[MEncoder-users] new doom9 codec comparission (submission)

[Rich Felker]

On Thu, Dec 15, 2005 at 07:04:52PM +0100, Matthias Wieser wrote:
> > > Come on, it seems you are trying very hard to find excuses for lavc
> > > slightly loosing against xvid.
> >
> > No, this is not "slightly", it's a HUGE difference.
> 
> Ateme's mpeg4 filter is not so bad. In addition, all other mpeg4 codecs 
> would be affected the same way if Atme would use a strange dct. Most 
> affected would probably be xvid because of xvid's additional "asm. vs. c" 
> dct bug. If Atme supports FOURCC=FMP4 I think they have had done at least 
> some basic quality assurance to make sure the output of their filter is 
> comparable to lavc's own decoder.

What would FOURCC=FMP4 mean? lavc has several different dct
implementations and the user can choose among them. AFAIK the selected
one isn't even stored in the file, and which one is used may depend on
the os and cpu platform.

> > For YEARS AND YEARS all mplayer developers have known from experience
> > that lavc is superior in quality to xvid,
> 
> This topic shows up on the mailinglists from time to time but I can't 
> remember having seen numbers showing that lavc is better than xvid. All I 
> have seen was "we all know that lavc is better than xvid and that's why 
> lavc is better". Btw. does lavc support GMC?

Decoding, of course. Encoding, no. GMC is not useful and wastes a huge
amount of resources at decoding time. IIRC even someone who was
involved in designing mpeg4 admitted it was a big mistake but I may be
wrong on that.

> > while windows users have 
> > known the opposite. This idct discrepency would explain it.
> 
> Hm, many people use ffdshow for playback. Probably some of them use 
> sometimes another mpeg4 decoder, too. If there really would be such a big 
> difference in quality beteen ffdshow and other mpeg4 decoders somebody 
> would have spotted this already.

It is NOT A BIG DIFFERENCE IN QUALITY!!!!!!!!!!!!!!!!!!

Like I said, if you watch movies like a normal sane person you will
never notice it. When you notice it is when you do idiotic things like
saving single still frames and zooming them 4x in an imageviewer!

When the thread about xvid dct came up on ffmpeg list, I could not see
the difference in the two images without looking very closely, but the
guy who posted the "bug report" thought it was very important and I'm
sure pedants like doom9 will too.

> > Just a few hours ago Dominik was reencoding a h264 file he had to make
> > it playable on his "slow" laptop. He tried both lavc and xvid and got
> > vastly superior results with lavc (2.32 avg quant versus 2.82 -- avg
> > quant is generally not a good metric of quality, but with this much
> > difference it probably is, and he reported the lavc file looked better
> > too).
> 
> Very scientific test. Maybe he has used good lavc but bad xvid options. 
> Maybe xvid does not like this particular video very much. Maybe ...

It's not scientific, but once it happens enough it becomes common
sense. This is the common sense of the developers which has built up
over years and years of examples. We see xvid files off the net from
"professional" encoding groups, and then make lavc encodes that look
much better, etc. etc. etc.

> > So please stop trolling and accusing me of grasping at straws to prove
> > lavc is better.
> 
> Please stop telling other people they are trolling only because they have 
> an opinion different to yours.

Having a different opinion is not trolling. Belittling my opinion and
accusing me of exaggerating and grasping at straws is. I want this
matter investigated properly, that's all.

> > and if  
> > it's performing badly in doom9's test we want to get to the bottom of
> > his mistakes in testing methodology rather than accepting incorrect
> > results.
> 
> I agree, but one has to accept that there is the possibility that there 
> are other reasons why lavc loses against competing codecs.

Yes and I'd like to see why. But judging from experience I'm nearly
sure that it's bad testing methodology of one form or another.

> If the decoder would be the single cause for this then it should be no 
> problem to capture some snapshots which show the difference between 
> ffdshow and atme's decoder. We don't need to guess if atme is good or 
> bad. We can measure it.

I never even heard of atme before this test, wtf is it?

> Additionally we can investigate which features or settings made the other 
> codecs look better - but just saying doom9 has done the comparison only 
> 99% correctly and taking this as an lame excuse does not help anybody.

There's no such thing as 99% correctly. Either a comparison is correct
or it's not. If it's not then the results are meaningless.

> > > For many years lavc did not even have an own FOURCC - it has been
> > > nothing unusual to use any compatible Mpeg4 codec for playback.
> >
> > Yes, and it will look fine for watching. Doom9 is not testing casual
> > watching. He's zooming in on single still frames.
> 
> But I doubt that slight dct differences might be the reason for losing 
> against other codecs.

This is because you're just flaming rather than knowing the theory and
reading the relevant threads and looking at the sample images that
SHOW THE DIFFERENCE.

> At least the last comparison showed quite big 
> differences - for example some video codecs did not show any raindrops 
> while other codecs preserved those fine details. It's highly unrealistic 
> that small, accumulated dct errors make raindrops disappear.

We're only talking about two codecs here, lavc and xvid. I haven't
seen doom9's screenshots. Would you care to point me at them.

Rich