XviD 05042002 -> EPSZ, EPSZ^2, MV hints!

[Koepi]

Tadaa!

Now this is worth a thread of it's own again!

-h added interlacing support and MV hints to XviD!

Grab it as long as it's hot (follow the XviD link in my signature)!

XviD-05042002-4:
- EPSZ and EPSZ^2 activated.
- New code supports MV hints, resulting in a 30% speed gain on 2nd pass!

Best regards,
Koepi

[-h]

And yes I know the vfw interface is getting ugly, with some controls enabled while others are disabled.

It's just more fun working on core stuff :)

I'll give vfw a polish this weekend hopefully, unless something more fun shows up (which will probably happen).

-h

[-h]

Er, I'd advise against anyone using this build for 2-pass stuff at the moment, it's impossible to do a 2nd pass *without* using hinted ME at the moment.

I just committed a fix to CVS, but Koepi will have to update his binary (again) ;)

Shouldn't be long, maybe 15 minutes.

-h

[Selur]

little question, will motion vector hinting some how degrade quality?
Don't think&hope so, or ?
(because if not, why should one not want to use it ? ;) )

Cu Selur

[Koepi]

The impact on the quality is very low.
First tests show that the filesize at fixed quantizers raised max ~1.2%.

But that's what we're doing here, developing and trying to find out those things.
Remember, it's not even beta stage.

Koepi

[Nic]

http://nic.dnsalias.com

for -h's corrected version.

-Nic

[Selur]

Yeah, know it's not even beta and I wouldn't use it if I wouldn't like testing around,...

and thx for the info ;)

Cu Selur

[Koepi]

XviD-05042002-5:
- Integrated -h's fix so 2nd pass can be done without MV hint file as well ;)

[FuFo]

I'm download , download, download, again download, .....he he he...
damn speed new xvid build, lol. Upgrade every 30 min...ufff

Great works all

[wing1]

:D u gotta luv these guys!

@-h
Just did some tests with interlace from uManiac's instant build, and it looks good.

@Nic
It seems that on win98 system your DS filter works fine using builds from 3/30/2002 and forward. However, on my W2K system the DS filter refused to work for some reason. It only works if I register bicubicResizer.ax.

[debris]

Just did some tests with hinted MVs (Koepi's 5th build from today) and at least with 2-pass/modulated quant/6-ultra high and everything else standard using hinted MVs gives heavy heavy artifacts not to say completely garbled output. Green, red and block blocks crawl over the screen, whole parts of the image jump up and down. No problems with same settings but without hinted MVs though.

But I must say: I really appreciate you guy's work!!! :)
Getting better every build and the builds come out on an hourly basis today ;)

[Koepi]

Uh, yeah, totally forgot about that:

For hinted MVs you should disable modulated quant type and use the quant type you used on first pass (either MPEG, H263 or Custom, but not at all modulated!).

Please test that again. Do the errors still occur?

If yes, try using search precision 5.

Are they gone now?

If still not, then EPSZ is incomaptible with the MV hints now and we (hehe, of course -h I mean) still has to work on it ;)

Best regards,
Koepi

[-h]

Just did some tests with hinted MVs (Koepi's 5th build from today) and at least with 2-pass/modulated quant/6-ultra high and everything else standard using hinted MVs gives heavy heavy artifacts not to say completely garbled output. Green, red and block blocks crawl over the screen, whole parts of the image jump up and down. No problems with same settings but without hinted MVs though.

Ouch, haven't seen that before! Modulated quant you say..

But I must say: I really appreciate you guy's work!!! :)
Getting better every build and the builds come out on an hourly basis today ;)

Heh at this rate I could probably commit any old crappy broken code and someone would say thanks ;)

-h

[-h]

@-h
Just did some tests with interlace from uManiac's instant build, and it looks good.

Yeah it'll still be a bit slow though. I'll add the assembly this weekend, and if that isn't fast enough, a checkbox that bypasses the field decision altogether and assumes everything on the screen is an interlacing artifact :)

-h

[rui]

Originally posted by Koepi
Uh, yeah, totally forgot about that:

For hinted MVs you should disable modulated quant type and use the quant type you used on first pass (either MPEG, H263 or Custom, but not at all modulated!).

Please test that again. Do the errors still occur?

If yes, try using search precision 5.

Are they gone now?

If still not, then EPSZ is incomaptible with the MV hints now and we (hehe, of course -h I mean) still has to work on it ;)

Best regards,
Koepi

:(

I really like the modulated. Oh well, i will have to live without MV hints for now.

[Doom9]

just in case.. I ran a small test... 20k frames, matrix.. same avs as in the codec comparison. also.. same xvid settings (besides the quantizers of course.. I used h.263 as modulated is not possible using mv files)

first pass: 13 minutes, 2nd pass: 10 minutes, average fps: 28.98fps..
but since I couldn't use the same quantizers it doesn't necessarily mean it's faster.. after all.. epsz^2 was used whereas in the test it was pmvfast. Also.. the same 20k frames take 12:13 using modulated quantizers in the 03/17 build.

[canadian_fbi]

some basic quality test results of the new hinted mv:

test video clip: the matrix kung-fu scene (it was on my hd)
test bitrate: 1000 kbps
curve compression used: normal

normal mv avg quant: 3.298
hinted mv avg quant: 3.351

so for this particular scene it gave about a 1.6% decrease in quality. i didn't time the encodings as i was doing other stuff at the time so it wouldn't have been terribly accurate either, but it seemed to be about 20% faster.

hmm, so now we have to choose between 1.6% worse quality or 20% faster encodes... :)