Consider Changing the VMAF Scale to [1, 101], or (0, 100]

[rfliam]

As noted in the best practices post, the faq, and the associated issue the harmonic, geometric or other means may be desirable to compute.

However the harmonic mean, power mean, and many others can not construct an underlying metric on a space including 0.

I currently map any 0 value to a slightly larger non 0 value to calculate these means. The "0" value doesn't seem to have any real "meaning" in VMAF, unlike say in resistors in parallel. Certainly one 0 frame does not imply the quality of all frames in the sequence have a "0" perceived value. This would not affect the arithmetic mean greatly either.

In particular I often see sequences such as this:

Where an encoding contains approximately the same average VMAF, but a much tighter distribution of frames. This maps to much better perceived quality from viewers. The harmonic mean captures this well, but can not be used on sequences with 0 frames as is.

[li-zhi]

@rfliam As you pointed out, a practical solution is to "+1" in your harmonic VMAF calculation. You can +1 for the input VMAF values, and -1 on the final output VMAF value. This way, you can maintain the VMAF scale to [0, 100].