[FFmpeg-devel] Scaling of flashsv stream

[Benjamin Larsson]

Brian Candler wrote:
> I have been using ffmpeg to process screencasts created by vnc2flv using the
> flashsv codec, and I've found a tweak that allows them to compress much
> better after scaling.
> 
> My test 9.5 minute screencast has a frame size of 992x608. After a null pass
> through ffmpeg, the video portion is 10543K:
> 
>   ./ffmpeg -i in.flv -vcodec flashsv -acodec copy -r 12 -g 120 out.flv
> 
> If I scale this down by a factor of 2:1, the video still accounts for 9048K,
> which is hardly any smaller at all.
> 
>   ./ffmpeg -i in.flv -vcodec flashsv -acodec copy -r 12 -g 120 -s 496x304 out.flv
> 
> Poking around the codec, I discover that the scaled blocks don't zlib
> compress as well, presumably because ffmpeg is doing such a good job of
> anti-aliasing that a much wider colour palette is being used.
> 
> So I tried quantising the palette to 16x16x16 with the following patch, and
> after this the video is only 4086K.
> 
> --- libavcodec/flashsvenc.c.orig	2009-11-06 22:13:00.000000000 +0000
> +++ libavcodec/flashsvenc.c	2009-11-06 22:23:27.000000000 +0000
> @@ -88,7 +88,7 @@
>          npfptr = pfptr+(i*stride)+dy*3;
>          for (j=0 ; j<w*3 ; j++) {
>              diff |=npfptr[j]^nsptr[j];
> -            dptr[j] = nsptr[j];
> +            dptr[j] = ((nsptr[j] + 8)/ 17) * 17;
>          }
>          dptr += w*3;
>      }
> 
> The screen quality is hardly any different - if anything, the text is
> subjectively a little sharper.
> 
> Is there an existing colour quantisation facility in ffmpeg that I could
> have used to do this? If so please accept my apologies, and I will move over
> to ffmpeg-user :-)
> 
> Otherwise, do you think this would be worth adding? How would it be enabled?
> Does it belong in the flashsv encoder, or is there somewhere more
> appropriate to plug it in earlier in the chain?
> 
> Thanks,
> 
> Brian.

We have the upcoming avfilter framework. I think this feature would suit
there. As the flashsv format is lossless I think the only job the
encoder should perform is to compress as well as it can. Thus any patch
like this would be rejected. But a colorquant filter would be nice to have.

MvH
Benjamin Larsson