[theora] NHW Project - lower quality settings

[Raphael Canut]

Hello,

I have re-tested -l4 high compression setting and it's clear that it lacks
of precision on degraded, rather blurred images. So I don't know if it is a
good idea to base the other lower quality settings (-l5,-l6,...) on -l4
setting.I have tested the NHW codec against x265, x264, Daala, WebP,
Rududu, DLI and it's clear that at high compression these very good codecs
have more precision.But they also tend to blur/decrease neatness-sharpness
of image where the NHW codec would enhance image neatness/sharpness.But
that's right that on degraded, blurred images where there is no neatness to
enhance, the results start to be not good and lack of precision.So for now,
I recommend to use -l4 setting on well-defined, good quality images.

I have also "simply" increased the wavelet coeff threshold of the dead-zone
for -l5 setting (to save -5Ko), but the quality drop is too big for now.The
other codecs (above) don't have that drop in quality from -l4 to -l5
compression.So apparently, there needs to be more smartness to pass from
-l4 to -l5 with the NHW codec.I'll try to work on this, but it's extremely
complex, for me...For example, Rududu doesn't have this problem, the codec
source code has been made available recently on GitHub, I have started to
study it and it is very powerful: advanced SPIHT tree-like coding, advanced
variance-based context modeling, advanced range coding, and other things...

So to sum up, all the very good codecs quoted above have a very good
precision at high compression but they tend to decrease neatness/sharpness,
on the contrary of the NHW codec, but on degraded images where there is no
neatness/sharpness, the NHW codec then lacks of precision...

Last but not least, for now the big advantage of the NHW codec I see, is
that it is very fast.The NHW codec is way faster than the codecs above, for
example it is still now H.264 technology that is used in mobile devices,
and the NHW codec compared to ultra-optimized x264 is x3 times faster to
encode and x1.2 times faster to decode, that is to say that with good
optimization (SIMD, multithreading) the NHW codec will be x12 times faster
to encode and x5 times faster to decode than x264...

Again working on high compression is very complex, and any help would be
very welcome.And I know there are experts out there, more skilled than me,
that could have very good ideas for the NHW Project, if they could find the
time...

Do not hesitate to contact me!

Many thanks!
Cheers,
Raphael
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.xiph.org/pipermail/theora/attachments/20180309/33ef0722/attachment.html>