0

I am looking for a VMAF-like objective user-perception video quality scanner that functions at scale. The use case is a twitch-like streaming service where videos are eligible to be played on demand after the live stream completes. We want to have some level of quality in the on demand library without having to view every live stream. We are encoding the livestreams into HLS playlists after the stream completes, but using VMAF to compare the post-stream mp4 to the post-encoded mp4s in HLS doesn't provide the information needed as the original mp4 could be of low quality due to bandwidth issues during the live stream.

JeffN
  • 1,575
  • 15
  • 26

1 Answers1

3

Clarification

Not sure if I get the question correctly. You want to measure the output quality of the transcoded video without using the reference video. Is that correct?

Answer

VMAF is a reference quality metric, which means it simply compares how much subjective distortion was introduced into the transcoded video when compared to the source video. It always needs a reference input video.

I think what you are looking for is a no-reference quality metric(s). Where you can measure the "quality" of video without a reference source video. There are a lot of no-reference quality metrics intended to capture different distortion artifacts in the output video. For example, blurring, blocking, and so on. Then you can make an aggregated metric based on these values depending upon what you want to measure.

Conclusion

So, if I were you, I would start searching for no-reference quality metrics. And then look for tools that can measure those no-reference quality metrics efficiently. Hope that answers your question.

Adithyan Ilangovan
  • 169
  • 1
  • 9
  • Fundamentally/mathematically, a naive computer can't distinguish between desired detail and unwanted artifacts (which is the idea behind measuring quality). So, you need to either tell the computer what the video is supposed to look like (with a reference video), or tell it what the artifacts look like (no-reference quality metrics). // Theoretically, you could train some ML to judge it the human way - based on rich life experience of what real objects with real lighting look like - but that's obviously out of the question for now. – sssheridan Jun 29 '22 at 06:17