[x264-devel] Re: x264 vs JM 7.5c encoding results

[Loren Merritt]

On Mon, 3 Jan 2005, Tuukka Toivonen wrote:

> On Thu, 30 Dec 2004, Loren Merritt wrote:
>
>> not CIF. You might see a different amount of improvement.)
>
> There aren't many standard sequences with higher resolution than CIF,
> it would be nice if others could repeat the tests (->standard sequences).

Repeatability is nice and all, But I'd rather test compression of the 
content that I'd actually like to compress -- real movies. And yes, that 
includes the ability to deal with MPEG2 artifacts.

>>> P.S. I wonder what are the major reasons for difference between the two 
>>> encoders?
>> Mostly that JM uses rate/distortion optimization, whereas x264 uses SATD.
>> JM also uses a slower motion search.
>
> So x264 doesn't use rate distortion to either mode selection nor the motion
> vector estimation? Why not? I don't think that RD-optimization is so slow,
> at least for motion estimation. It's just a table lookup and an addition to
> the SAD/SATD per one checking point.

x264 does use the "rate" part of rate/distortion: It takes into account 
the bits needed for various decisions (including motion estimation).
But it doesn't do the "distortion" part: For mode selection, the residual 
is compared with rate + SATD, whereas JM fully encodes all modes and 
compares rate + quantization error. And for decimation (cbp, p_skip 
detection), x264 just looks at the quantized residual, not the 
quantization error.

I am currently trying to do full rate/distortion, and even that is only 
about 2x slower than the current code. But it doesn't yet improve 
compression, so there must be a bug somewhere that I haven't tracked down.

--Loren Merritt

-- 
This is the x264-devel mailing-list
To unsubscribe, go to: http://developers.videolan.org/lists.html