmafModel **model, VmafModelConfig *cfg,
const char *path);
void vmaf_model_destroy(VmafModel *model);
A VMAF score is a fusion of several elementary features which are specified by a model file. The next step is to register all feature extractors required by your model or models with vmaf_use_features_from_model()
. If there are auxillary metrics (i.e. PSNR
) you would also like to extract use vmaf_use_feature()
to register it directly.
int vmaf_use_features_from_model(VmafContext *vmaf, VmafModel *model);
int vmaf_use_feature(VmafContext *vmaf, const char *feature_name,
VmafFeatureDictionary *opts_dict);
VMAF is a full-reference metric, meaning it is calculated on pairs of reference/distorted pictures. To allocate a VmafPicture
use vmaf_picture_alloc
. After allocation, you may fill the buffers with pixel data.
int vmaf_picture_alloc(VmafPicture *pic, enum VmafPixelFormat pix_fmt,
unsigned bpc, unsigned w, unsigned h);
Read all of you input pictures in a loop with vmaf_read_pictures()
. When you are done reading pictures, some feature extractors may have internal buffers may still need to be flushed. Call vmaf_read_pictures()
again with ref
and dist
set to NULL
to flush these buffers. Once buffers are flushed, all further calls to vmaf_read_pictures()
are invalid.
int vmaf_read_pictures(VmafContext *vmaf, VmafPicture *ref, VmafPicture *dist,
unsigned index);
After your pictures have been read, you can retrieve a vmaf score. Use vmaf_score_at_index
to get the score at single index, and use vmaf_score_pooled()
to get a pooled score across multiple frames.
int vmaf_score_at_index(VmafContext *vmaf, VmafModel *model, double *score,
unsigned index);
int vmaf_score_pooled(VmafContext *vmaf, VmafModel *model,
enum VmafPoolingMethod pool_method, double *score,
unsigned index_low, unsigned index_high);
For complete API documentation, see libvmaf.h. For an example of using the API to create the vmaf
command line tool, see vmaf.c.
To write a new feature extractor, please first familiarize yourself with the VmafFeatureExtractor API documentation.
Create a new VmafFeatureExtractor
and add it to the build as well as the feature_extractor_list[]
. See this diff for an example. Once you do this your feature extractor may be registered and used inside of libvmaf
via vmaf_use_feature()
or vmaf_use_features_from_model()
. To invoke this feature extractor directly from the command line with vmaf
use the --feature
flag.
VmafFeatureExtractor
is a feature extraction class with just a few callbacks. If you have preallocations and/or precomputations to make, it is best to do this in the .init()
callback and store the output in .priv
. This is a place for custom data which is available for all subsequent callbacks. If you allocate anything in .init()
be sure to clean it up in the .close()
callback.
The remainder of your work should take place in the .extract()
callback. This callback is called for every pair of input pictures. Read the pixel data make some computations and then write the output(s) to the VmafFeatureCollector
via the vmaf_feature_collector_append()
api. An important thing to know about this callback is that it can (and probably is) being called in an arbitrary order. If your feature extractor has a temporal requirement (i.e. motion
), set the VMAF_FEATURE_EXTRACTOR_TEMPORAL
flag and the VmafFeatureExtractorContext
will ensure that this callback is executed in serial. For an example of a feature extractor with a temporal dependency see the motion feature extractor.
If the VMAF_FEATURE_EXTRACTOR_TEMPORAL
is set, it is likely that you have buffers that need flushing. If this is the case, .flush()
is called in a loop until something non-zero is returned.