[FFmpeg-soc] [PATCH] xiph packetizer

[Martin Storsjö]

On Thu, 29 Jul 2010, Josh Allmann wrote:

> > Also, there's an issue with vorbis timestamps if you stream copy from an
> > ogg file - not all packets have pts/dts set when reading them, and for
> > vorbis, you'd need to decode it to figure out the proper length of each
> > frame. In ffmpeg.c, lines 1622-1623, the pts is updated using an estimate
> > of the frame size, which doesn't turn out to be right (in my case at
> > least).
> >
> > If playing only vorbis over RTSP/RTP, it still worked, but the timestamps
> > advanced slowly from 0 to perhaps 1, then jumped to 5-6 seconds, advanced
> > slowly there again, and jumped onwards when the ogg demuxer returned a
> > packet that actually had a timestamp. Combined with a video stream, it
> > doesn't work too well in that setup of course.
> >
> > If live transcoding into theora/vorbis, and sending it out over RTSP/RTP,
> > it seemed to work really well, though!
> >
> 
> That's pretty strange, I would like to see your sample. With Big Buck
> Bunny, Vorbis standalone timestamps are fairly linear.

Can you give an url to this test clip?

> Using Theora and Vorbis together with RTSP/RTP, things work well.
> Theora standalone is still full of artifacts around motion with TCP,
> and UDP is a slideshow because of all the dropped packets. Tomorrow,
> I'll copy over the packetization routine from Feng and see if the
> results are more sane.

Hmm, that sounds really strange. Can you give the command lines that 
you've used for these test setups?

// Martin