WebM format for Record

[Jon Nettleton]

On Sat, Apr 9, 2011 at 11:16 AM, Benjamin M. Schwartz
<bmschwar at fas.harvard.edu> wrote:
> On 04/09/2011 02:00 PM, Daniel Drake wrote:
>> Ben, what are your thoughts on Theora vs WebM for Record? You
>> mentioned at one point that Theora developers were considering
>> focusing on the low-power arena (being displaced by WebM on the
>> high-power high-res front), is that direction being taken?
>
> Essentially, yes.  The WebM encoder (libvpx) has always been much slower
> than libtheora because VP8 is much more computationally intensive than
> Theora.  The libtheora devs have been working to expand this difference by
> adding encoder speed optimizations.

Google has also worked on speed optimizations for the vp8 encoder.
The version of libvp8 included in Fedora 14, might be in
updates-testing, includes these optimizations.  You will see in my
patch I am using speed setting 2 which is a single pass encoding meant
for live streaming.  I am so far very impressed with the system
demands as well as the quality.  With speed set to 2, and quality at 5
( middle of the road ) I was able to record full 2 minute clips at Low
Res, High Res, and also did some testing at 640x480.  400x300
recordings were using about 70-80% cpu utilization encoding in the
WebM format.

Google is also working on ARM optimizations that will be important for
the XO 1.75

>
> The latest versions of libtheora (1.2 alphas, and maybe even more in SVN
> head) contain a lot of work to speed up encoding.  I have recommended
> several times that OLPC ship 1.2-prerelease libtheora in the default
> build, due to the significant speed improvements.

I did test this version out but found the improvements on our hardware
to be marginal.  I know there was talk about constant bit rate vs vbr
possible needing to be tweaked.  I didn't have the time to look into
it then.

>
>> So encoding speed is quite a big deal
>
> Indeed.  I don't see WebM as being a realistic option for Record until
> OLPC gets WebM encoders in silicon, which will presumably be after XO-3.

You can see in this video  http://www.youtube.com/watch?v=ilmMrtlzSRg
that the WebM on the XO 1.5 is really working better than our current
release.  You will hear a few pops in the audio, those points were
when we would previously lose all sink between audio and video.  In
this video the sink isn't perfect but relatively close.

Oh and there was a question about backwards compatibility.  I know
there would be work testing it, but I see no problems with backward
compatibility.  We would just have to check file types and adjust the
pipeline accordingly.  We would also have to adjust the mimetype
properties to any Activities that may hard code them.

-Jon