[FFmpeg-devel] [RFC] libavfilter-soc and variable-frame-size video streams

[Aurelien Jacobs]

On Sat, 13 Dec 2008 12:45:21 +0100
Stefano Sabatini <stefano.sabatini-lala at poste.it> wrote:

> On date Friday 2008-12-12 16:54:26 -0800, Jason Garrett-Glaser encoded:
> > > But maybe I'm just confused and there is a far simpler solution, or
> > > maybe we could just decide that supporting variable-frame-size video
> > > streams is not worth that hassle.
> > 
> > This sounds reasonable to me: thinking through what every filter would
> > have to do to support such a thing, it sounds completely
> > unmaintainable, especially for filters that act temporally: how will
> > they deal with the fact that cached temporal data from previous frames
> > (such as for temporal denoisers) is no longer the correct size
> > matching the current frame?!
> > 
> > Avisynth gets away just fine without variable frame size support: I
> > have actually never seen such a thing as a feature request either.  I
> > don't think libavfilter needs it either.
> 
> My only problem with this is that this way we cannot support video
> stream such as VP6, where the size is detected just when we parse the
> first decoded frame:
> http://thread.gmane.org/20080830132832.GA6663 at geppetto

If this really helps avfilter, I can modify the VP6 decoder so that
it don't zero out the container width/height before decoding first
frame. But that won't change the fact that width/height can change
later.

> I'm thinking about some mechanism for which an input filter may detect
> that the input size changed, and notify the output filter of that.

Without knowing libavfilter very well, I can imagine a simple
solution to resolution change.
Every filter should have a flag to declare wether it supports resolution
change or not. Then when a resolution change happens, the whole filter
chain is inspected. As soon as a filter which don't support resolution
change is detected, a scale filter is automatically inserted before
it. Some people may not like automagic stuff, so this feature could be
optionnal, and when not enabled, automatic insertion of the scale
filter would be replaced by a hard failure.

Aurel