Slashdot Mirror
More On The BBC's Codec 'Dirac'
TioHoltzman writes "El Reg is reporting about a new codec that is built on top of wavelet technology and seems to offer performance that is "roughly in line with the Video Codec 9" from Microsoft. The project has been released as open source on SourceForge. This looks like it might be really interesting." (Previously mentioned a few weeks back.)
Last time I checked, wavelet compression methods were burdened by many patents: google search. What does that mean for users of the codec?
"I love my job, but I hate talking to people like you" (Freddie Mercury)
The standard way to compress both audio and video is with the Discrete Cosine Transform, or DCT. MPEG audio and video are based on DCT.
The basic idea of DCT is to transform the data into a series of waves, which tends to concentrate the data. Then you throw away part of the data, and then use lossless encoding on what is left. If you just threw away pixels, the result would be obvious in an image; but if you throw away part of the wave specification data, the results are not as obvious.
With DCT, consistent data sets compress very well (e.g., a blue sky or a white wall). Pictures with lots of sharp little edges (e.g., a field of blades of grass) compress much less well.
My understanding is that potentially wavelets will compress even better than the DCT. However, they are not enough better to be a huge win at the moment.
steveha
lf(1): it's like ls(1) but sorts filenames by extension, tersely
Not to mention this other canny Brit.
as for wavelet compression being a novel codec, what about apple's pixlet technology?
Some drink at the fountain of knowledge. Others just gargle.
This is from 1998.
http://www.seyboldreports.com/SRIP/wavelet/
steveha
lf(1): it's like ls(1) but sorts filenames by extension, tersely
While wavelets doesn't offer a breathtaking advantage in data rate vs. quality factor, it does appear to lend itself to a simpler implementation than does DCT, and unlike MPEG, which is very intensive on the encoder, wavelets places symmetrical burdens on encoder and decoder.
It was a core assumption in the design of MPEG that the world market for encoders was quite small (where have we heard that theme before???) Clearly, the assumption was false, and one advantage of switching to a wavelets technology would be reduced cost per unit for encoders.
--- Bill
Dirac (pronounced Dih-RAK) was a physicist and mathematician. His name appears in this context because of the "Dirac delta function," otherwise known as an impulse function. It lies at the heart of linear signal theory, including wavelet theory.
Nothing to do with the government. The BBC is granted a charter from Parliment, but is not government run or funded. The BBc is funded by a compulsory license fee for owning equipment capable of recieving and decoding their broadcasts such as a TV or tuner card. Basically it's a tax on virtually every household and business in the UK. There is a discount for black & white TV's, pensioners and those with vision based disabilities. In the 'old days' you used to need a 'wireless licence' as well for radios!
This means that when information is dropped in each block (according to the compression required), the edges of blocks suffer in a way unrelated to the edge of adjacent blocks. The result -- as the quality decreases, the edges between blocks become more and more obvious, and the whole image becomes 'blocky'.
I believe this is one way that wavelet technology improves -- the individual wavelets are spread over the whole image, without regard for any blocks, and so the compression degrades much more gracefully.
As you say, the DCT converts each 8x8 block into a series of cosine waves, both horizontally and vertically in the block. Then, when it needs to reduce the space, it drops the higher-frequency coefficients first -- this is why sharp edges, with lots of high frequency information, suffer most. (You tend to find that lower-frequency coefficients try to compensate, giving the characteristic ripples near sharp edges.) Areas that are relatively smooth, with only low-frequency information to start with, suffer much less.
Another way JPEG loses information is by colour. The human eye is much more sensitive to fine changes in brightness than it is to fine changes in colour; so the picture is transformed from RGB into a brightness channel and two colour channels, and the brightness channel gets a greater share of the limited space. It's quite interesting, if you're, er, interested in that sort of thing...
Ceterum censeo subscriptionem esse delendam.
Pixlet is designed for real-time editing, so it has minimal artifacts and no interframe compression. Dirac is for broadcast, so it is much more agressive about compression and can take advantage of motion compensation and other computationally expensive compression techniques.
You are right, however, that wavelets are not at all a new compression technology. People started playing with it at least 10 years ago and JPEG-2000 uses wavelets for still photo compression. I think that the computational load has prevented their use in video until recently.
Potential for profit may go down, but I'm talking innovation, not wallets.
In theory, the BBC hasn't been all that interested in profits, being a non-profit taxpayer funded organisation. I was starting to wonder what I payed a license fee for, but if they carry on like this I'll be quite happy to keep paying it.
I welcome the BBC's foray into OSS, and I hope it'll be the first of many OSS sucesses for them.
mplayer plugin plays whatever mplayer plays. and mplayer plays quicktimei have been watching apple trailers for quite some time
BSD is for people who love UNIX. Linux is for those who hate Microsoft.
Just one question. Have you even used Mozilla or Firefox lately? It certainly use to be a buggy piece of garbage that I avoided for years (after Netscape 4.7) but I'd argue its now better and more stable than IE (which isn't hard).
Mark my words, within five years DIRAC will be bigger than MP3 is now.
You're all too willing to predict the future. I wish I had your prescience. You may be right but like Bill Gates and Larry Ellison have been to the tune of many millions and they're the successful ones.
These posts express my own personal views, not those of my employer
The BBC have a long history of supporting computer technology.
Many of us English folks grew up with the BBC Micro computer in schools.
liqbase
The BBC hasn't been a monopoly since 1955, and no, it's not a function of Government, there is no minister for the BBC nor is there a chain of command within government in charge of managing the Beeb. We have agencies, lots of them, but the BBC isn't amongst them.
It's an independent public body incorporated via a Royal Charter. Just because it appears soft-left doesn't mean this is encouraged or engineered by Government, it's much to their annoyance in fact, take the Dr. Kelly affair or their war coverage from last year for example.
ogg is the container. Just like AVI and quicktime, any number of codecs for audio and video (or other) can be encapsulated in ogg.
Vorbis is the audio codec we know and love so well.
Unlike an 8x8 DCT (for example), fractal compression is generally scale independent. A block of pixels is represented by a contractive mapping that can be applied to ANY size block. The mapping is applied iteratively and can be proved to converge, regardless of the initial values or the size of the block. If applied to a bigger block than in the original, the algorithm is simply "making up" detail information that wasn't in the original. In some cases, the detail looks quite realistic. In other cases, it doesn't. Researchers, of course, tend to show images that make their algorithms look good (that's not a criticism, just an observation).
Fractals are not the only class of algorithms where this can be done. It is possible to extrapolate additional levels of detail (high-frequency) information from a wavelet-encoded image. I have seen very effective image sharpening techniques based on wavelets that provide more detail than in the original image.
H.264 is going to become "MPEG4 AVC", Advanced Visual Codec. This is one of the 3 compression standards due to be approved (or maybe actually approved by now) for HD-DVD. The other 2 are WM9 (love it or hate it) and MPEG2 (for those that have more bits than sense).
Again, and AFAIK, MPEG4 is more of a "wrapper" format than anything else. Thus, it's becoming a little confusing as "MPEG4" won't mean the same thing once H.264 finally comes out.
I saw some 8Mbps H.264 (Quicktime and others) at NAB 2004 in Vegas and it looked AMAZING. It was 720p on a huge plasma. I was VERY impressed although saddened to see those smooth areas still had that weird problem with posterization (or whatever it's called). Ya know, smooth skies and so on. However, fast movement was perfect. It uses a crap-load of CPU (especially for encoding!) but I think it will win out overall. Looks similar to some HD 8Mbps WM9 that I've seen....
This concept is referred to generically as "honorifics." It is not unique to Japanese. But you can certainly judge a person's competence in Japanese by their ability to juice up the honorifics. For instance, you score more points for saying, "Would you graciously grant that my weak feable eyes might bask in the glory of your amazing garden." In the same sentence you abase yourself while you elevate the garden, creating as much space as possible between the two. More points are awarded for making it passive: "Could my meager feable eyes be most graciously commanded to humbly partake of the glory within your majesty's splendid garden."
Here's the relevance to video codecs: In Japanese, the word for look or see is _not_ different in each sentence. The same word is used; it is just conjugated differently. It can go from base to super honorific to passive using the same character. So the the video presentation of this lesson could be achieved with a GOP that transmits the kanji for miru (look/see) once, and only the changes in conjegation are transmitted for subsequent frames. Since conjugations are written in simple kana, and not kanji, the cost of the extra frames gets even less expensive.
But you can certainly judge a person's competence in Japanese by their ability to juice up the honorifics. For instance, you score more points for saying, "Would you graciously grant that my weak feable eyes might bask in the glory of your amazing garden."
Oh no, it's certainly nothing that simple. If you're too polite, you start to sound sarcastic. At the very least you'll make the owner of said garden uncomfortable, which should be avoided at all cost. You have to be exactly the right level of polite, or maybe slightly more so.
But I doubt any of this will do Feynman any good.
It is relatively early days in terms of Wavelet and Fractal technology. Looking at JPEG DCT vs JPEG2000 Wavelets vs Fractals... What makes an image look bad in terms of image quality is the ability of our brain to pick out unnatural patterns from an image. The simplest way to represent an image with less information is to reduce the number of pixels that form the image. Problem is that if we try to do this too much our brain picks up little squares that form the image. What happened with JPEG was that instead of using unnatural looking squares, images were broken down into natural looking sine waves, which when used in an image looked quite convincing. So when you have lots of data to use for image storage, say compression levels of 1:10, there is not much difference between Wavelets, Fractals, DCT and the original image. But when you start to go to compression levels of 1:50 and beyond JPEG image quality falls off because you loose the high frequency components of the image that give the image fidelity. But wavelet patterns are much more complicated and therefore harder for our brain to predict so images based on this complicated patterns look much better. Fractals could possibly have infinite complexity and possibly represent just about any image with very little data. But its like decryption where finding the right key is like finding a needle in a thousand-million haystacks. Fractal compression has the most potential but currently fractal compression works with very simplified mathematical models. I predict that we are starting to hit a technological barrier in terms of DCT and video, so we are starting to move to exploiting wavelets. After we start to exploit the limits of wavelets then fractals will be the answer. Actually I read a paper that showed a mathematical link between wavelets and fractals. The paper was a bit beyond me, but from what I understood wavelets have an interative nature that fractals also have, and the paper somehow demonstrated this mathematically.
Yet another ironic recursive statement.
The general method of detection is that they send you a letter if you haven't bought a licence for your house, such as when the previous owner moves out and takes their licence with them.
It's at that point where you buy a licence, or tell them you don't own a TV.
Obviously, I've never tried this personally, but scuttlebutt (and TV adverts) indicate they send someone out to your house to see if you really don't have a TV. If you're muppet enough to have your TV visible from the road when you're using it, then expect a nastygram threatening (civil) court proceedings. Up to 400 quid fine, I believe.
If you refuse to let them in (which you can do of course, they won't have a warrant), and they still suspect you of having a TV (aerial or satdish on the roof, for example) then they can send a TV detector van.
Theoretically, an operating TV also acts as a low power transmitter which can be picked up with a close range directional aerial. Vans wandering through the street looking for non-payers is a myth - they have to sit right outside your house with a directional aerial looking for the signal. I've never seen one personally.
Note, you can't go to jail for not paying and getting caught. It's a civil offence, not a criminal one. It's a bit like getting caught for not paying your road tax...
Remember kids, it's all fun and games until someone commits wholesale galactic genocide.
The BBC is a quasi-autonomous non-governmental organisation (QUANGO). IIRC they are empowered by their royal charter to collect the licence fee, which in theory is the source of all tax raising powers in the UK. The board of governors are appointed by the government and are charged with serving the interests of the viewers, and maintaining editorial independence. On of their duties as laid out in the charter is to perform R+D in line with their overall obligations.
http://www.bbc.co.uk/info/policies/charter/
You can't win Darth. If you mod me down, I shall become more powerful than you could possibly imagine