@Bond & (other who may be interested to know his answer ) - What do you think.

[iwod]

After some busy weekend and work on programming, ( Hopefully it won't be long till i start to code my competeting product to Surreal UI )
And i have time t read doom9 again.

I have been following this New AV codec section some of you may know. However i have NEVER EVER seen Bond reply to so many thread.
Especially when these Thread are either Vp6 or RV10.

Bond, from what i know by reading other part of Doom9, is a complete fans of Xvid. ( Although not as crazy as Marc FD :D ) So it is interesting to see him posting here about Vp6 and Rv10.

What i am interested to know is. What is your view on VP6 and RV10, Bond? ( Sounds like james bond :cool: )

Its quality, Vp6 vs Rv10 and Xvid......? ( Since quality is very subjective so it doesn't matter if he have done or not any test. I am only interested in his point of view and answers.)


P.S sorry speed was never in my list of questions, if Rv 11 could compress a AVI in 225Kb and look = 1200 Kb Divx Pro i wouldn't mind spending 20x longer than real time encoding speed to encode.

[Sirber]

hum... we'll be competitors huh? don't wanna join me instead? mouhahahahahahhaa :D

[Koepi]

iwod:

if you're a fan of a certain codec and thus don't use the other codecs, it's better not to write anything about the other codecs. I try to do my best not to write about other codecs than XviD.

Only that way we can have as much "freedom of choice" and relative "freedom of speech"(no, this forum isn't democratic or public domain) as we have now.

I hope you understand that.

Regards
Koepi

[bilu]

I think is just interested in feedback from using other codecs coming from an experienced MPEG-4 user. To some extent bond's opinion could be considered MPEG-4 biased, but the fact he's been showing continuous curiosity about the others codecs may be considered an impartiality factor :)

EDIT: IMHO Sagittaire and SeeMoreDigital could provide more experienced opinions. Bond's post I've seen in RV10 and VP6 are more related to licensing and to AAC than video. I may be wrong of course :)


Bilu

[bond]

what? me mpeg-4 biased? :D

seriously to answer iwods question (tough i am little bit surprised that you are asking especially me):

1) of course i tested/test all codecs i am interested in (which includes vp6 and rv9(+ehq)) as good as i can, i dont care about psnr (cause it once gave me other results than my eyes, which is enough to dont trust it anymore) so dont ask me about it
2) i dont have the time to make tests which are really representing all kinds of movies (dark/bright, easy/hard to compress aso)
i never do animes and i only do 1cd rips

3) my findings:
first of all the most important thing:
for the types of encodes i do i still have to see a codec that is clearly better than XviD, period
(combining this high quality with my opinion that open standards are more than _very important_ (which includes the ogg framewark too of course) some might call me mpeg-4/xvid biased if he/she likes)

to rv9 and vp6:
i didnt test rv10 but in rv9(+ehq) the good thing i found was the behaviour doom9 describes as "foreground high detail, background low detail", especially in the matrix sample (which i also use for my tests), where neo was using his mobile phone (white background) rv9 had as good as the same amount of details as xvid!!
the bad thing about rv9 was the really low quality in high action scenes, according to karl_lillevold (who i really respect very much :) ) he already tried to change this, tough i didnt test it till now (i will make a new round of testing when xvid 1.0 is out) - but anyways this high action low quality behaviour was enough for me to not use rv9
when it comes to low bitrate i cant really say much about rv9s behaviour, lotr2 is still waiting on my hd for xvid 1.0 to get encoded :D (but i saw parts of a pearl harbour encode by cruncher with xvid and i have to say xvid performed _very well_ even on low bitrates)

about vp6: normally i wouldnt have said it that way, but if you are asking me that directly about my opinion, i think its better than rv9 (once again i didnt do any representative tests!!!)
edit: hm, looking at the clips again i dont know anymore why i remembered vp6 being better than rv9, so lets say they are on par :D
i remember that the sampling features (or how they are called) hurted the quality more than they did any good, so i disabled them (maybe they are more for very low bitrate streaming or so)
in my tests it brought good details (also on high mo), maybe comparable to divx5, but again no chance against XviD


all-in-all i can only repeat the consens we have on doom9, which is "do your own tests"! quality is more than subjective
i recommand everyone for his/her comparison to use the great avscompare tool, which allows to zoom into the picture, switching between different zoomed clips aso (and this without showing which codec is displayed atm, avoiding "self-biasing" ;) )...

[Sirber]

good reply, I agree with you, making our own test is the best way to go. Since encoding is for yourself, use the codec you like more :D

I'm pretty happy myself with RV9/10, but I won't say: don't use XviD. But... I might say... don't use WMV9 lol :D

[iwod]

Um.......

I think Koepi....... you misunderstand me abit...... All i was trying to get is information from people who uses other codec and see what is their opinion on it. It doesn't matter if they hate it or not.
And it just happen i pick Bond. ;)

Example. Many of you may have been here longer than me. And you will all know that once Vp3,4,5 was never ever mention here. They only said to be "bad" and other things.

Real wasn't any good either. People were more of Divx 4 and Xvid those days.

I came here one year before MacFD pretty left the community to play PS2 :D . I was a newbie and didn't know much ( all though i still don't know enough today ;) ) What is Avisyth or filter .. etc... I know NONE. However a fact i know is that there are a small amount of Clips on internet are using Real. They are small and good.
At the time MarcFD was pretty much my mentor.... taught me many many things. I PM him pretty much about every question i had. ( I know it must be annoying... )
He, as he say was a total, uncondition Xvid fans. He said not only does Xvid quality counts, it is because it is open standard. ( Yes Bond.... what you said remind me of him )

RV wasn't popular as it was today. Thanks to many people pushing it. ( i can only remember Sirber quoting "300Kbps in Rv9 give you perfect quality :D ) And the arrival of Karl, which helps to solve problems and improve so many things.

Vp6 has been getting huge momentum lately with a personal editon of codec. ( 3 years late in arrival... )

The only reason i ask is to see how other people react to Rv now and what is their new and up to date view on it. 1 or 2 years ago i would expect a answer of Real is crap, spyware... etc....

But now people changed their view on it. That is why on my first post i worte it doesn't really matter if you have test it or not. Is his view that counts...........

By the way, i am not a fans of Real. I am just enjoying Rv9 / 10 giving the quality i want. I just like to analysis things a bit and see what is going on. Thinking back of it now this post is pretty pointless, it is much more like my blub.... ;)

But feel free to join in and chat though ;)

[Sagittaire]

For me ...

The VP6 and the RV10 crush everyone for the difficult encoding. The best mathematic test is SSIM because it's a psycho-visual test: VP6 are the best (and with PSNR, Overall PSNR) for this test for all bitrate ...
http://www.cns.nyu.edu/~zwang/

For HDTV resolution RV10 requires too many resources but the VP6 with its adjustable PP is fantastic ...

[Manao]

SSIM is not a psycho visual test. It's just a way of computing distorsion with taking a more parameters in count. Instead of just taking the mean square error, it also computes the covariance.

It doesn't take into account psycho visual heureustics in order to compute the distorsion.

Neither does it take into account the motion.

And most important, for me, it is still too kind with smoothing. It will penalize more a sharp picture with a slight block than an overall unblocky but smoothed picture. Since I don't like bluriness, I don't agree with SSIM results.

[Doom9]

If you want quality metrics, the JND thingie from Sarnoff would be the way to go. But even that, with all the well paid programmer and scientific (or not) research behind it, it still rated bad codecs like VP3 way too good so I'll keep on relying on my eyes only. But hey.. it's like with beers.. everybody has his favorite.

[DevilsChild]

Someone should compare the codecs with Video PQoS:
http://www.genista.com/videopqos.htm

[karl_lillevold]

bond and iwod: thanks for your thoughtful replies and kind words.

Re various quality metrics They all have their quirks. Especially JND has some weird preferences that do not match my visual preference, and it is too slow and cumbersome, but it may fit someone else. I have not tried SSIM, but will do.

Re the problem of constant quality with 2-pass VBR in RV9/10:

Generally I try to adhere to the same principle as Koepi, that is not commenting on codecs I do not frequently use, but at least in this case, I have used recent XviD builds more than any other codec besides RV, and I know it does an outstanding job in keeping the quality constant and in preserving detail. In fact, I repeated a well known comparison recently discussed much in this forum, and I am very impressed. Congratulations to everyone who contributed with the XviD development and improvements! And thanks to doom9 for his detailed work!

These comparisons and XviD performance, along with the feedback here served as enough motivation that I have worked on improving RV10 and its 2-pass VBR mode this weekend. The main problem I am facing is how so much of our code inherits from and still has to support various forms of CBR streaming, not to forget the complex SureStream handling. Compared to the XviD 2-pass algorithm, which is such an elegant way to achieve constant quality, ours is quite uhhm different.

Still, for my 8 minute test clip (low + high action), I have much improved visual quality during the high action. This translates to 0.4 dB improved Overall PSNR and +0.15 dB for Average PSNR, and 4% lower bitrate. For the full length clip, I have very solid constant quality, and significantly improved during the most famous scene. However, still a few MB over target, so I am not sure when these changes will be ready to test, but I will provide details when they are. Mainly I have completely redone the adaptive B frames algorithm, such that it now varies from 0 to 3 instead of 1 to 3, and evened out the B frame QPs when the reference is hard to compress. Many B frames were previously too blurry during high action. I have also approximated constant base level QP for the whole clip. The previous improvement I did helped more for even higher bitrate, shorter clips than this test, since it still operated with a 10 minute window length for its calculations.

[iwod]

@Karl

I don't understand how this thing limit the way of RV could be improved on. Couldn't one just make a Mode ( Something like EHQ )that will take all forms of CBR limitation out of the questions and only focus on VBR?
Or will that break the bit stream of RM. ( Sorry what is sure stream? )

[Koepi]

iwod:

depending on the rate control impementation it might not be possible to take such things out for a special mode except for rewriting and reconcepting it from scratch, so it can be a major job which can introduce buggy behaviour. This isn't necessarily a good target to aim at ;)

Surestream is a realmedia technology to have multiple bitrate streams combined to a single stream/file which at playback time can get switched depending of the internet connection quality(you know, bandwidth variation / network congestion due to server load and such stuff).

Regards
Koepi

[iwod]

Is that what is Surestream? I read about it long time ago..... I like the idea of it for streaming purpose....... Gosh it is there stopping all the Rv VBR quality.........:devil:

[karl_lillevold]

Thanks, Koepi. You explained it better than I would have myself.

SureStream is there stopping all the Rv VBR quality
Not stopping it, just making it more complicated. Some nice improvements on the way, just not a complete re-write...
There is nothing stopping a re-write either, except the time it would take. Maybe it will happen one day, but with only two video codec engineers, myself included, we have to pick our projects carefully. Unfortunately there are many other duties we have to tend to than encoder side improvements, that these turn into somewhat of a side-activity, even though that's the really interesting part.

[iwod]

I am to sorry to say.......... But why such big company like Real which reply heavily on their video and audio codec for their business only have 2 people on Video Codec......

Because from what i read it is just like there is a lot of idea in your ( Karl ) head that could improve Rv but lacking the resources.....

Like the EHQ 100, better pp......... etc..........

[iwod]

Originally posted by Sirber
hum... we'll be competitors huh? don't wanna join me instead? mouhahahahahahhaa :D

Well....... decided I didn't have enough time to jump to the delphi ...........

My best is still on Java. However i belive there will be nearly NONE who want to use a Java program........ ( Actually i am not sure of that, may be i could do a poll later.... coz people generally don't want to have a C# program, simply because they have to install .Net framework..........same goes to Java...... however i should expect everyone have java in their computer ^_^ )

[Sirber]

Java is cool, I use it to play Go online :D

[MfA]

Originally posted by Manao
SSIM is not a psycho visual test. It's just a way of computing distorsion with taking a more parameters in count. Instead of just taking the mean square error, it also computes the covariance.

It doesn't take into account psycho visual heureustics in order to compute the distorsion.

The difference with most other measures is that it isnt derived systematically from some HVS model. Given the amount of simplifications and assumptions underlying something like Watson's work I dont see that as a big problem though. IMO all local frequency space based HVS models are confusing cause with effect, we are edge sensitive not frequency sensitive ... researchers are lured by the correlation, and the ability to use measurements of overly simplified stimuli to train the model.

One of the underlying assumption of SSIM is that covariance is perceptually relevant. Performance wise it seems as good or better an assumption than seperability of perceptual error into spatial channels.

I expect the multiscale SSIM will be more adapt at detecting blurring, and taking motion into account would also be nice ... but all those things require automated training using the VQEG (and other such) sources. That is just a little too much work.