[theora] NHW codec - progress 2

[Raphael Canut]

Hello,

Just a quick message to correct what I have said in my previous post.It
finally seems not that easy to code small wavelet coefficients to have a
good precision, and SPIHT is certainly a very good algorithm for that, -and
I know there are SPIHT experts that can make very good SPIHT-based
codecs-.I have chosen a new and different approach, the multistage residual
coding, so I will try to keep it and get the best of it.

I estimate that I need an additional 10Ko in average to code low contrast
information (and avoid artifacts).So I will need to remove 10Ko of
information in high contrast and give it to low contrast.I think in a first
time, I will only use residual coding for the low contrast area errors.I
will try to do this next month if I can find time, and will let you know
what it gives.

I am also aware that with all this, it will be very very difficult to beat
HEVC, VP9/10, Daala.The entropy coding is certainly the most interesting
part of my codec.

Cheers,
Raphael

[Raphael Canut]

Hello,

Sorry for the late answer.I don't have advanced with low contrast
information coding in the NHW codec, primarily because I can not target
specifically these areas... But I think it's a choice, between neatness and
precision, and it's not the same algorithms for each.I have chosen a new
approach: neatness, to (try to) bring a new point of view for image
compression, and it can be half-successful actually, because at -h2 setting
for the nhw codec where there is full-neatness, results are visually better
for me than BPG (HEVC intra), WebP, and other codecs...

But the NHW codec doesn't scale very well to high compression.In this
domain, AOM AV1, HEVC and VP10 are monsters of compression, while still
keeping the CABAC entropy coding to ensure a "fast", reasonnable
decoding.Very quickly, is that true? I have read on another forum that
Google VP9/10 and then AOM AV1 use the CABAC entropy coding scheme as in
HEVC.I have checked the code rapidly, and I can see the subexp (exp-golomb)
binarization, the lookup-table-based arithmetic coding,... But this must be
untrue because AOMedia and Google claim their works are royalty-free (while
the CABAC scheme is heavily patented)... Any clarification would be welcome.

To quickly finish, the other great advantage of the NHW codec is that it is
fast, a lot faster than these codecs above... But can not compete with them
for high compression...

Cheers,
Raphael

Post by Raphael Canut
Hello,
Just a quick message to correct what I have said in my previous post.It
finally seems not that easy to code small wavelet coefficients to have a
good precision, and SPIHT is certainly a very good algorithm for that, -and
I know there are SPIHT experts that can make very good SPIHT-based
codecs-.I have chosen a new and different approach, the multistage residual
coding, so I will try to keep it and get the best of it.
I estimate that I need an additional 10Ko in average to code low contrast
information (and avoid artifacts).So I will need to remove 10Ko of
information in high contrast and give it to low contrast.I think in a first
time, I will only use residual coding for the low contrast area errors.I
will try to do this next month if I can find time, and will let you know
what it gives.
I am also aware that with all this, it will be very very difficult to beat
HEVC, VP9/10, Daala.The entropy coding is certainly the most interesting
part of my codec.
Cheers,
Raphael