[MPlayer-users] Re: Tools for mpeg* to DVD

[Corey Hickey]

rcooley wrote:

>>B frames require big motion estimate search radii and are
>>unfriendly to predictive motion estimators 
> 
> 
> I think it's pretty much universal opinion of the more knowledgable
> users on this mailing list, that B-frames are useless.  I personally
> haven't yet seen a case where B-frames improve quality and/or reduce
> filesizes.
> 
> 

I don't know about mpeg-2, but with mpeg-4 a single b-frame usually
increases PSNR slightly (while more than one is almost always bad).

The table below is for a one-minute clip I usually test on, that I
selected to be an even mix between fast and slow motion scenes.

+------------------+-------+-------+
| params           | psnr  | fps   |
+------------------+-------+-------+
| :vmax_b_frames=1 | 41.48 | 63.11 |
| :                | 41.34 | 74.68 |
| :vmax_b_frames=2 | 41.26 | 59.68 |
| :vmax_b_frames=3 | 41.01 | 57.79 |
| :vmax_b_frames=4 | 40.71 | 56.97 |
| :vmax_b_frames=5 | 40.42 | 56.83 |
+------------------+-------+-------+


As for the price/performance ratio, see the table below for a comparison
to trell, mbd=2, and mbd=1.

psnr_% = the psnr for each test divided by the psnr of ":"
fps_%  = the fps for each test divided by the fps of ":"
asdf   = the psnr increase for each test with respect to ":",
          divided by the fps decrease with respect to ":"
          (so, this is "psnr gain for each fps lost")

+------------------+-------+-------+--------+-------+-----------+
| params           | psnr  | fps   | psnr_% | fps_% | asdf      |
+------------------+-------+-------+--------+-------+-----------+
| :trell           | 41.63 | 58.98 | 100.70 |  0.79 |  0.001015 |
| :mbd=2           | 41.54 | 56.94 | 100.48 |  0.76 |  0.000821 |
| :vmax_b_frames=1 | 41.48 | 63.11 | 100.34 |  0.85 |  0.000334 |
| :mbd=1           | 41.43 | 62.29 | 100.22 |  0.83 |  0.000234 |
| :                | 41.34 | 74.68 | 100.00 |  1.00 | -0.000000 |
| :vmax_b_frames=2 | 41.26 | 59.68 |  99.81 |  0.80 | -0.000264 |
| :vmax_b_frames=3 | 41.01 | 57.79 |  99.20 |  0.77 | -0.001270 |
| :vmax_b_frames=4 | 40.71 | 56.97 |  98.48 |  0.76 | -0.002582 |
| :vmax_b_frames=5 | 40.42 | 56.83 |  97.77 |  0.76 | -0.003809 |
+------------------+-------+-------+--------+-------+-----------+


While the performance isn't exactly stellar compared to "the big two",
trell and mbd=2, it does edge out mbd=1 (note that that's 334 vs 234, if
your eyes are trying to trick you like mine just did to me).

Don't kill me if I messed up the math. :)

-Corey