[MPlayer-users] cbp tests

[Corey Hickey]

Michael Niedermayer (high overlord of lavc mpeg-4) just added a
"Rate distorted optimal coded block pattern" option. I don't understand
the particulars of what that means, but I decided to give a few tests of
trell+cbp vs. mbd=2 vs. all the other options I usually use.

The 200-second clip I used has low graininess but lots of fast motion
and moderate amounts of chaotic changes. It's probably reprentative of
non-grainy movies with lots of fast motion and moderate amounts of
chaotic changes. Your results WILL vary. :)

All the encoding lines look like this, plus more lavcopts:
for i in 1 2 ; do mencoder test.vob -ss 205 -endpos 200 -aid 128 -vf \
crop=714:352:2:62,scale=688:288 -sws 1 -ovc lavc -lavcopts \
vbitrate=800:vcodec=mpeg4:vpass="$i":psnr -oac copy -ofps 23.976 -o \
small"$i".avi ; done

For clarity, I'll only report the additions to lavcopts below; thus, the
first test has no changes from the command above. Also for clarity, I
will collectively reference the following options as <stuff>:
v4mv:precmp=3:cmp=3:subcmp=3:vmax_b_frames=1:qpel:vlelim=-2:vcelim=7:\
lumi_mask=0.05:naq:vqcomp=0.7:mbcmp=2

The PSNR line listed is for the second pass.

<no changes to lavcopts>
PSNR: Y:39.13, Cb:43.22, Cr:43.91, All:40.17

mbd=2
PSNR: Y:39.44, Cb:43.56, Cr:44.31, All:40.49

trell
PSNR: Y:39.36, Cb:43.26, Cr:43.97, All:40.37

mbd=2:trell
PSNR: Y:39.62, Cb:43.56, Cr:44.34, All:40.64

trell:cbp
PSNR: Y:39.37, Cb:43.09, Cr:43.82, All:40.35

mbd=2:trell:cbp
PSNR: Y:39.52, Cb:43.27, Cr:44.05, All:40.51

<stuff>
PSNR: Y:39.78, Cb:43.24, Cr:43.96, All:40.72

<stuff>:mbd=2
PSNR: Y:40.08, Cb:43.64, Cr:44.41, All:41.04

<stuff>:trell
PSNR: Y:40.07, Cb:43.31, Cr:44.05, All:40.97

<stuff>:mbd=2:trell
PSNR: Y:40.08, Cb:43.64, Cr:44.41, All:41.04


NOTE: After I ran the following two tests, I noticed that the encoding
was seriously screwed up, and reduced the problem down to vlelim/vcelim.
I reported the bug, Michael fixed it, and I re-ran the two tests. The
results below are from after the bugfix.

<stuff>:trell:cbp
PSNR: Y:40.10, Cb:43.30, Cr:44.03, All:40.99

<stuff>:mbd=2:trell:cbp
PSNR: Y:40.33, Cb:43.65, Cr:44.42, All:41.25


My perception is neither discerning nor unbiased enough to properly
detect visual differences between the tests; I showed them to my friend,
and the best I could get out of him was that the early tests looked good
and the later tests looked really good. Regardless, I can certainly
notice no qualitative loss when cbq is enabled.

Decide what you want based on the PSNR figures. I think it's interesting
that cbp contributed a 0.21 PSNR increase to the last test, though. :)

Depending on how much time I have, I might do some more tests with
different clips, and throw mv0 into the mix. Is anyone else testing
these options?

-Corey