5

I've been encoding some clips and then analyzing them with the VMAF tool from Netflix, but I don't know how to understand the results.

This is the overall mean section of one of my tests.

VMAF_feature_adm2_score:0.982934
VMAF_feature_motion2_score:2.262897
VMAF_feature_vif_scale0_score:0.727423
VMAF_feature_vif_scale1_score:0.946992
VMAF_feature_vif_scale2_score:0.975502
VMAF_feature_vif_scale3_score:0.986295
VMAF_score:94.240835

3 Answers 3

3

In general, the higher the VMAF score, the higher the quality of the video will be perceived by the viewer. Lower VIF component and ADM scores indicate problems caused by intra-frame compression artifacts. A lower motion score indicates problems caused by inter-frame compression artifacts.

3
  • What scale is VMAF_feature_motion2_score on? I can tell the others are from 0-1, but not what the maximum of VMAF_feature_motion2_score is.
    – chew socks
    Jun 5, 2018 at 20:49
  • I don't know what the scale is. It doesn't seem to be described in the blog post, but you should be able to determine it from the source code. Jun 6, 2018 at 19:16
  • 1
    I found it in the README, it varies from 0 to 20.
    – chew socks
    Jun 12, 2018 at 23:05
3

The VMAF models are actually trained by humans who rated the quality on a 1-5 point scale, watching the movie either on an 1080p screen from 3x screen height away, or on a 4k screen from 1.5x screen height away. This rating is mapped to VMAF scores 0, 20, 40, 60, and 80, while 100 is when they compare the source video to itself. Your score of 94.240835 is somewhere between what humans rated the highest and lossless.

The VIF scores refer to Visual Information Fidelity but I don't know what the scale numbers mean there. Don't know about adm2 or motion2 scores.

2

The numbers don’t mean anything. They only have value when compared to other values. You can compare the results between two encodes of the same file and determine which encode produced better results.

5
  • How do I decide what a good threshold is? Like, If I make encodes with varying bitrates, I know more is better, but how much is enough?
    – chew socks
    Jun 5, 2018 at 20:54
  • That’s subjective. You get to decide what is “good enough” for your case.
    – SlimSCSI
    Jun 5, 2018 at 21:57
  • If you wanted to encode something so that your average non-expert wouldn't be able to tell the difference from the original, what numbers would you target? ;)
    – chew socks
    Jun 6, 2018 at 3:31
  • Like @SlimSCSI said, the scores are relative to the source material. There is no specific number to shoot for. If you are tweaking the encoder on a case by case basis, you can compare different settings by score. If you are setting up a "factory" workflow, run several test files through to find the better settings. If you are doing a QC workflow, you will have to experiment with your threshold for your specific set up. Jun 6, 2018 at 19:12
  • 1
    I found on the mux.com blog that a difference of 6 is barely noticeable. "A 6-point difference is VMAF is considered a Just-Noticeable-Difference."
    – chew socks
    Jun 12, 2018 at 23:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.