[FFmpeg-user] live transcoding alignment

[Carl Lindqvist]

2014-07-28 9:38 GMT+02:00 Claudiu Rad <jazzman at misalpina.net>:
> hello all,
>
> using ffmpeg i want to live transcode an input live stream into multiple
> variants and push all of them (including the original) to the client.
> everything works all right except for one thing: i want all keyframes to be
> properly aligned between streams but it seems that over time, they get
> misaligned which is probably as expected because it would make sense for
> ffmpeg to treat each output and do its keyframe generation quite
> independently. however, the first question:
>
> 1. how can i enforce keyframe alignment between output transcoded
> (reencoded) streams in ffmpeg? i don't need P/B or other frame types to be
> aligned, but I-frames should be exactly on the same frame numbers on all
> streams
>
> to extend this even further:
>
> 2. how can i enforce keyframe alignment also with input which i would like
> to simply -c copy ? this means that reencoded outputs should have the
> exactly same keyframe placements as the input, thus, internal codec
> heuristics should be disabled on this matter.
>
> i am talking about a h264/x264 input/output.
>
> thanks much.
>
> --
> Claudiu

I do live encoding with ffmpeg via rtmp to Akamai. I have not been
able to use the original stream, but that was not a problem in my case
since it was an mpeg2 stream from multicast or raw video from a
capturecard and, needed to be transcoded anyways. I have had more
problems with the bug where rtmp and issue #1604.

But I have had good success on the current git version with something
like this, if you can live with transcoding the original stream
(example in PAL SD format):

ffmpeg  -i udp://X.X.X.X:XXXX^
 -filter_complex "[0:0]
yadif=0:0,split=4[v1][v2][v3][v4];[v1]scale=720:576[v1o];[v2]scale=720:432[v2o];[v3]scale=640:360[v3o];[v4]scale=384:216[v4o];[0:1]
asplit=4[a1][a2][a3][a4]"^
 -map "[v1o]" -map "[a1"] -pix_fmt yuv420p^
 -r 25 -c:v libx264 -preset veryfast -x264opts keyint=50:min-keyint=50
-crf 18 -maxrate 3372k -bufsize 7000k -profile:v high -aspect 16:9^
 -c:a libvo_aacenc -b:a 128k -ac 2^
 -f flv "rtmp://p.epXXXXXX.i.akamaientrypoint.net/EntryPoint/XXXXX_1_3500@XXXXXX
flashver=FMLE/3.0\20(compatible;\20FMSc/1.0) live=true pubUser=XXXXX
pubPasswd=XXXXX"^
 -map "[v2o]" -map "[a2"] -pix_fmt yuv420p^
 -r 25 -c:v libx264 -preset veryfast -x264opts keyint=50:min-keyint=50
-crf 18 -maxrate 2372k -bufsize 5000k -profile:v main -aspect 16:9^
 -c:a libvo_aacenc -b:a 128k -ac 2^
 -f flv "rtmp://p.epXXXXXX.i.akamaientrypoint.net/EntryPoint/XXXXX_1_2500@XXXXXX
flashver=FMLE/3.0\20(compatible;\20FMSc/1.0) live=true pubUser=XXXXX
pubPasswd=XXXXX"^
 -map "[v3o]" -map "[a3"] -pix_fmt yuv420p^
 -r 25 -c:v libx264 -preset veryfast -x264opts keyint=50:min-keyint=50
-crf 18 -maxrate 1404k -bufsize 3000k -profile:v main -aspect 16:9^
 -c:a libvo_aacenc -b:a 96k -ac 2^
 -f flv "rtmp://p.epXXXXXX.i.akamaientrypoint.net/EntryPoint/XXXXX_1_1500@XXXXXX
flashver=FMLE/3.0\20(compatible;\20FMSc/1.0) live=true pubUser=XXXXX
pubPasswd=XXXXX"^
 -map "[v4o]" -map "[a4"] -pix_fmt yuv420p^
 -r 10 -c:v libx264 -preset veryfast -x264opts keyint=20:min-keyint=20
-crf 18 -maxrate 252k -bufsize 500k -profile:v baseline -aspect 16:9^
 -c:a libvo_aacenc -b:a 48k -ac 2^
 -f flv "rtmp://p.epXXXXXX.i.akamaientrypoint.net/EntryPoint/XXXXX_1_300@XXXXXX
flashver=FMLE/3.0\20(compatible;\20FMSc/1.0) live=true pubUser=XXXXX
pubPasswd=XXXXX"

I hope it can be of use to someone. Keyframes seem aligned at every 2
seconds, and HLS from Akamai seem to work fine. The rates above are
not fine tuned, I have mostly been experimenting.

Your question is interesting, though. It could be really useful for
adaptive streaming to align keyframes to a source and keep that as the
most high quality stream.

/Carl L