[HD-DVD Challenge] MPEG2, VC-1 and H264 with real uncompressed source movie

[benwaggoner]

Originally Posted by CruNcher

Yes and that's also why the PEP Encoder is not 100% the same like the consumer encoder it is especialy tuned for Film Content the same for Cinevision and Cinemacrafts Encoder and in those regards X264 is still @ the start it is more tuned for Anime then anything else tough it's high bitrate visual results with Film Content improving constantly since some time now (and also it's very flexible due to it's wide variety of settings that can be adapted for certain source encoding scenarios)

Actually, the VC-1 Encoder SDK is a pretty close derivative of PEP (although it was forked before the most recent PEP release). There's a lot of the film-tuned quality improvements in there, even if not all apps chose to expose them. But even a prosumer-esque app like Expression Encoder 2 exposes the PEP-derived features like Adaptive Deadzone, Differential Quantization, and our noise reduction filter.

VC-1 Encoder SDK is absolutely meant to enable consumer/professional apps to provide VC-1 Blu-ray encoding quality like using PEP in 2-pass mode.

[Sagittaire]

Originally Posted by benwaggoner

Actually, the VC-1 Encoder SDK is a pretty close derivative of PEP (although it was forked before the most recent PEP release). There's a lot of the film-tuned quality improvements in there, even if not all apps chose to expose them. But even a prosumer-esque app like Expression Encoder 2 exposes the PEP-derived features like Adaptive Deadzone, Differential Quantization, and our noise reduction filter.

VC-1 Encoder SDK is absolutely meant to enable consumer/professional apps to provide VC-1 Blu-ray encoding quality like using PEP in 2-pass mode.

Well I have big difference at low bitrate between VC1 from SDK and VC1 from PEP. I use this command line with VC1 SDK:

@REM -----------------------------------------------------------
@REM
@REM Profil BluRay 1080p23.976 2 passes extra high quality
@REM
@REM -----------------------------------------------------------


@REM Source file name (suffit de mettre la source ici)
set E_SRC=Encodage_HD_NTSC_1080p.avs

@REM Set of bitrates (ici le bitrate)
set E_BR=10000

@REM Set of max bitrates (ici le bitrate max)
set MAX_BR=24000

@REM Set of Buffer (ici le buffer)
set BUF_BR=3750000

@REM Profil (ici le nom des fichiers de sortie)
AVS2ASF.exe -i %E_SRC% -o azerty.vc1 -rate %E_BR% -peakrate %MAX_BR% -vbv %BUF_BR% 
-framerate 23.976 -ratecontrol 3 -profiletype 2 -maxkeydist 24 -bframes 1 -bdeltaqp 1 -adaptiveGOP 
-keyPop 2 -inloop 1 -overlap 1 -complexity 5 -motionsearchlevel 2 -mesearchmethod 1 -mbcost 1 
-mvcost 1 -mvrange 4 -adaptivequant 1 -dquantoption 1

pause

and equivalent setting for pep/pse. Bug in avs2asf.exe ... ?

[Golgot13]

Hi Ben,

NAB is finish, so don't forget us we wait the video.
About LV, the Wynn hotel is very nice and there is a golf !!!
I hope you tested it.

About H264 encoder, did you see the new Sony and Thomson encoder ?
My contact said me this encoder are good, so better than PeP/PSE,
because he use it but not for long time now...

[benwaggoner]

Speaking of which, here's my presentations from NAB:

http://www.on10.net/blogs/benwagg/22040/

[benwaggoner]

...and here's a torrent of an ATSC-compliant encode of the 1080i source, for a better example of what a final clip looks like.

http://216.99.212.233:6969/torrents/...B3AB54B0B8D367

[Sagittaire]

Originally Posted by benwaggoner

Speaking of which, here's my presentations from NAB:

http://www.on10.net/blogs/benwagg/22040/

Well my 2 cents:

MPEG-2
• Standard for HD broadcast for 10+ years
• Well understood
• Mature tools
• Approaching limit of compression efficiency
• Transparent compression requires much higher data
rates than other codecs
• Needs 50-100% more bitrate than VC-1 or H.264
• Limits quality for longer titles on SL discs
• But MPEG-2 can look great @ BD‟s 40 Mbps peak
• Remains useful for transcoding consumer content
• From HDV or ATSC without recompression

In fact with good encoder MPEG2 at HD resolution/bitrate never needs 100% more bitrate. MPEG2 is a really good codec for BD50 encoding. Can be a really good codec for BD25 encoding with medium/short movie.

VC-1
• VC-1 is SMPTE designation (SMPTE 421M-2006)
• Licensing handled by MPEG-LA
• Standardization of Windows Media Video 9
• Microsoft works closely with major studios to support
VC-1 encoding within their workflow
• Best codec for preserving texture detail at lower bitrates
• We expect wide support in HD authoring tools
• VC-1 Encoder SDK free for license
• Used on most HD DVD titles, roughly 1/3rd of Blu-ray
• Can make single VC-1 encode for both HD DVD and BD
• Warner VC-1 BD titles use same encode as HD DVD

I have Rate Control problem with VC1 SDK. VC1 from pep/pse produce by far better quality with equivalent setting.

H264 is by far the best codec in high quantisation encoding scenario. Cabac, more powerfull inloop, wpred, multiref produce are big advantage in this case.

You can use the same encoding for VC1, MPEG2 and H264 for HDDVD and BD.

AVC/H.264
• AVC, H.264, MPEG-4 Part just different names
• Defined by MPEG and ITU
• High 4:2:0 profile supported in BD/HD DVD
• Adds 8x8 blocks to Main Profile‟s 4x4 blocks
• Improves ability to handle textures
• Main Profile behind VC-1/MPEG-2 in DVD Forum tests
• Deliver low artifacts, but may lose texture detail
• Highest processing requirement for decoding
• A concern for software players, not CE devices
• BD spec requires at least 3 “slices” be used to ease
CABAC decode

H264 preserve texture and detail even at low bitrate. H264 can use low deadzone or adaptative threshold exactly like VC1.

At this time the best software decoding use less CPU for H264 than for VC1 (coreavc decoder outperform MS VC1 DMO decoder). VC1, MPEG2 or H264 decoding are not a problem with actual GPU generation.

My opinion, for 24p film content
• 480p: 1.5 Mbps
• 720p: 4 Mbps
• 1080p: 8 Mbps

For the same quantisation level (aka quality by pixel) there are a simple and empirical equation for that:

bitrate / ( H * W * Fps ) ^ N with N ~ 0.75

For the same source with the same average quantizer value:
• 480p: 1.5 Mbps
• 720p: 3.1 Mbps
• 1080p: 5.8 Mbps

And compression artefact for 1080p are less noticable simply because the relative size of artefact are less important for 1080p. 8x8 structure for all codec imply same absolute artefact size for 1080p or 480p. That's mean that equivalent subjective quality bitrate are even more close. For equivalent subjective quality IMO N ~ 0.6 is a good value for the previous empirical formulation.

Well now the bitrate equivalence for MPEG2 if my reference is 480p at 6 Mbps:
• 480p: 6.0 Mbps
• 720p: 12.5 Mbps
• 1080p: 23.0 Mbps

Nice things about encoding for download
• Don‟t have to worry about optical limits
• Can use longer GOPs for more efficient encoding
• Less risk of keyframe popping
• Average to Peak bitrate ratios often much higher
• Reduces segment reencoding needs

Optical limit for BluRay are really not a problem for ~ 8 Mbps "download encoding"
- GOP at 1 sec is not a really short gop. infinite GOP (keyframe only for scenecut) imply not a big improvement. It's easy to fight keyframe popping if you use really low ratio between intra-inter frames.
- Average bitrate at 8 Mbps with max at 40 Mbps with buffer at 30 Mbits is in practice unconstrained rate control encoding.

[MarcioAB]

Adding a BPP perspective:

Benwaggoner
480p: 1.5 Mbps = 0.15 BPP
720p: 4 Mbps = 0.18 BPP
1080p: 8 Mbps = 0.16 BPP
(hum...)

Mpeg4 (empirical equation)
480p: 1.5 Mbps = 0.15 BPP
720p: 3.1 Mbps = 0.14 BPP
1080p: 5.8 Mbps = 0.12 BPP
These are almost the BPP numbers I normally see.
I work with 0.01 BPP above those.

Mpeg2
480p: 6.0 Mbps = 0.61 BPP
720p: 12.5 Mbps = 0.57 BPP
1080p: 23.0 Mbps = 0.46 BPP
(obs: a well encoded Mpeg2)

[benwaggoner]

Originally Posted by MarcioAB

Adding a BPP perspective:

Benwaggoner
480p: 1.5 Mbps = 0.15 BPP
720p: 4 Mbps = 0.18 BPP
1080p: 8 Mbps = 0.16 BPP
(hum...)

Mpeg4 (empirical equation)
480p: 1.5 Mbps = 0.15 BPP
720p: 3.1 Mbps = 0.14 BPP
1080p: 5.8 Mbps = 0.12 BPP
These are almost the BPP numbers I normally see.
I work with 0.01 BPP above those.

Mpeg2
480p: 6.0 Mbps = 0.61 BPP
720p: 12.5 Mbps = 0.57 BPP
1080p: 23.0 Mbps = 0.46 BPP
(obs: a well encoded Mpeg2)

Okay, I can't claim I did a whole lot of emperical testing on those. The numbers I provided conflated a number of things.

1. Video and audio data rates
2. Assumptions of audience quality expectations for each tier
3. Content with film grain, where bpp doesn't seem to scale to 0.75 as closely as for cleaner content

[IgorC]

Sorry for offtopic.

benwaggoner

WMA Pro 10 outperforms AC-3 2-3x at low rates

Since when VC1 is optimizied for low rates to be used with audio multichannel tracks at low rates?

Is there any public test that shows it?
there is only one http://www.ebu.ch/CMSimages/en/tec_d...tcm6-53801.pdf

And there is no proof for such statement.

There is no 2x improvement even over MP3 at some usefull range of bitrate. Compare state of art HE-AAC 64 kbit/s vs LAME 128 kbit/s.

[benwaggoner]

Originally Posted by IgorC

Sorry for offtopic.

benwaggoner

Since when VC1 is optimizied for low rates to be used with audio multichannel tracks at low rates?

Is there any public test that shows it?
there is only one http://www.ebu.ch/CMSimages/en/tec_d...tcm6-53801.pdf

And there is no proof for such statement.

There is no 2x improvement even over MP3 at some usefull range of bitrate. Compare state of art HE-AAC 64 kbit/s vs LAME 128 kbit/s.

VBR WMA Pro 10 stereo @ 48 Kbps outperforms AC-3 stereo at CR 160 Kbps. Part of the practical advantage for WMA Pro for VOD is that we've got 2-pass VBR implementations, while AC-3 is almost always implemented as CBR. Rate-controlled VBR audio can be a big with with soundtracks, since they vary so much in complexity.

[Dark Shikari]

Originally Posted by benwaggoner

VBR WMA Pro 10 stereo @ 48 Kbps outperforms AC-3 stereo at CR 160 Kbps.

I find this highly doubtful, given that 160kbps AC-3 is basically transparent, while 64kbps WMA sounds atrocious.

Then again, this is yet another case of "ours is better than everyone else's despite all indicators to the contrary" syndrome, a marketing concept that has persisted for decades despite the fact that almost nobody falls for it. Of course, the mere fact that VC-1, VP7, and RV30/40 exist as marketable products shows that a fool and their money are indeed soon parted.

[benwaggoner]

Originally Posted by IgorC

Is there any public test that shows it?
there is only one http://www.ebu.ch/CMSimages/en/tec_d...tcm6-53801.pdf

Also, I note that only tested WMA 10 Pro at 192+ Kbps. That excludes the frequency synthesis modes used in the 32-96 Kbps range where the biggest efficiency improvements come.

But even still, WMA 10 Pro's 128 Kbps VBR with 5.1 can produce quite good quality with soundtracks, at a bitrate AC- struggles at with stereo and can't reach with 5.1.

[IgorC]

Originally Posted by benwaggoner

VBR WMA Pro 10 stereo @ 48 Kbps outperforms AC-3 stereo at CR 160 Kbps.

Really? Link to public test?

[benwaggoner]

Originally Posted by Dark Shikari

I find this highly doubtful, given that 160kbps AC-3 is basically transparent, while 64kbps WMA sounds atrocious.

That doesn't match my experience, but this is an emperical question. Can you share what source you're thinking of, and what settings you used?

Note I'm speaking of 48 Kbps 2-pass VBR WMA 10 Pro, and with soundtrack content. Both CBR encoding and test clips of more consistant complexity would be less advantageous to WMA 10 Pro.

[benwaggoner]

Originally Posted by IgorC

Really? Link to public test?

No, that's data coming internally, and my own anecdotal experience.

I would like to see a new hydrogen audio test for low bitrates soon. Their last round that included WMA 10 Pro wasn't a full apples to apples, as it compared CBR WMA 10 Pro to fixed-quality modes in the other codecs. I'd love to see the results of an independent double-blind test using either all CBR or all VBR for the advanced codecs. I think we'd do well, but it's always good to have community validation.

[IgorC]

Originally Posted by benwaggoner

No, that's data coming internally, and my own anecdotal experience.

Internally? hm....

Thank you for information.
Good bye.

[Dark Shikari]

Originally Posted by benwaggoner

That doesn't match my experience, but this is an emperical question. Can you share what source you're thinking of, and what settings you used?

OK, I just did a quick test. Since I don't have an AC-3 encoder handy, I compared AAC-HE 2pass 48kbps to WMA Pro 2pass 48kbps in a blind test using Winamp (put both in playlist, randomely skip a few dozen times so I don't know which is which, and then listen to the two in order).

WMA sounded atrocious. It was literally painful to listen to. Nero AAC sounded just fine, very few noticeable issues with the sound.. LAME MP3 128kbps sounded just fine too, and I suspect that is similar to 160kbps AC-3.

Source was "Jupiter", from Gustav Holst's "The Planets."

[IgorC]

I open this topic in HA forum. Let's see what audio professionals say.
http://www.hydrogenaudio.org/forums/...howtopic=62889

[Golgot13]

@ Dark Shikari, @ IgorC and at all,


After many request, from this forum and thread, they changed many information on MS VC1 website
like "VC1 is much better than MPEG2/H264" (and they updated last year, 2007, with new from 2004-2005)....
Don't forget MS prefer to say like "internally test" because "external re-test" could not prouve it...


If MS or Ben want to prouve officially someting about VC1 (best codec for... make presentation or may be make bretzel )
you can participate at MSU annual comparison codec (MS don't want since comparison between MS Photo and JPEG2000 ?).


Ben, what it's the low bitrate for VC1 ("best to preserve detail texture....) for you.
Can you explain the internally test to compare H264 (from Mainconcept you said) and VC1?
- VC1 file made exactly with VC1 core encoder only or core encoder + video preprocessing
- H264 file made only with H264 encoder (no preprocessing ?)
- Preset of H264
...

[benwaggoner]

[QUOTE=Dark Shikari;1128880]OK, I just did a quick test. Since I don't have an AC-3 encoder handy, I compared AAC-HE 2pass 48kbps to WMA Pro 2pass 48kbps in a blind test using Winamp (put both in playlist, randomely skip a few dozen times so I don't know which is which, and then listen to the two in order).[QUOTE]
CBR or VBR?

WMA sounded atrocious. It was literally painful to listen to. Nero AAC sounded just fine, very few noticeable issues with the sound.. LAME MP3 128kbps sounded just fine too, and I suspect that is similar to 160kbps AC-3.

Atrocious? I wonder if something else went on there. Did you use WME for rate conversion perhaps?

Also, note I was talking about movie soundtrack encoding above, not orchestral music. Orchestral music, which is more consistent than a soundtrack, would have less advantage for WMA Pro's VBR mode versus CBR codecs.