[theora] NHW Codec - new study on 30 images

[Raphael Canut]

Hello all,

It's me again.Just a quick message to tell that I finally do not agree that
HEVC (don't have tested AOM AV1 for now) is better than the NHW codec at
mid compression.

Last night, I dowloaded 30 images of Star Wars Rogue One on the web and I
made another comparison with x265 (HEVC) for mid compression (-l2 to h2
settings for the NHW codec).
Visually, I prefer the results of NHW codec on nearly 18 images out of 30
(60%), I find that more neatness is more pleasant, I prefer the results of
x265 on 11 images (35%), more precision is more pleasant, and on 1 image
the result of NHW codec is quite bad (looks terrible, lots of aliaising,
annoying imprecisions).

First, do you agree with these results? Second, is it acceptable that a
codec is better than another in 60% of the cases but perform quite badly in
5% of the cases? I would be very interested in your opinion on this subject.

Very quickly to finish, my codec really need improvements at some stages.It
really needs to have better YUV420 down/upsampling algorithms (don't know
about image Super-Resolution upsampling with machine learning or
convolutional neural network, they're excellent but too slow for image
compression I fear..., there are also other algorithms that are very
good)... For the lower quality settings (-l1,-l2 mainly, don't use -l3,-l4
they're not good), the quantization and rounding of the wavelet
coefficients in the higher order according to the residual error of the
first order wavelet DC image is incomplete and so there is a lack of
precision.This is on the TODO list...

Really, I would be very interested in your thoughts on can a codec perform
good or very good in 95% of the cases but perform bad in the 5% remaining
cases.

Many thanks!
Cheers,
Raphael
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.xiph.org/pipermail/theora/attachments/20170224/f510c751/attachment.html>