new wavelet codec

[easyfab]

here is a new wavelet codec :
http://rududu.ifrance.com/rududu/indexen.htm

What do you think out it ?

[celtic_druid]

I don't think this is new as I have had it installed for some time... or is there a new version??

[Sigmatador]

i've done a quick test. the result: using the parameters for best quality, the video looks worst than a divx 3.11 cbr video, and the filesize is bigger ! using parameters for the same filesize than the divx file, it's so ugly !
but it's not a final release ;) so continue working :cool:
one good thing: the wavelet compression is very good for the compression of dark scene and with a low bitrate (p-frames are very big, i-frames are very small :sly: ). the dark scene is not very beautiful with divx (all divx&xvid) so... to be continued :D (wavelet compression could be a good alternativ for dark scene coding)

one question plz: have you used the ogg tarkin source code ?
i've tested the tarkin codec, the video wasn't ugly like rududu but the filesize was bigger (and no directshow filter so just for test ;) ) good for dark scene coding too (just tested with high bitrate).

[Sigmatador]

erf... no reply :confused: (@rududu je c ke ta pa le net mai fai un effort ;p)
and what about helping the developpement of XViD (wavelet in xvid?¿? :D )

[Tommy Carrot]

Last month a new version came out.

It has no such artifacts than the mpeg derivants (blockiness, mosquito noise, etc). Qualitywise it's a bit better than vp3, gives more details. Ofcourse nowhere near to xvid/divx, but it's progressing.

It has no practical use yet, because of the limitations (resolution have to be dividable with 64), and serious bugs (altering colors, vibrating picture).

But it's nice to see a (surprisely well) working wavelet codec. If they developing it more, and fix the bugs, it will be very promising.

(sorry for my english :))

[ChristianHJW]

Interesting !! No time to test now, maybe you can answer :

- how man frames will it load into buffer and perform wavelet transformation on ? Only one or more ?

- Is one complete frame compressed into one set of coefficients or does it break down the frame into lines or macroblocks ?

Thanks for clarifying

[Tommy Carrot]

Originally posted by ChristianHJW
Interesting !! No time to test now, maybe you can answer :

- how man frames will it load into buffer and perform wavelet transformation on ? Only one or more ?

- Is one complete frame compressed into one set of coefficients or does it break down the frame into lines or macroblocks ?

Thanks for clarifying

Obviously i could not look into the sources, so these are just guesses, but it uses some kind of motion estimation (p frames), not 3d wavelet. I doesn't buffer the frames.

From the "resolution have to be multiples of 64", there has to be macroblocks, but they cannot be seen.

Watch out! There are 2 version, and the link above points to the older one. The newer is a lot better, i'm speaking/writing about that.

[ProfDrMorph]

IIRC Wavelet compression divides the picture into something like 4 subpictures ( each being 1/4 of the size of the original ) and one of those subpictures is divided into 4 subpictures again and so. Since 64 is a power ( I hope that's the correct english word ;) | I mean the general term for square like 2²=4 ) of 2 ( 2^6 = 64 ) this may indicate the number of "subpicture creation steps" the codec uses. The reduced size of the subpictue is the result of reducing each the height and width of the picture to it's half.

My english is rather poor when it comes to explaining "scientific" or "technical" stuff ( because we don't learn about this in school ). I hope I made myself clear even though I had problems expressing what I meant. ;)

[rududu author]

Hello all !

"But it's nice to see a (surprisely well) working wavelet codec. If they developing it more, and fix the bugs, it will be very promising."

Thank you, but I'm alone :)
The next version will be better !

"Interesting !! No time to test now, maybe you can answer :

- how man frames will it load into buffer and perform wavelet transformation on ? Only one or more ?

- Is one complete frame compressed into one set of coefficients or does it break down the frame into lines or macroblocks ?

Thanks for clarifying"

I don't use B frame if it's your question, and I probably will never. The wavelet transform is performed after motion compensation (overlapped block motion compensation) on the residual coeffs.
I don't use the same blocks as MPEG, there is no macroblock, the only blocks I use are the motion estimation/compensation blocks. The wavelet transform is performed on the whole frame.


" IIRC Wavelet compression divides the picture into something like 4 subpictures ( each being 1/4 of the size of the original ) and one of those subpictures is divided into 4 subpictures again and so. Since 64 is a power ( I hope that's the correct english word | I mean the general term for square like 2²=4 ) of 2 ( 2^6 = 64 ) this may indicate the number of "subpicture creation steps" the codec uses. The reduced size of the subpictue is the result of reducing each the height and width of the picture to it's half."

That's true, but it's not the reason for the multiple of 64 limitation, my wavelet transform could be performed on odd image dimensions. The limitation is due to some motion estimation/compensation blocks, I will remove this, but I'm trying other wavelet schemes for the moment.

Thank you all for your support.

Nicolas

[Tommy Carrot]

xiph.org (the creators of ogg vorbis) is working on a wavelet codec, called Tarkin. They are experimenting with many kind of wavelet, and there are adept developers, i'm sure they could give you good advices.

[Neo Neko]

Originally posted by Tommy Carrot
xiph.org (the creators of ogg vorbis) is working on a wavelet codec, called Tarkin. They are experimenting with many kind of wavelet, and there are adept developers, i'm sure they could give you good advices.

I am aware of Tarkin. But AFAIK there has only been some preliminary work done on it and some code donated to it. None of which dealt with wavelets IIRC. Due to school etc. I have not had contact with Emmett or his crew for a while now. Longer than I would have liked. I have seen no mention on the mailing list last time I checked of a decision to go with wavelets.

[Tommy Carrot]

Originally posted by Neo Neko
I am aware of Tarkin. But AFAIK there has only been some preliminary work done on it and some code donated to it. None of which dealt with wavelets IIRC. Due to school etc. I have not had contact with Emmett or his crew for a while now. Longer than I would have liked. I have seen no mention on the mailing list last time I checked of a decision to go with wavelets.

Yes, they more or less abandoned tarkin, because of the higher attention to theora (which i think is not a too good move, because theora is based on VP3, which is far from competitive), but there still are knowledgeful people, who could help IMO.

[sam_b]

Xiph.org's page pretty much says they've barely got a monkey with a typewriter on tarkin at the mo, as they're concentrating on theora. I would very much like to see tarkin progress, but rududu's codec seems very promising. I think I could learn to live without DCT artifacts :)

It seems Xiph.org want to get a stream format out the door for theora, and tuning will be secondary.

[TheXung]

Personally, I really admire the people at xiph, but sometimes I look at their mission statement and I think they're crazy. To build media codecs without infringing on any current patents requires doing years of research and that time frame is far longer than the developement period of codecs that reuse older technology. I think we're at the point where xiph is trying to prove that it is more than just vorbis by jumping into the game with an "old" codec given to them. And I fear they may come up with a lesser reputation than what they could have been if they had concentrated on really fine tuning vorbis and being known for dominant quality audio. You can only ride the shoulders of one gifted programmer to such an extent.

[sam_b]

I suspect they are desperately trying to get a foot in the RV/WMV door before these two shut it. Probably for streaming. Once they have some Ogg infrastructure in place then they will probably go back and pay more attention to things such as tarkin. People are less likely to write complete Ogg stuff if there is no video to go with the well-integrated vorbis. I know what you mean about them not wanting to get bogged down with vorbis, but I don't have a good feeling about theora. Yet.

[gldblade]

Just wondering: Does wavelet technology (word choice?) have any patents yet?

[Tommy Carrot]

Originally posted by gldblade
Just wondering: Does wavelet technology (word choice?) have any patents yet?

The ME/MC is certainly. Almost everything is patented, which is related to video compression.

EDIT: The wavelet tech itself is iirc old enough to be patent-free, but if you want to get decent results with it, you have to use patented algorithms.

[rjamorim]

Originally posted by gldblade
Just wondering: Does wavelet technology (word choice?) have any patents yet?

Yes. Lots and lots of patents.

That's the main reason why jpeg2000 is so heavily patented.

Try a google search "wavelets patents"

Regards;

Roberto.

[Neo Neko]

Originally posted by Tommy Carrot
Yes, they more or less abandoned tarkin, because of the higher attention to theora (which i think is not a too good move, because theora is based on VP3, which is far from competitive), but there still are knowledgeful people, who could help IMO.

1. Tarkin is not abandoned. It is still in the development stages. Even if full attention were given to tarken now xiph would quote you at least a 2 year time to completion.

2. More attention is being given to Theora or VP3-XIPH because VP3 is a proven/working codec. Although a bit slow. Xiph to this point has been concerned solely with audio. Theora is their first major chance to get their feet wet in the video pool. They will take some of the technologies they got from VP3 and the knowledge they gain from tweakin it and apply that to Tarkin. This furthers the belief that Tarkin will not be wavelet based.

It all makes good sense. Especially in the realm of competition. Either Xiph could stumble along for several years to develop their first codec. While in the mean time everyone else releases one or more new versions of their codec. Or they take an existing codec like VP3 and use it as a springboard to their future cutting the development time drastically and giving them many lessons without them having to learn the hard way. ;)

[Tommy Carrot]

I think the problem of tarkin was that it became clear that 3d wavelet is not the way to go, and (beyond the patent problem) there isn't a known and working ME algorithm for wavelets. All of the wavelet codecs i saw had very bad artifacts, smoothed the image, and the texture was somewhat flowing (sorry, i cannot express it better). I really lost my faith in the wavelet codecs.

Rududu's codec is a pleasant surprise, because there are almost no artifacts (except some bugs, but i think those are fixable), and the quality is quite good. And i hope it can be even better.