[x264-devel] Global optimization of keyframes

[Jason Garrett-Glaser]

On Mon, Dec 27, 2010 at 5:51 AM, Nicolas George
<nicolas.george at normalesup.org> wrote:
> Le sextidi 6 nivôse, an CCXIX, Jason Garrett-Glaser a écrit :
>> This idea has a few problems.
>>
>> x264 currently makes all frametype decisions in the first pass.  It
>> also makes various decisions *based on* these frametype decisions,
>> such as MB-tree.  Changing frametypes in the second pass would require
>> either ignoring the inaccuracy of the other decisions, or redoing them
>> all.
>
> That makes things difficult indeed.
>
>> Secondly, I don't really get your example.  Why would you want
>> keyframes at all in that case?  Keyframes are for scenecuts, and your
>> described video has no scenecuts.
>
> Unless I am mistaken, if the scene is too long, the encoder is expected to
> add keyframes at regular intervals to allow seeking (and, for some codecs
> but not H.264, tu avoid drifting due to rounding errors). Or am I using the
> wrong words.
>
> A stupid encoder would just put these periodic keyframes as far apart as
> possible, but a good encoder would place them sometimes early if it detects
> a better place. That is the selection of these "better places" that I think
> would benefit from global optimization.

This is less useful with long keyframe intervals (like x264 uses by
default).  It's also less useful when one isn't relying on keyframes
as a "quality boost" mechanism.

x264 uses MB-tree, which does exactly what you're looking for -- it
tracks references between frames and raises quality on areas of the
frame that will be referenced more in the future.  This effectively
solves that problem.

Jason