0

I used to check and adjust the quality of my video encodings with an Avisynth script calculating the SSIM index. For some reason, since moving to Windows 10, the performance is much lower. Then I found out about the ffmpeg ssim filter, which runs at least ten times faster. But I have these issues:

  • I can't find what algorithm is used by either method or what are the differences,

  • More importantly, when I plot the results of each method for a number of video files, the corrrelation is virtually zero. With the avisynth SSIM, I find a good correlation between the h.264 CRF factor (as expected), but that correlation disappears with the ffmpeg SSIM.

Has anyone verified whether the ffmpeg SSIM filter really works?

EDIT: I need to investigate further on the second point. I actually found no correlation for ffmpeg over a small sample of different videos, however, there is an excellent correlation for both methods between CRF and SSIM for a single video sample.

JMor
  • 47
  • 1
  • 11
  • According to source code both are based on algo by Zhou Wang et al. FFMPEG, however, [approximates](https://github.com/FFmpeg/FFmpeg/blob/master/libavfilter/vf_ssim.c#L23) overlapped 8x8 block sums, rather than the original gaussian weights. – wOxxOm Oct 29 '16 at 20:22
  • Thanks for the reply, that would explain why it's so much faster. But it seems that, given the significant differences in results, the ffmpeg approximation leads to less accurate results. – JMor Nov 05 '16 at 15:05

0 Answers0