Getting Theora competitive

[Kurtnoise]

Could somebody please upload a copy of the ffmpeg2theora build with avs support for me. Kurtnoises site seems to be down unfortunately, so I can't get that.
http://kurtnoise.free.fr/misc/

Your version with avs support is very useful. It is also supposed to support subtitle since it has kate built in. However, it does not support subtitles.
use either my last build from this month (20091113) or oggz-merge from the oggztools package + kateenc. Same link as above...

[donaldtone]

Thanks Kurtnoise!
The last ffmpeg2theora (20091113) works very well with avs script and kate subtitle encode.

There is just a concern, the kate lib version built in is just 0.3.1.

Would it be possible to change to kate lib 0.3.6(last official version)? Thanks.

[donaldtone]

BTW: I could not find kateenc binary for windows anywhere.
Could anybody do me a favor?

[Kurtnoise]

There is just a concern, the kate lib version built in is just 0.3.1.

Would it be possible to change to kate lib 0.3.6(last official version)?
build from today updated to the libkate 0.3.6. Anyway, you won't see any differences.

btw, I don't know why you're so interesting by the srt support in libkate. The rendering is poor to me (italic not supported by example)...


BTW: I could not find kateenc binary for windows anywhere.
Could anybody do me a favor?
http://kurtnoise.free.fr/misc/libkate-0.3.6.zip

[donaldtone]

Thank you very much.

Agree with you, the kate can not support <i>, <b> tags right now.

But Kate is in its early age, let's see its progress.

And Kate is the only subtitle format in new OGV format. I am not sure OGV support srt or not even though ogm supports.

[donaldtone]

new edition of kate has just released.
It fixes a lot of issues and add support of SRT style(that means italic, bold style can be rendered correctly.).

0.3.7 - 15 november 2009
new encoding functions to specify timing in granule rate units
fix SRT timing precision problems
new -M option to kateenc to allow HTML-like markup in SRT files
allow the last event in an SRT file to not be followed by an empty line

[iwod]

No news at all? Or have they given up theora and just move to X264

[Rasi]

No news at all? Or have they given up theora and just move to X264

? look one post above yours?

[Astrophizz]

That's just for Kate, which is used for subtitles. That wasn't a post about theora.

[iwod]

That is nearly 4 months without update.

[tuqueque]

Actually, since yesterday, a new development branch was created for the eventual 1.2 release... Check it out https://trac.xiph.org/timeline

And there is already some nice improvements.

[ricardo.santos]

ffmpeg2theora 0.26 released

[nurbs]

I did a small test (http://www.megaupload.com/?d=GLRFRB7J) of ffmpeg2theora 0.26 and x264 r1400. I only two sources and the default settings on both encoders, with x264 encoded at the same and half the bitrate the Theora encoder used. Exact settings, links to source files, PSNR, SSIM and comparison pics can be found in the download. I can upload encoded samples should anyone be interested.

Theora still needs a lot of work. Even at half the bitrate x264 produces clearly better results.
I'd be interested why the background in the score area of the Touhou video is so blurry in Theora. Since it is completely static it shouldn't be so low quality.

[Dark Shikari]

I did a small test (http://www.megaupload.com/?d=GLRFRB7J) of ffmpeg2theora 0.26 and x264 r1400. I only two sources and the default settings on both encoders, with x264 encoded at the same and half the bitrate the Theora encoder used. Exact settings, links to source files, PSNR, SSIM and comparison pics can be found in the download. I can upload encoded samples should anyone be interested.

Theora still needs a lot of work. Even at half the bitrate x264 produces clearly better results.
I'd be interested why the background in the score area of the Touhou video is so blurry in Theora. Since it is completely static it shouldn't be so low quality.MB-tree is the reason x264 performs so well on the score area: without an algorithm that performs a similar purpose, it's impossible to compete on that particular area of the frame (except by simply cheating and setting a really high ipratio). This isn't really something to criticize Theora heavily on, since x264 is the only encoder with MB-tree ;)

By the way, you should test the latest Theora SVN; they added adaptive lambda (rough equivalent of x264's AQ), which should help quality in flat areas a lot.

[nurbs]

I will test it if anyone compiles it for me. AQ is important so it will be interesting to see the results when it's released.

[ricardo.santos]

@nurbs
You can try the ffmpeg2theora nightly builds
http://firefogg.org/nightly/

last update was today so im guessing they're compiled using latest build

[hellfred]

@nurbs
You can try the ffmpeg2theora nightly builds
http://firefogg.org/nightly/

last update was today so im guessing they're compiled using latest build
The build is based on Theora 1.1 Thusnelde.D:\Manga\Maian>ffmpeg2theora.exe
ffmpeg2theora 0.26+svn16879M - Xiph.Org libtheora 1.1 20090822 (Thusnelda)
The latest branch should annouce itself as "ptalarbvorm". Maybe ask at the theora mailing list for a build.

Yours,
Hellfred

[Astrophizz]

Good lord, where did they get that name?

[mariush]

ogg is the container
vorbis is the audio codec
theora is the video codec
thusnelda is the release name
Stephen R. Savage is the .....

[Midzuki]

Astrophizz wrote:

Good lord, where did they get that name?

I have no idea. :) Anyway, for the curious, here goes the URL to the "offending codename": :p

http://svn.xiph.org/experimental/derf/theora-ptalarbvorm/

[ricardo.santos]

so the ffmpeg2theora nightly builds arent actually based/compiled using the latest build from the SVN?

[valgor]

Well, I built rev.16883 theora-palarbvorm.
The same video source, the same settings (v -6).
Thusnelda 1.1.1 produced 673 Kbps video, Palarbvorm - 704.
Visually, Palarbvorm produced a bit sharper video, but the difference is too small, sometimes no diffs.

[nurbs]

Could you upload the build somewhere, please? :thanks:

[valgor]

I'm on Slackware Linux, so no binary for Windows.

Anyway, every one can assess visually:

Xiph.Org libtheora 1.1 20090822 (Thusnelda)
Xiph.Org libtheora 1.1+ 20100126 (Ptalarbvorm)

I used better source than previous post.
100 first seconds of Elephants_Dream_HD.avi (http://www.elephantsdream.org/download/), resized to 640x360:

mplayer /home/video/Elephants_Dream_HD.avi -vf scale=640:360 -endpos 100 -nosound -vo yuv4mpeg:file=/tmp/ed640.yuv


both encoded with encoder_example:

./encoder_example -v 6 /tmp/ed640.yuv > ed640.ogv


both video in single tar-archive, Ptalarbvorm_vs_Thusnelda.tar, 18 MB
http://shareflare.net/download/8268.83264053348a5744538bb6b9f/Ptalarbvorm_vs_Thusnelda.tar.html

[hellfred]

Did you really build rev. rev. 16883 (https://trac.xiph.org/changeset/16883) with the latest bug fixes from 02.07.2010, as the date in your naming schema Xiph.Org libtheora 1.1+ 20100126 (Ptalarbvorm) somewaht contradicts the revision number. The date in the later one implies a build without the bug fixes.

[valgor]

Sure, this define hardcoded in lib/internal.h, line 74-75
http://svn.xiph.org/experimental/derf/theora-ptalarbvorm/lib/internal.h


/*This library's version.*/
# define OC_VENDOR_STRING "Xiph.Org libtheora 1.1+ 20100126 (Ptalarbvorm)"

[Midzuki]

OK, I've just found a Windows binary for "ptalarbvorm" :) :

http://www.mediafire.com/?rwiwnnhjezt

(contains the encoder_example, libtheora, libvorbis and libogg).

[Stephen R. Savage]

http://img706.imageshack.us/img706/9547/lolfailt.th.png (http://img706.imageshack.us/i/lolfailt.png/)

Theora: keyint=300, speed=0

Xvid: max_keyint=300, max_bframes=6, vhqmode=1, cartoon, bvhq, bquant_ratio=100, qpel, gmc, nopacked, lumimasking

x264 1000k: keyint=300, minkeyint=30, bframe=6, badapt=2, bpyramid=normal, ref=8, deblock=-1:-1, aq=2:1.0, nombtree, partitions=all, direct=auto, me=umh, merange=24, subme=9, psyrd=1.0:0.0, trellis=0, no-fast-pskip

x264 500k: keyint=300, minkeyint=30, bframe=6, badapt=2, bpyramid=normal, ref=8, rclookahead=300, aq=2:1.0, partitions=all, direct=auto, me=umh, merange=24, subme=9, psyrd=1.0:0.0, trellis=0, no-fast-pskip,

Not too bad. Theora is competitive with Xvid, and at least 25% as good as x264. Some more frames here: http://www.sendspace.com/file/azews9

Edit: I noticed I was probably misrepresenting Xvid and low-bitrate x264, esp due to deblocking in the 500 kbps test and cartoon mode in Xvid. I have updated the screenshot to show that Theora does in fact not beat Xvid (though it still beats x264 @ 1/4 bitrate).

[Dark Shikari]

That's basically what AQ does for you. Notice how the sharp lines are far better in Xvid: that's because Xvid is wasting all its bits on the sharp lines.

[Stephen R. Savage]

The stranger part is that I used the Xvid version with your VAQ patch, but I still see the typical blurring of flat areas. Perhaps it is related to the H.263 matrix used in Xvid.

[Dark Shikari]

The stranger part is that I used the Xvid version with your VAQ patch, but I still see the typical blurring of flat areas. Perhaps it is related to the H.263 matrix used in Xvid.My VAQ patch for Xvid was never very good, nor properly tuned for Xvid's quantizer scale.

[Stephen R. Savage]

This is off-topic, but I have been wondering about a few things:

1) I read a quote from Manao that MPEG-4 ASP is only "20% more efficient than MPEG-2." Is this accurate? Another commonly cited figure is AVC having "twice the theoretical efficiency of MPEG-2". We all know that x264 is several times better than the last-generation competition, so is this statement still true? Could an optimized MPEG-2 encoder be 50% as effective as H.264 (thereby beating Theora in the screenshot above)?

2) How far are we from the theoretical best compression for any general 2-D image? Can we reasonably expect a 2x or similar increase in compression in the future?

[Dark Shikari]

This is off-topic, but I have been wondering about a few things:

1) I read a quote from Manao that MPEG-4 ASP is only "20% more efficient than MPEG-2." Is this accurate?ASP has a few advantages:

1) Better MV prediction
2) Better VLC tables, especially for low bitrates
3) Better quantization at very low bitrates (H.263 quant)
4) 4MV
5) Intra AC pred (0.1%, basically useless)
6) GMC (basically useless)
7) Qpel (<5% probably, potentially counterproductive)

It also has a few disadvantages, like the retarded B-skip rules (if a block is skipped in a P, it has to be skipped in the colocated block in the relevant B-frame(s)).

Overall, I'd say it could be a good bit above 20% at lower bitrates. Do remember the other big advantage: MPEG-4 ASP usually uses a far shorter keyframe interval, which is why Xvid could outcompress DVDs so well. This works because generally one can get exact iDCTs with MPEG-4 ASP, which is harder with MPEG-2, where the error can accumulate very fast.Another commonly cited figure is AVC having "twice the theoretical efficiency of MPEG-2".Do remember these figures depend heavily on bitrate. The lower the bitrate, the more the advantage, the higher the bitrate, the less the advantage. It's probably possible to construct a reasonable situation in which x264 beats MPEG-2 by a factor of 10, and it's also possible to construct one in which it barely wins by 30% or less.We all know that x264 is several times better than the last-generation competition, so is this statement still true? Could an optimized MPEG-2 encoder be 50% as effective as H.264 (thereby beating Theora in the screenshot above)?MPEG-2 is probably overall a bit inferior to Theora. It has no loop filter, and while it has B-frames and better MV prediction, it also has probably a significantly worse entropy coder. I don't think the gap is huge though. At a medium bitrate, I suspect an extraordinarily good MPEG-2 encoder (e.g. the same level of advancement as x264) could probably beat the current Theora encoder by a significant margin.2) How far are we from the theoretical best compression for any general 2-D image? Can we reasonably expect a 2x or similar increase in compression in the future?Especially at low bitrates, at least that is possible if you're willing to sacrifice performance. Some very promising ideas in that realm include Markov-chain intra coding and decoder-side motion vector derivation.

[LigH]

Information Theory (based on Claude Shannon) states that there is an entropy of information, a limit of of bitrate, which applies to any lossless compression.

For lossy compression instead, the limits are rather vague, and move from objective ratings ("How much loss is calculated?") to subjective ratings ("How annoying is the loss?").

It has already been discussed a lot for audio compression: Modern compression algorithms like AAC or Vorbis don't need a remarkable amount of objectively less loss than MP3 for the same bitrate; the loss just "sounds less noticable" than the one of most usual MP3 encoders. And LAME was able to keep track mainly by a very well tweaked psycho-acoustic model, not because it was able to get more out of the MP3 specs.

Something similar applies to AVC and VC-1, even to RM: Of course they use a more efficient encoding technology than MPEG2 and MPEG4-ASP; but even more important - especially for lower bitrates - are their technologies of suppressing noticable artifacts (e.g. "In-loop deblocking"): They do not only have "less loss", but especially also "less annoying loss".

And that is hopefully also the chance for Theora to grab its own piece of the big cake.

[dragsidious]

On a side note... just for the record.. etc
I did not see anybody mention it here and I expect all the comparisons done are in two-pass mode so it's probably not relevant, but here is a little note from a Theora developer on one of the details of comparing Theora vs H.264

from
http://www.streaminglearningcenter.com/articles/ogg-vs-h264---round-one.html?comment_id=377#comment_377

ffmpeg2theora* -V 500 -k 300 -d 99999*** (the only additions provided to the commandline above is -d 99999 to remove the buffering constraint).
Two-pass also disables the buffer constraint by default, but I wanted to make it clear the the improvement is from removing the buffering constraint, not from switching to two-pass.** If you're not live-streaming two-pass is highly recommended.* Many other encoders use a significant amount of lookahead in one-pass mode, while in one pass theora is a live-streaming-sutiable zero latency encoder.* The only way to do a apples to apples comparison is to either run both encoders as two-pass or both as zero latency,* and make sure to use the equivalent buffering constraints on each.

So I guess when your running ffmpegtheora by default it's designed to suit the needs of people using it for streaming. While the h.264 encoders are typically doing high quality. Making a somewhat unfair comparison.

Not that I think that this will make a serious improvement in Theora comparisons against H.264, but it may help out quite a bit if your trying to compare it against Mpeg-4 ASP or MPEG-2.

[benwaggoner]

Something similar applies to AVC and VC-1, even to RM: Of course they use a more efficient encoding technology than MPEG2 and MPEG4-ASP; but even more important - especially for lower bitrates - are their technologies of suppressing noticable artifacts (e.g. "In-loop deblocking"): They do not only have "less loss", but especially also "less annoying loss".

And that is hopefully also the chance for Theora to grab its own piece of the big cake.
Except is there anything that Theora can do that isn't applicable to other more efficient codecs?

Psych tuning is great, but for all the great work done with LAME, the most basic HE AACv2 implementation would mop up the floor with it below 64 Kbps.

Good compression is a combination of psych tuning WITH the available bitstream tools.

[LigH]

Unfortunately, some of the most efficient algorithms are patented (starting e.g. with the "Arithmetic Code" which has an advantage over the other integer entropy codes like Huffman; see CABAC vs. CAVLC), and therefore can't be used by projects which claim to stay patent-free.

It was a heap of fortune that the Vorbis algorithms turned out to be so efficient and convenient (if I understood them correctly, they are based on a kind of vector quantization). Like a music chart one-hit-wonder: The success of the second song release is not comparable... ;) - And so you can not expect Theora to have the same advantage over MPEG4 as Vorbis had over MP3. Video is not simply "like audio with an additional dimension". :D

[Dark Shikari]

Unfortunately, some of the most efficient algorithms are patented (starting e.g. with the "Arithmetic Code" which has an advantage over the other integer entropy codes like HuffmanNo it isn't. Range coding was invented in 1979. Some specific implementations designed to be faster are patented, but range coding as a whole is not.

[dapperdan]

For more on the patents on Arithmatic Coding see this post by one of the Dirac designers:

http://lwn.net/Articles/272973/

short version: the basics are too old to be patented, but many optimizations are. But many of these optimisation techniques are now so old as to be redundant. They seem to think their optimisations are pretty spiffy, based around making it easy to parallellise and were going to patent them, but then just published instead.

I guess a relevant question for this thread is to ask whether this can be transplanted into Theora in an advanced profile or new version and what kind of benefit that would provide.

[kosmonaut]

Looks like the situation just got more interesting.

Google to Open-source VP8 for HTML5 Video (http://newteevee.com/2010/04/12/google-to-open-source-vp8-for-html5-video)

[CruNcher]

Yep it's exciting to think about especially as it will be for sure directly sponsored by Google so Summer of Code this year could get really interesting if it's not just a rumor @ all, but it looks legit and in reality many expected it to happen :)
I guess everyone @ Xiph is currently Dancing on the Desks getting such a major update and support directly from Google in form of On2s Engineers vs MPEGs H.264, though most probably a lot visual tweaking has to be done before it could stand against H.264 (it's very well known how almost possessed ON2 was with the word PSNR) well especially side by side to X264. Also some Xiph guys maybe wouldn't need to imagine that the Visual quality is on par anymore, because it might be true by then who knows ;)
Beating Dark_Shikaris Psy work he done over the years on H.264 will be for sure a Adventure for On2(Google), Xiph and whoever joins by then :)

Also im pretty sure we gonna see alot of new faces appearing and the reason is simple who doesn't want to be as creative and work on a Video Technology that is supported by Google and especially a pretty new codebase that Theora didn't offered to begin with, it's crazy to think about having a Open Source free Patent alternative to MPEG with a pretty new codebase to work on for the future, new fresh ideas can emerge from that and not so limited like with theoras current base i hope :)

Think about it as Googles Project Mayo, though the main difference is no one stands above it and gives the direction to the people bellow, this is gonna create the direction actually it's up to Google how far they wanna go with this :)

Though MPEG is already far with the research on H.265 but im pretty confident that VP8 Open Sourced could drive up fast , though when you work on this you have to throw away a lot of things and completely free your mind and come up with new unpatented ideas :) that's a real research challenge in the Signal Processing field and especially Video (especially vs a such big entity of knowledge that MPEG stands for).

[iwod]

I am not getting it. Open Sourcing VP8 would means i could contributed to VP8 using patented software techniques and theoretically sue myself or Google for it just because i hate VP8 so much.

It would be better if RMVB opening up since they have much higher usage on the Internet.

[Mr VacBob]

Google's CLA (http://code.google.com/legal/individual-cla-v1.0.html) covers that.

[littleD]

Hmm what about decoding theora in hardware? Blog (http://code.entropywave.com/leonora/)
Results of theora playback (http://blog.mjg.im/2010/04/16/theora-on-n900.html)

[MfA]

I still have my doubts about how absolutely necessary b-frames are. I think video codecs are just lacking the tools to do pure forward coding right ... we need a forward only direct mode for forward coding with multiple references for instance (which extrapolates motion, like it now interpolates it for the bidirectional mode).

[LigH]

So far I believed that the "opening door" example is quite convincing... but you surely did more tests than me.

[kosmonaut]

Looks like Steve Jobs is threatening to let the other shoe drop: group organizing to enforce MPEG-LA patents against Ogg Theora (http://newteevee.com/2010/04/30/apple-may-be-gunning-for-open-source-codecs/).

If it was limited to Ogg Theora, and not the rumored royalty-free VP8 from Google, I doubt Jobs (or anybody else) would really care, so this has to be ultimately about VP8, imo. Surely Google did due diligence to determine if VP8 really could be royalty-free though?

[Dark Shikari]

Surely Google did due diligence to determine if VP8 really could be royalty-free though?Microsoft did that with VC-1 and look where that got them.Looks like Steve Jobs is threatening to let the other shoe drop: group organizing to enforce MPEG-LA patents against Ogg Theora (http://newteevee.com/2010/04/30/apple-may-be-gunning-for-open-source-codecs/).

If it was limited to Ogg Theora, and not the rumored royalty-free VP8 from Google, I doubt Jobs (or anybody else) would really care, so this has to be ultimately about VP8, imo. I'm pretty sure Steve Jobs is just trolling freetards. Remember, this was just sent as a short response to an email, not a big official statement. Not that there's anything wrong with a good troll, and Steve Jobs is a veritable master of it.

In short:
Odds Steve has any knowledge about such a plot: low.
Odds such a plot exists: medium.
Odds everyone is just getting trolled: high.

[kosmonaut]

I'm pretty sure Steve Jobs is just trolling freetards.

I suspect whether it's a trolling or not, he is going to provoke a big reaction out of it, that's for sure.

Also, while it was just one email, I don't think Jobs does anything (in public) without thinking it through.

[smok3]

Maybe Jobs should troll on that, practical example:

Running chromium on karmic (with non-free codec set installed) will return me theora version of the video in most cases, even where h.264 version is available, so Jobs has my support on this one (just this one). :devil:

edit: http://www.apple.com/hotnews/thoughts-on-flash/