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 libvmaf, we provide a C++ executable vmafossexec that has no dependency on Python.
To build vmafossexec, follow the instrunctions to build libvmaf.
You will find vmafossexec in libvmaf/build/tools/vmafossexec, run vmafossexec as:
libvmaf/build/tools/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 \
--psnr --ssim --ms-ssim \
--thread 0 --subsample 5
For VMAF v0.6.1, the model file is model/vmaf_v0.6.1.pkl.
The options --psnr, --ssim and --ms-ssim also allow reporting PSNR, SSIM and MS-SSIM results, respectively. The option --thread specifies the number of threads to use. Apply --thread 0 to use all threads available. The option --subsample specifies the subsampling of frames to speed up calculation. For example, --subsample 5 calculates VMAF on one of every 5 frames. The following plot shows the trend of how the subsample number impacts the processing speed (based on the Netflix Public Dataset of 1080p videos, with PSNR, SSIM and MS-SSIM calculation enabled):
Optionally, one can test vmafossexec by running the vmafossexec_test.py script
(this requires Python 3), run the following command
(which creates a virtual environment, and installs vmaf with all its dependencies into it):
rm -rf .venv/
python3 -mvenv .venv
.venv/bin/pip install pytest -e python/
.venv/bin/pytest python/test/vmafossexec_test.py
Expect all tests pass.
You can alternatively use python2 (not recommended), by changing the 2nd line above to this,
but you will need to install virtualenv module yourself:
python2 -mvirtualenv .venv