Chinese AVS codec, rivaling H.264 is rising

[Piper]

This just posted on digg.com

Chinese AVS codec, rivaling H.264, is rising
http://www.digitaltvdesignline.com/news/180205429;jsessionid=PTIFVSDOCM3NOQSNDBECKH0CJUMEKJVN

[IgorC]

are there any specifications/white papers?

[Sirber]

I read about it long ago IIRC, vaporware.

[SpaceV]

They got On2 VP6 source code about 2 years ago for the EVD project. Then they dumped On2 and now have their "own" AVS codec.

Hmmm, seems a bit fishy to me. I doubt we will see a white paper. I cant believe they didnt violate any patents.

Btw, whats with AT&T claiming MPEG patents, what a mess!!!

[bond]

first of all i believe it when i see the specs or a working implementation of that standard

than i am sure the chinese reverse engineer as much as they can. they do this is in many fields
that said that avs codec, if ever released, will propably be a clone of the mpeg standards
maybe they also ripped off on2, but who knows

about the patent status, patents are only valid in the countries where you have registered the patent. if no h.264 patent has been registered in china, noone in china has to pay for using it

[yanyani989]

not sure if these help

a presentation in ITU-T VICA Workshop on 22 Jul 2005:
http://www.itu.int/ITU-T/worksem/vica/docs/presentations/S3_P4_Gao.pdf

some sample clips:
http://www.avs.org.cn/fruits/

left side from the top:
1. WMP plugin for AVS playback (?)
2. Distributed Encoding Tools for AVS
3. MediaViewer (?)
4. speak for itself
5. SDK sample
6. ?
7. Sample Clips Copyright Claim
below are 9 AVS clips, each 150MB

edit:
forgot to say, some of the above DLs require ur info:
1. check the 'Agree' box on the left
2. fill in ur organization, name and email in the 3 boxes respectively
3. press continue to DL

[bond]

very interesting, thx!

is there any reference encoder available, useable on own streams?

[bond]

ok there is a gui encoder included which needs vobs as input
the first time i encoded it told me "mux: error" at the end
the second time it didnt but i cant play the resulting file with the avs dshow filters, meaning the source filter doesnt want to connect to the decoder

so somehow the whole thing doesnt really seem to work :>

[yanyani989]

i can only play the sample clips with WMP (10 or 6.4) but not MPC
the intro page (http://www.avs.org.cn/fruits/download/AVSPlayer_v1_Intro.asp )
point out that:
1. it only support up to 720*576
2. it's a plugin for WMP (the bottom say WMP 9.0+ anyway)

[gogoliu]

They open decoder's source under GPL now.
Homepage: http://cosoft.org.cn/projects/avsdec/
source code (2006-03-15): http://cosoft.org.cn/download.php?g_u_n=avsdec&f=avsdec_source.zip&file_id=7263
sample: http://cosoft.org.cn/download.php?g_u_n=avsdec&f=avs_test_stream.zip&file_id=7260

[Sirber]

What about quality?

[SeeMoreDigital]

Can anybody provide the direct-show filters?

[celtic_druid]

I compiled the above dll, doesn't appear to be a dshow filter though.

[SeeMoreDigital]

I compiled the above dll, doesn't appear to be a dshow filter though.Interesting...

How are users able to play the test samples in Windows Media Player without a direct-show decoder?

I'm confused :confused:

[celtic_druid]

Dll works ok with the test app. I don't know, I guess the samples pack linked to earlier contained a dshow decoder? Even with a dshow decoder filter, the .avs extension doesn't help.

[gogoliu]

This open source decoder was developed by beijing united source coding company (http://www.usc.com.cn), not AVS Workgroup.
And the version is Pre-Alpha.

[stefang]

Did anybody get that encoder to run?
If yes, it would be great if clips encoded with various options could be tested in MPlayer (see thread called "OpenAVS" for details) and if anything does not work, make the clip available and PM me, as I maintain the AVS decoder (not OpenAVS, but a new one in ffmpeg).
I don't have Windows, so can't try myself.

Regards
Stefan Gehrer

[mediator]

Their deblocking has 4 boundary strengths (like AVC), they have indexA and indexB (like AVC), they have Alpha and Beta (like AVC).... Hmmmmm....

I did not have a look on other parts, but I can hardly believe that this Codec does not infringe any patents that apply for AVC.

[videomixer9]

As I mentioned earlier already in another thread, it seems that the codec is a ripoff that just tries to take away customers from MPEG LA with cheaper licensing. Also note that recently China has seen a wave of patents that were plain technology copies from western technologies that were already patented, but not in China due to missing possibilities to patent the technologies that were just recently introduced. That's how they can brag about patents being registered mostly in China while mostly anything is stolen in real. Too bad this scam here is supported by some people. I'd rather have MPEG LA get the money as they at least spent it on development and research and not to alone spent it on playing mr. rich guy.

[diogen]

...China has seen a wave of patents that were plain technology copies from western technologies that were already patented, but not in China due to missing possibilities to patent the technologies that were just recently introduced...The advantages of being a country with 1 bln consumers. :p
They will probably be prevented from selling those abroad - but who needs it when you have enough of your own customers and some $20 less in licensing fees.
And if $100 Chinese players start playing Blu-Ray/HD-DVD, they will find their way to westerners.

I'd expect the negotiations (MPEG-LA and China) to be very hush-hush: no western company/organization wants to get into trouble with upsetting the Chinese government. Just look at "Don't do evil" Google.

Diogen.

[videomixer9]

What do I have from Chinese stealing our stuff producing it cheaper themselves, only unemployment here at our cost. And western companies are pretty pissed off doing buisness with chinese. The general opinion about China is deep down below, they only want their cheap work and their money, the stealing of technology pisses them plenty, and they'll only look away for the time they can still make profits without too much loses by cheap copies. But international pressure on China. Honestly I wish China the worst possible, their stuff is only cheap cause it's pirated, they enslave their workers and they disregard anything about ecology.

China is just a big fat country full of assholes without innovations that just copy anything and use the large market to pressure companies to get their stuff stolen for some profits, this will end pretty soon when Chinese stole the last things too and produce it as copies and noone will buy originals anymore. Not even talking about all the communist bullshit.

Copycats may rot in hell. The Chinese copy everything policy pisses me off enough, only bunch of losers need to copy. But in China they seem to not even see that when you kill off profits for innovating companies, you cannot copy them anymore and your all so smart "let others innovate and copy it for free" concept will not work anymore. Of course all the chinese companies that don't need to pay for research produce cheapass with their slaves working in factories that do not need to fulfill strict regulations as in the civilized world. Only thing they do is to kill everything the workforce in Europe and the US fought for by doing cheap work. But most Chinese I know are selfish assholes anyways and here at university are plenty of them. Their english sucks ass, they behave like arrogant asses in general.

[diogen]

Take it easy mixer, don't get so upset.
I think every country, no matter how developed _today_, went through a lawless phase like this.
...they only want their cheap work and their money...No need to say anything else.
It will overrule any other argument, moral or otherwise. You can't change the nature of capitalism.

Diogen.

[avih]

Guys, take it easy please, and moderate your responses.

[celtic_druid]

There is a market for original stuff though... People with money in China don't want to buy cheap stuff that was made in China.

Example:
If I had money and I was in China building a new kitchen, I would use the original Eurpean handles. If I was in Europe, I would use the cheap Chinese handles.

[SeeMoreDigital]

I hear you mate...

My kitchen came from Ikea (Sweden) even the door handles (or so I thought) which on closer inspection came from "The Peoples Republic of China"..... Together with some chopping boards, cutlery set, saucepans and under cabinet lighting. The tea towels were made in Bangladesh though ;)

[Nic]

VideoMixer9 striked for a rule 4.

Any country/race bashing from any other user will result in similar penalties.

Keep thread on topic or it will be closed.

-Nic

@Mods: IMHO VideoMixer9's thread should be edited to remove profanity/insults etc.

[conando]

rule 4? what's that mean?
- ah ok.. got it... but there's nothing about what kind of penalties you'd get from that... heavy spanking? :)

[SeeMoreDigital]

conando, I suggest you read all the forum rules again. Including the bit about "sanctions" after rule 16!

[Guest]

rule 4? what's that mean?
- ah ok.. got it... but there's nothing about what kind of penalties you'd get from that... heavy spanking? :) Struck for rule 11: "no use" post.

Direct such matters to a moderator via PM.

[stefang]

@videomixer9
Chinese stealing innovations from other countries should not bother you. Your words make clear that you are not the kind of person that will ever make an innovation, so nothing will be stolen from you.

But back to the AVS standard: It is a public standard and companies holding patents that are used in this standard are involved in licensing talks. That includes all non-Chinese companies, and looking at http://www.avs.org.cn/en/membership.asp you can see companies like Intel and Philips among others. Apart from that, patents involved in video coding are software/algorithm patents anyway, which have often shown to hinder innovation (like the JPEG-patent or the one-click-buy, see http://www.nosoftwarepatents.com for better examples). True, there is a lot of technology in AVS that also can be found in other standards (mainly AVC), but the same patent holders get rewarded for this, only by another licensing body. The claim that the MPEG LA would put the money back in research and innovation is not true. They only distribute the money back to the patent holders, what they do with it is up to them. Same goes for the AVS consortium.

[diogen]

...Your words make clear that you are not the kind of person that will ever make an innovation, so nothing will be stolen from you.Very mature... and innovative.

Diogen.

[*.mp4 guy]

I couldn't care less about the patents in this case, Chineses or otherwise. What is bothersome is having another standard that is basically a slightly less flexible clone of another standard; with an already widely used extension no less (Why they couldn't have picked anything else is beyond me). There isn't any use for another mpeg4 avc clone, we already have VC-1, and I'm not even sure if there is any point in it either. If VC-1/AVC?AVS isn't enough there is also VP7, which is, last time I checked better then everything but high end AVC encoders. IMO we would be better off with just AVC/VP7, ASP/VP6, MPEG2/VP3/Theora, there isn't any point in adding more video standards if they don't do enything better then one of the ones we already have.

[bond]

stefang, as the thread title is "rivaling h.264" and you seem to know avs, what features does avs have that are not in other formats, like mpeg-4 asp, avc or vc1? :)

[stefang]

Okay, as I stated before: many things are AVC-like, just trimmed down to be less computationally expensive, but I guess a few things are worth pointing out that seemed novel to me:

"Simplified dequantization": State-of-the-art encoders of other formats have the feature "trellis quantization". For each coefficient, they search for the optimum value in terms of rate-distortion, knowing how the decoder will dequantize it. If you already do that effort, what is the point of having a complicated dequantization on the decoder side? So in AVS, there are no matrices for dequantization. Each coefficient is scaled up the same way, which makes it easier for optimizations. All the cleverness has to be on the encoder side.

"2D VLC": the coefficients of one block are in the bitstream back to front, from the last (which is expected to be small, but with a high distance to the next) to the first (which is expected to be large, but close to the one before). Whenever a coefficient exceeds a certain size, the next Huffman/VLC table is taken. Something similar can be found in AVC CAVLC mode (i.e. when CABAC is off), but there it is algorithmically really complex, while in AVS it only takes a few lookup tables.

"Symmetrical B blocks": In AVS, no block of a B-frame has both forward and backward motion vector coded. To the usual forward/backward/direct block types that can be found in MPEG4 ASP and AVC, they add the type "symmetric". Imagine a group of pictures with two B-frames between each I/P frame (IBBPBBPBBPBB in display order). In the first of the two B-frames, the forward motion vectors (to the previous I/P frame) are often half the size and opposite direction as the backward motion vectors (to the I/P frame which will be displayed two frames later in time). So symmetrical means that the forward motion vector is predicted/coded as usual and then the backward motion vector is mv_bwd = -2*mv_fwd in the first B-frame and mv_bwd = -1/2*mv_fwd in the second B-frame.

"Geometrical median": Other standards use the arithmetic median to find a motion vector prediction from surrounding motion vectors. AVS uses the "geometrical median". You can read in http://www-ee.uta.edu/dip/Courses/ee5359/AVSslides.pdf on page 86 what that looks like. This paper is a good introduction to AVS anyway.

Regards Stefan Gehrer

[Manao]

Geometrical median prediction is a way to avoid the patented classical median prediction. I find the trick a nice workaround to avoid the patent, and the efficiency must be quite close. Yet, I wish that patent didn't exist in the first place...

Each coefficient is scaled up the same way, which makes it easier for optimizationsHardly. Scaling by a matrix requires only an additionnal multiplicationn. The overall slowdown is meaningless ( especially in comparison to, lets say, deblocking ).

I also fail to see how trellis can replace custom matrices. Trellis is all about dumb and brutal PSNR optimization. Custom matrices is all about psychovisual impression ( they generally lower the PSNR ).

[foxyshadis]

"Symmetrical B blocks": In AVS, no block of a B-frame has both forward and backward motion vector coded. To the usual forward/backward/direct block types that can be found in MPEG4 ASP and AVC, they add the type "symmetric". Imagine a group of pictures with two B-frames between each I/P frame (IBBPBBPBBPBB in display order). In the first of the two B-frames, the forward motion vectors (to the previous I/P frame) are often half the size and opposite direction as the backward motion vectors (to the I/P frame which will be displayed two frames later in time). So symmetrical means that the forward motion vector is predicted/coded as usual and then the backward motion vector is mv_bwd = -2*mv_fwd in the first B-frame and mv_bwd = -1/2*mv_fwd in the second B-frame.

Isn't this essentially the same as a temporal b-skip, except for requiring an MV?

[xyloy]

there is also VP7, which is, last time I checked better then everything but high end AVC encoders.
Yup, it's better than anything, except AVC High Profile wich is far far away(a long time ago.. oops, hum.. no) better for low quality encodes.

[Manao]

Isn't this essentially the same as a temporal b-skip, except for requiring an MV?No, in AVC, you either provide no vectors, in which case the vector(s) are guessed from the surrounding ( either spatial or temporal ). Or you provide two vectors. AVS provides only one vector, and infers the other one by the formula given by stefang. It's quite close in spirit to the direct mode in ASP.

[stefang]

Geometrical median prediction is a way to avoid the patented classical median prediction. I find the trick a nice workaround to avoid the patent, and the efficiency must be quite close. Yet, I wish that patent didn't exist in the first place...
I think it's not only for avoiding a patent. The arithmetic version does it component-wise, so predicted x is median of the surrounding three x components, and predicted y is median of the three ys. This has the disadvantage that you might choose x and y from different vectors, therefore constructing a new vector which might not have much to do with the motion actually happening. However, in the geometrical version you are always sure to take one of the candidate vectors completely. I believe that's an advantage.

[akupenguin]

"Simplified dequantization": State-of-the-art encoders of other formats have the feature "trellis quantization". For each coefficient, they search for the optimum value in terms of rate-distortion, knowing how the decoder will dequantize it. If you already do that effort, what is the point of having a complicated dequantization on the decoder side? So in AVS, there are no matrices for dequantization. Each coefficient is scaled up the same way, which makes it easier for optimizations. All the cleverness has to be on the encoder side.
Doesn't optimize anything. The reason H.264 uses dequantization matrices is that it allows IDCT to be non-normalized, and thus faster to compute with integer math. Once you're doing that, custom matrices cost no speed at all.

"Symmetrical B blocks": In AVS, no block of a B-frame has both forward and backward motion vector coded. To the usual forward/backward/direct block types that can be found in MPEG4 ASP and AVC, they add the type "symmetric". Imagine a group of pictures with two B-frames between each I/P frame (IBBPBBPBBPBB in display order). In the first of the two B-frames, the forward motion vectors (to the previous I/P frame) are often half the size and opposite direction as the backward motion vectors (to the I/P frame which will be displayed two frames later in time). So symmetrical means that the forward motion vector is predicted/coded as usual and then the backward motion vector is mv_bwd = -2*mv_fwd in the first B-frame and mv_bwd = -1/2*mv_fwd in the second B-frame.
Meh. Better would be to code one mv, predict the other using this method (maybe. Spatial direct works better than temporal direct most of the time.), and then code a mvd which is hopefully often zero. But that requires cabac...

However, in the geometrical version you are always sure to take one of the candidate vectors completely.
That's something which can be tested. Don't trust your intuition about which of those is a better predictor.