[FFmpeg-devel] donation for snow
Fri Nov 7 00:49:34 CET 2008
On Thu, Nov 06, 2008 at 10:30:58AM -0800, Jason Garrett-Glaser wrote:
> On Thu, Nov 6, 2008 at 10:04 AM, Lars T?uber <lars.taeuber at gmx.net> wrote:
> >> > > Ive no doubt that snow could beat x264 given a few determined and smart
> >> > > developers
> >> > I highly, highly doubt this, assuming you stick to the basic current
> >> > idea of Snow. I don't doubt that you or I could make a better format
> >> > than H.264 if I tried--there are dozens of places one could make
> >> > improvements--but what I do doubt is the ability to make a better
> >> > format *using overlapped wavelet*, since thousands of people have
> >> > tried that for something on the order of two decades and failed
> >> > miserably.
> >> I had no intent to stick to wavelets once someting else had been implemented
> >> that work better, actually i was aware that existing wavelets with existing
> >> ntropy coders perform poorly on inter frames before even writing the wavelet
> >> code.
> > Then what about using wavelets for intra frames only and other algorithms for the inter frames?
> Because ordinary wavelets are even worse for intra coding than for
> inter, since they don't have spatial prediction.
> There are some ideas floating around in the realm of directional
> wavelets and so forth, but nobody's come up with anything that can
> beat spatial prediction yet, at least as far as I know.
last i heard jpeg2000 beats h264 in intra coding quality per bitrate.
So from actual comparissions (which arent recent i have to admit though)
normal wavelets do beat h264s spatial prediction.
As you claim the opposite iam curious upon what that is based? If you want
iam pretty sure i could find the old comparissions.
about directional wavelets i must agree, iam not aware of anything
successfull in that area either.
If you want something better quality per bitrate wise, its not particularely
hard to define an adaptive transform that codes low to high resolution and
adapts its basis functions depending on correlation estimates based on the
already decoded lower resolution/frequency components.
The problem with this is just that it is computationally not too close to
Above is pretty much a KLT that is not based on a trivial correlation
of pixels but one where the KLT is "redone" with the remaining unknown
orthogonal components, after each frequency component
is decoded and thus more information becomes available about the image
which can be used to refine the predicted pixel correlation pairs.
Such transform, should end up with basis functions that would follow
spatial correlations and this would not be limited to squares and 8
directions, but really could handle anything like curves or even
discontinuous patterns. Thus in theory should vastly outperform anything
that could be done with h264 style spatial prediction.
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
The educated differ from the uneducated as much as the living from the
dead. -- Aristotle
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 189 bytes
Desc: Digital signature
More information about the ffmpeg-devel