[MPlayer-users] short mbd=2/mv0/trell/cbp with and without qprd benchmark

[Corey Hickey]

Tuukka Toivonen wrote:
> You made interesting tests but it's quite difficult to understand them
> because when both the bitrate and PSNR decrease it's quite impossible
> to say whether it's better or worse.

I've actually put some thought into this. Right now I'm too lazy to turn
words into math and write it into my benchmark script, but it shouldn't
be too hard.


1. For any given set of lavc parameters that is functioning properly,
the relationship between bitrate and PSNR is a logarithmic curve.
An exception to this would be when trying to encode a clip at bitrates
that are way too low; in this case the curves jump around a bit until
the bitrates are sufficiently high. For the purposes of testing a
parameter at reasonable bitrates, the exception can probably be ignored.

2. When comparing encodes that differ only slightly in bitrate (say,
+-5 kbit/sec), we can safely assume that bitrate vs. PSNR is linear. The
changes in slope over such a range are quite minimal.

3. When the set of lavc parameters is altered in any way (except for
anything that alters bitrate), the slope of the curve at any given
bitrate changes. I don't know if this is even worth accounting for, but
the slope could be re-calculated after each parameter change if
necessary.


So, we should be able to estimate, with reasonable precision, what the
PSNR of each test would be if all the average bitrates were to have been
exactly the same. This "adjusted PSNR" might be slightly more
representative of how effective a certain parameter is.

Comments?

-Corey