Quant 1 Mpeg smaller than H263?

[soujir0u]

Hi, I encoded two small clips with Constant Quantizer setting at 1, with H263 and Mpeg Quantization types. I found that H263 actually produces bigger filesizes than Mpeg for both the clips! (About 1MB more, the clips are only about 15 seconds) But for Constant Quant 2 and above, H263 files are smaller than Mpeg. Kinda strange...

[Koepi]

This sometimes happens and mislead -h to announe that mpeg quant type produces smaller output files. This is NOT the case. MPEG quant type is very sensitive to noise, and thus h263 quant type usually results in smaller clips as most sources are slightly noisy.

You just found a clip where this is true (very low noise) - you should play the lottery today, you're under a lucky star :)

Regards,
Koepi

[Hanty]

soujir0u

More interesting, how does the quality of the different clips compare?

[unplugged]

Originally posted by Koepi
You just found a clip where this is true (very low noise) - you should play the lottery today, you're under a lucky star :)

Tried 3 minutes matrix scene (720x424 avisynth+Vdub, YUV) with fixed quant 1, H.263 is 70% bigger than MPEG !!!!!

Tested with 2, this doesn't happen.

[unplugged]

I have made another (quadruple) test with matrix "plaza" action scene (reception attack) from frame 145419 to 150107 (3'07''), crop 720x424, fixed quant 1, Koepi's XviD-20020517 compile.

MPEG2AVI direct YV12-->YV12 results:
XviD_H263.AVI 346 Mb
XviD_MPEG.AVI 277 Mb

Avisynth (YUV2) + mpeg2dec.dll + VirtualDub 1.4.10 (fast-repack, YUV) results (same as GKnot package):
XviD_H263.AVI 340 Mb
XviD_MPEG.AVI 275 Mb

I think H.263 quantization matrix suffers a bit by that rare :) quantizer (1), could anyone make similar test with other stuff?


P.S.: small test with quant 2 H.263 gives slightly lower size than MPEG, negligible.

[Koepi]

Well, it seems that the image isn't well compressable when DCT'ed with a "constant" instead of a matrix like in mpeg at quantizer 1.

Nice to know, but useless.
I did _1_ encode which contained quant 1 and it is unwatchable on slower machines - it even sometimes stutters on my duron700.

I can't recommand the use of quant 1 - just want to mention this explicitly.

starting with quant 2, mpeg is in 99.999% of all cases bigger than h263 quant type.

I hope this helps,

regards,
Koepi

[soujir0u]

At Quant 1, I can't notice much difference between H263 and MPEG, but like Koepi said, Quant 1 isn't really very practical to use. But there are some scenes where Quant 2 is just too blocky, so I use Quant 1 to encode those scenes using the Credits tab option.

[Acaila]

starting with quant 2, mpeg is in 99.999% of all cases bigger than h263 quant type.
I've been running some tests on constant quantizers (even before this thread started) and so far with every scene I've done quants 15-31 are always bigger with H.263 than with MPEG.
I'm still running to see how the other quantizers perform (I started at 31...).

Also, quants 1-21 follow a size reduction with a decreasing pattern, however the quants from 22-31 are pretty much screwed up, no pattern whatsoever.
A scene encoded with quant 28 is always larger than even the same scene at quant 17.
But quant 31 still results in the smallest size though (I was happy to see that, because I always use 31 for my credits).

Ps. Always used Koepi's builds (if that even makes a difference with this).

[Koepi]

The code in XviD is a little weird when it comes to quantizers >20.

This is why you can't draw any conclusion fom tests with those quantizers.
For a more precise explanation head over to the xvid.org-forum, it was discussed recently.

Regards,
Koepi

[Hanty]

Well, does it really need any attention?
I can't personally think of any situation where >20 quantitizers are used that still produce satisfying results.

[Nic]

Only if you want to understand the reasoning...also I think gruel stated in a post here that even with quants >15 the codec is not appropriately "tuned". (I believe it was to do with the inter/intra decision making, but please, as koepi said, anyone interested look it up :)

-Nic