VMAF Re-implementation on PyTorch: Some Experimental Results

Note, that these estimates in principle less susceptible to such preprocessing. The original VMAF have to be computed over the sample of images. Instead, the algorithm was implemented in C [3] and no effort is known assumption is made that the estimates can be computed over to us to re-implement it fully, i.e., including all its sub-metrics the patches ([4], section IV; [5]) using some ML framework. One of the reasons for that is VIF is computed on four scales by downsampling the image; the claimed non-differentiability of this metric. We propose four values per frame are used as features for final score an implementation of VMAF using PyTorch and analyze regression. The original version of VIF included the wavelet its differentiability with various methods. We also discuss transform, but the same authors released another version of potential problems related to the computation of this metric VIF in the pixel domain [6]. VMAF uses only the pixel domain in the end of the paper.