[FFmpeg-devel] [PATCH] WebP native muxer bugfix: frames should have alpha blending off

Michael Niedermayer michaelni at gmx.at
Fri Jun 26 03:08:47 CEST 2015


On Thu, Jun 25, 2015 at 08:07:06PM +0000, Urvang Joshi wrote:
> On Thu, May 28, 2015 at 8:30 PM Michael Niedermayer <michaelni at gmx.at>
> wrote:
> 
> > On Thu, May 28, 2015 at 05:52:35PM +0000, Urvang Joshi wrote:
> > > On Wed, May 27, 2015 at 5:33 PM Michael Niedermayer <michaelni at gmx.at>
> > > wrote:
> > >
> > > > On Wed, May 27, 2015 at 03:10:05PM -0700, Urvang Joshi wrote:
> > > > > All the frames that the native muxer gets are fully reconstructed
> > > > > frames,
> > > >
> > > > wrong
> > >
> > >
> > > >
> > > > >  and they should not be alpha-blended with the previous frames.
> > > > >
> > > > > As per the WebP container spec
> > > > >
> > https://developers.google.com/speed/webp/docs/riff_container#animation,
> > > > > ANMF chunk should specify blending method = do not blend (and
> > disposal
> > > > > method = do not dispose).
> > > > >
> > > > > However, the native muxer was wrongly setting blending method = use
> > > > > alpha blending.
> > > > > This bug can be reproduced by converting a source with transparency
> > > > > (e.g. animated GIF with transparency) to an animated WebP, and
> > viewing
> > > > > it with vwebp.
> > > > > ---
> > > > >  libavformat/webpenc.c | 2 +-
> > > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > >
> > > > this breaks the encoder completely
> > > > the testcase is the same as previously but probably any testcase that
> > > > enables encoding multi frame animations will do
> > > > try  -cr_threshold 10000 -cr_size 16 for example
> > > >
> > >
> > > Ah, the problem seems to be for sources which don't have alpha, and then
> > > alpha is introduced by 'cr_threshold' and 'cr_size' for example. [There
> > are
> > > other cases too, but this is one example].
> > >
> > > I believe the logic for cr_threshold / cr_size is incorrect then,
> > > unfortunately. Here's why:
> > > 1. The original frame that the encoder gets (before possibly being
> > modified
> > > based on cr_threshold and cr_size) is fully-reconstructed, and should NOT
> > > be alpha-blended with the previous frame.
> > > [Yes, this is true. You can repro this bug by converting this GIF to WebP
> > > before this patch: http://dhelemann.de/images/Flug1.gif]
> > >
> > > For example, if this frame had a transparent pixel, it should be shown as
> > > transparent pixel and should NOT see-through the corresponding pixel from
> > > the previous frame. This would be achieved by setting blending method =
> > "do
> > > not blend".
> > >
> > > 2. On the other hand, based on the cr_threshold and cr_size, some pixels
> > > which are 'similar' to the corresponding pixels in the previous frame are
> > > modified to be transparent. So, this logic expects that the frame is
> > > alpha-blended with the previous frame (to see-through pixels from the
> > > previous frame).
> > >
> > > Clearly, the two requirements are contrasting and cannot be met.
> > >
> >
> > > One solution I can think of:
> > > (1) By default, set blending method = "do not blend"
> > > (2) Some pixels can be modified to become transparent ONLY IF the
> > original
> > > frame doesn't have any transparent pixels. And if some pixels are made
> > > transparent, we set blending method = "blend".
> >
> > sounds possible unless iam too tired and misunderstand
> >
> 
> Alright, I'm implementing this behavior then.
> 
> One question:
> - Whether the original frame had transparent pixels or not is determined in
> the encoder; but
> - The 'blending method' flag is written in the muxer.
> 
> Given this, how do I pass along this info (selected "blend method" for each
> frame) from encoder to muxer?

I would try to seperate demuxer and decoder in such a way that
the output from the encoder is self contained and can be fed to the
decoder and decode correctly. That is i would make the chunks that
are needed for decoding part of data that belongs to the
decoder/encoder layer, if that is possible



[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

The real ebay dictionary, page 1
"Used only once"    - "Some unspecified defect prevented a second use"
"In good condition" - "Can be repaird by experienced expert"
"As is" - "You wouldnt want it even if you were payed for it, if you knew ..."
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20150626/68d0ffb9/attachment.asc>


More information about the ffmpeg-devel mailing list