XviD's performance in the codec comparison

[diNgdonG]

Watching the amazing XviD encoded screenshots in the latest codec comparison, I immediately realized the desperate wish to expand my encoding horizon.
After quite a bit of reading ... this is exactly what happened to my machine:


Athlon XP1800+, 512MB DDR333, GF4 Ti4200, XPproSP1
Source: MPEG2 720x576, 25fps, interlaced
DVD2AVIdg
AviSynth 2.53 Filters: Lanczos, Field Deinterlace
GordianKnot 0.28.7
VDubMod 1.5.10.1 build2366
2-Pass encoding, AVI, Target Bitrate 1400 kbps, speed 14-20fps
XviD-1.0-Beta3-26122003

Profile AS@L5: MPEG, Qpel, b-frames: max(2) ratio(1.50) offset(0.75), closed GOV
2 Pass Tuning: I-frame: boost(10), reduction threshold(13), reduction factor(25), max overflow 25/25, curve compression 0/0/10
Advanced Def.: 6-UltraHigh, VHQ-4, chroma motion, max I-frame interval(250), turbo, quantization: 2,31,2,31,2,31
XviD Decoder: ZoomPlayer v2.80, XviD-Dec-1.0-Beta3 DS-Filter, deblocking activated
... everything else was off or default!


My aim was to repeat the codec comparison on my own, creating a maximum quality result within a reasonable 2-3 CD size limit.
The outcome of my encoding session was very nice, but less impressive than what I was yearning for.
XviD was in deed a bit more detailed but the picture had far more blocks and noise as well.
Even after some heavy tuning efforts I couldn't get XviD to beat the good old 3.11 codec at equal bitrates.

Maybe this is related to what people in the forum said about the Xvid codec not being specially designed for high bitrate encoding
and known quality loss with interlaced source material.

As my 3.11 encoding quality gets close to what I've seen in the codec comparison I still have no clue how they made XviD perform so much better.

[temporance]

Did you use the exact same parameters as doom9 published in the comparison? I think your settings and scripts are different. If you used a higher bitrate then it's going to get harder to see differences between the codecs. Also what was your source movie?

I tried to replicate some of doom9's encodes (using identical scripts and codec settings) and my results were pretty good. I tried DivX and xvid which both looked very nice on the Matrix test and my encode was very close to the comparison's screenshots.

I can't say I can come to the same conclusion as doom9: in fact I can't really call a winner between the two codecs I tried. I think it's too difficult to say there was any real difference DivX vs. xvid. Anyway, the two codecs had completely different distribution of bitrate throughout the movie - maybe the best codec is the one that leaves the best impression after watching the whole movie (!).

My speedy conclusion:

http://forum.doom9.org/showthread.php?s=&postid=423390#post423390

If I were you I'd try to replicate doom9's encodes precisely (same source, scripts, settings) to see if you agree with his conclusion or can add any useful feedback. Then move on to new tests.

[mikeX]

maybe the two pass tuning is to blame
also if you are looking for max quality, disabling turbo might be a good idea, the devs say it gives unnoticeable quality loss, but you never know

[diNgdonG]

As I don't own any of the 3 DVDs doom9 used as source I have to do my encoding on interlaced DVB MPEG2 material,
so I won't be able to create a comparable sequence - sorry.

I tried a very lot of different settings including -no turbo-, leading to the 2ndPass tuning settings,
wich I extracted from the forum for high quality encoding. Reduction- threshold(13) and factor(25) were explained to be ideal for 25fps.
A low MaxOverflow value(20-30) is ought to help as well, but I'm not sure wich OverflowControlStrength works best.

Even with the exact settings doom9 used my encodings refuse to make any valuable progress and I still get a much too noisy picture.
It looks a bit like the decoding isn't that optimized. Don't know if it makes sense to try a different decoder ?
In comparison the divX output is a bit blurry, but very smooth, has fewer blocks and leaves the far better overall impression.

But in my opinion the codec comparison showed that there is still a way to make perfect encoding happen, so I refuse to give up on XviD so easily!

[Koepi]

Hit "load defaults" again and just change VHQ=4 (1 would suffice IMO), use trellis quant - since you don't give the end resolution I think you might use a very high one, so you may disable qpel as well(qpel makes sense if you resize to 640xYYY). Use Chroma ME, don't use turbo. Ah, set iframe boost to 5% as well. MPEG quant type is nice, you can also try custom quant matrix "HVS GOOD PICTURE".

That should give you pleasant results.

Regards
Koepi

[Didée]

Originally posted by diNgdonG
Athlon XP1800+, 512MB DDR333, GF4 Ti4200, XPproSP1
Source: MPEG2 720x576, 25fps, interlaced
DVD2AVIdg
AviSynth 2.53 Filters: Lanczos, Field Deinterlace
...
Ouch.

Before drawing any conclusions, you should begin with *proper* de-interlacing.

If your source is telecined film, you must perform IVTC, e.g. with 'telecide'.
If your source is truly interlaced video, then FieldDeinterlace is one of the approbriate solutions ... but then you should not at all compare the picture quality to those screenshots Doom9 shows in the comparison, because these are two different kind of shoes. Simple as that.

- Didée

[Siku]

Originally posted by Didée
Ouch.

Before drawing any conclusions, you should begin with *proper* de-interlacing.

If your source is telecined film, you must perform IVTC, e.g. with 'telecide'.
If your source is truly interlaced video, then FieldDeinterlace is one of the approbriate solutions ... but then you should not at all compare the picture quality to those screenshots Doom9 shows in the comparison, because these are two different kind of shoes. Simple as that.

- Didée

But the source is 25 FPS. Is there such thing as 'telecined film' in PAL movies? I have always performed Field Deinterlace to interlaced PAL movies. If there is better solution, please share your knowledge.

Regards,
Siku

[diNgdonG]

It's interlaced 25fps PAL DVBs material - recorded from satellite, so field deinterlace should be fine.
I chose -no blend- assuming one can leave that to the codec ?

My end resolution is 544x368 meaning it is not really big and I think Didee is right and the quality shouldn't be compared to the codec comparison, due to it's interlaced PAL origin.
But once again this sounds like XviD has a true weakness when it comes to interlaced sources.
If I'm able to improve quality by encoding from DVD this could proof that it's obviosly better to use divx for DVBs TV recordings?
So far 3.11 gives me the perfect result, but who isn't eager for a change...

I'll try soon, Koepi's settings are next, thanks!

[zettai]

Originally posted by diNgdonG
But once again this sounds like XviD has a true weakness when it comes to interlaced sources.

This statement makes no sense. You are DEinterlacing your footage and thus your source is not interlaced... so exactly where is the weakness?

Of course, saying you can't reproduce doom9's results when you are not using the same source is like saying you cant make a meal taste as nice as someone else with the same oven when all the ingredient you have used are cheap artificial substitutes.

I'm just saying there may be more flaws to your encoding process than the xvid settings.

[DevilsChild]

Haven't had any trouble with XviD and interlaced sources. I just select IVTC through Gordian Knot. There is also an "interlacing" option in the XviD profile tab.

1400 kbps should give excellent quality, even at a high resolution. I would just load the defaults and use 2-pass with VHQ1, Chroma Motion, 1 max b-frame, trellis, adaptive quantization, and "Turbo" mode.

[Doom9]

Futurama is both PAL and Interlaced as well. The AviSynth script is available on page one on the codec comparison. But obviously you can't compare a DVB broadcast with a DVD source... DVDs are usually of better quality and that does have an effect.

[diNgdonG]

As someone suggested I rechecked my testing environment and found some way to improve it.
Encoding from a different source without turbo and using Trellis quantization the performance is now getting close to what I was looking for.
I never used Trellis before, cause documentation said it would only be helpful with H.263. How about the HVS Good Picture matrix ?

The XviD encode is very nice now, specially in high detailed scenes, like rain forrest, there is a holy lot of detail.
And I got the impression that the background is now very stable, meaning there are far less shifting elements,
compared to divx en/decoded scenes.
Only in some low detail areas like walls the additional smoothness of the divx decoding sometimes looks better.
There are still some problems with mosquito noise so I'm going to try this filtering script, wich was recommended by Koepi.

LoadPlugin("C:\PROGRA~1\GORDIA~1\mpeg2dec3.dll")
LoadPlugin("C:\PROGRA~1\GORDIA~1\UnFilter.dll")
LoadPlugin("C:\PROGRA~1\GORDIA~1\UnDot.dll")
LoadPlugin("C:\PROGRA~1\GORDIA~1\Convolution3d.dll")
mpeg2source("H:\DVD\Projects\FOREST_VTS_02_PGC1\FOREST_ForceFilm.d2v")
crop(4,62,716,348)
LumaFilter(-2)
Convolution3d(preset="movieHQ")
UnFilter(-5,-5)
UnDot()
LanczosResize(640,256)

But maybe some mosquito type artefacts are the price to pay for such an amount of detail. You can't have it all !
It's getting very interesting...