[FFmpeg-devel] FFv1.3: Final testing - Results/Multipass

Peter B. pb at das-werkstatt.com
Mon Nov 5 23:14:00 CET 2012


2 things:

1) FFv1.3 results - distilled:
I've read Michael's mails about how to summarize the results again and
they focused on GOP-size and multipass results, so I figured that it'd
make sense to leave only one combination of coder/context in the
condensed table.
As coder=1 is required for encoding >8 bit per-component, I thought I'd
only consider coder=1,context=1 for the summary.

Objections?

2) Multipass results:
For testing multipass performance compression-wise, I took 3 longer
videos (1min, 10min), as the videos from Derf's-, or SVT's-collection
are very, very short.

I've prepared 3 videos for multipass tests from our archive material,
where the rights should be cleared:
1) 1min VHS vertical pan over Vienna's Heldenplatz
2) 1min DigiBeta documentary (different movement)
3) 10min VHS (medium-bad condition) of a home-video with very different
movement patterns

As posted previously, I think the results are quite impressive: The
2pass files are around ~50 MiB per minute smaller. Wow!
With slices it is such that more slices may be faster to compute, but
increase the filesize. When applying 2pass, GOP=1 files are getting as
small (or even smaller sometimes) than their equivalents with GOP=300.

The passlog files are quite huge, though: ~70 MiB per minute. So my 10
minute test video has a pass logfile of more than 700 MiB.

The filesizes for multipass with GOP=1 and GOP=300 are available as
semicolon-separated CSV here:
http://download.das-werkstatt.com/pb/mthk/ffv1_stats/20121105-ffv1_stats.csv

Feedback welcome,
Pb

PS: I've added the encoding duration instead of encoding FPS, as it
shows an approximate of the speed of FFv1.



More information about the ffmpeg-devel mailing list