[FFmpeg-devel] evaluating the experimental status of ffv1 version 3

Peter B. pb at das-werkstatt.com
Tue Oct 2 10:57:49 CEST 2012

Quoting Michael Niedermayer <michaelni at gmx.at>:

> On Sat, Sep 29, 2012 at 06:53:54PM +0200, Peter B. wrote:
>> I've now repeated the HD-and-beyond tests with the current git-head and
>> it seems that there actually are some problems with 2160p files (not
>> just "crowd_run").
>> When selecting 30 slices, the file can't be transcoded to FFv1, because
>> ffmpeg is segfaulting.
> fixed

HD-and-beyond tests with videos from Derf's collection at xiph.org [1]  
are now finished and every generated FFv1.3 file matches its original!  

This means that YUV (8bit, 420p, 422p, 444p) has now been tested for  
resolutions from qcif (176x144) up to 2160p (3840x2160).

The testsuite execution code is posted on github [5] and was run with  
the following parameters for FFv1 encoding:

FFV1_GOP_SIZES="1 300"
FFV1_SLICES="4 6 9 12 16 24 30"

For each input file, 112 output files were generated and their  
framemd5 frame-checksums were matched against the original input  
file's framemd5 checksums.

As far as I've seen, the material from Derf's collection is purely  
non-interlaced and 8bit-per-component only.
For 16bit tests, I will use videos from the Swedish Television (SVT)  
collection [2] and excerpts of Sintel's 1k/4k files [3], but I have to  
read them from image sequences. SGI files for SVT and PNG for Sintel.

For YUV422 10bit, I'll use the a converted set of the SVT collection  
from the "Video Quality Experts Group" (VQEG) FTP server [4].

Material missing for the YUV tests so far is:
YUV420p10, YUV444p10 (both LE and BE), as well as 420,422 and 444 with  
9bits (also both LE and BE).

I could generate this material from 16bit sources, using ffmpeg.

Feedback welcome,

== References:
[1] http://media.xiph.org/video/derf/
[2] http://media.xiph.org/svt/
[3] http://media.xiph.org/sintel/
[4] ftp://vqeg.its.bldrdoc.gov/HDTV/SVT_exports/SVT_YUV10_Exports_

More information about the ffmpeg-devel mailing list