[XviD-devel] What B-frames give

[Radoslaw 'sysKin' Czyz]

Hi

I updated my p/b/i decision and did some more tests.
I-frame decison is now based on the old sad/deviation comparsion and
works very good - just as good as the old ME-based version. I'm not
going to play anymore with it, if someone writes something better
(like histogram-based) then great.

As for P/B decision, I tried sum of squared SADs again (it just looks
so much more logical to me) but it gave me worse results - so it stays
the way it was yesterday.

Some tests. All testes were made with dvd-resolution and noisless
(compared to vhs) clips.
I encoded them at fixed quant of 3, with b-frame ratio=134% (quant 4 -
I hope) and compared filesizes:

1. matrix lobby
max_frames = 0 : 23857 kb
max_frames = 1 : 21197 kb
max_frames = 2 : 20783 kb
max_frames = 3 : 20754 kb

2. warcraft III trailer
max_frames = 0 : 22659 kb
max_frames = 1 : 20240 kb
max_frames = 2 : 19894 kb
max_frames = 3 : 19873 kb
max_frames = 4 : 19904 kb
max_frames = 5 : 19947 kb

3. a bit of anime (taken directly from divx so the test isn't very
   serious)
max_frames = 0 : 12710 kb
max_frames = 1 : 11638 kb
max_frames = 2 : 11159 kb
max_frames = 3 : 11098 kb
max_frames = 4 : 11112 kb
max_frames = 5 : 11157 kb

The visual quality of b-frame encodings is a bit higher imho (!)
because I don't see any smearing artifacts. I don't see any
degradation of b-frames' quality (well the quantizer is a bit higher
after all) either.

The file increase with max_bframes -> infinity shows that my P/B
decision is not perfect - a perfect one wouldn't allow any file
increase when given less limitations.

What do you think?

Radek

PS. I forgot - the gain is higher with quant ratio == 200%, while I
still don't see any quaity degradation. I just wanted more fair
comparsion.