4. Why are you releasing Dirac Open Source?
6. What are the license conditions?
7. Do you infringe any patents?
8. What will you do if you infringe patents?
9. Is the BBC going to stream video using Dirac?
10. Are you going to broadcast TV using Dirac?
13. Why did you choose the techniques that you're using?
14. What's with your picture file format?
15. How do I play compressed video?
The BBC Dirac team is led by Tim Borer, Principal R&D Engineer in the BBC Research and Development Department. The main algorithm developer is Thomas Davies, Senior R&D Engineer, who devised the Dirac algorithm. In the next couple of months our little team will grow to incorporate software engineers to develop Dirac into a real software system. BBC R&D has always been involved in video coding research as it's central to what the BBC does. Thomas began experimenting with compression techniques in about 2001, and developed various prototypes of coding tools. In January 2003 he had a rough research system in place which seemed to perform well. Since then he's been working to translate the software into something that could be the basis of a working codec, and to streamline the design. This involved rewriting the codec entirely in C++ to exploit the flexibility and modularity of the language. Dirac is named after the British physicist (of Swiss extraction) Paul Dirac, who was one of the great figures in modern physics. We've seen some good explanations on the 'net, suggesting that we chose the name because the Dirac delta function is fundamental to signal processing and that Dirac's work underpins wavelets. Sorry, but the stories aren't true. We thought of hundreds of names, most of them awful, some of them good, and most of the good ones had been used already. Tim suggested Dirac. It sounded good. It stuck. It was going to be called Camel, so maybe we had a lucky escape. You can find out more about Paul Dirac here.4. Why are you releasing Dirac Open Source?
The BBC has always advocated open standards, and has tried to use them where possible. So far, streaming has been dominated by proprietary systems and existing licensing regimes for standards-based systems have not been as attractive as they might be for large-scale broadcasting, particularly for Public Service broadcasters. You can download the latest version from the Sourceforge project page. You can also look at the developing code on the CVS there.6. What are the license conditions?
Dirac is released under the Mozilla triple license. This is an Open Source license that allows both free and commercial use of the software. It also allows for relicensing under the GPL or the LGPL.7. Do you infringe any patents?
The short answer is that we don't know for certain. We haven't employed armies of lawyers to trawl through the tens of thousands of video compression techniques. That's not the way to invent a successful algorithm. Instead we've tried to use techniques of long standing in novel ways. Where we think we're novel, we're in the process of getting patent protection ourselves, which will invalidate others' claims of priority. There are some areas that are more heavily patented than others. Arithmetic coding is one such, even though the technique itself has been around for 25 years. We're keeping an eye on the situation, and we'll adopt alternative techniques if we have to.8. What will you do if you infringe any patents?
Code round them, first and foremost. There are many alternative techniques to each of the technologies used within Dirac. Dirac is relatively modular (which is one reason why it's a conventional hybrid codec rather than, say, 3D wavelets) so removing or adding tools should be possible until it goes Beta and the bitstream and decoder functionality are fixed. But we also need to see where in the world potentially problematic patents apply, and how important they are.9. Is the BBC going to stream video using Dirac?
Not any time soon. The software is nowhere near ready for that - we need real-time decoding, integration with players, a bitstream spec and a transport stream format. So this will probably have to wait until Dirac is close to beta.10. Are you going to broadcast TV using Dirac?
Broadcasting or multicasting over the web is likely to develop over the next few years, and the technology we use for this is likely to be the same as for standard streaming. There is no likelihood at all of replacing our existing Digital TV infrastructure, based on European DVB standards and using MPEG-2, with anything new - the installed base of millions of customers is too large. We're glad that it looks like Theora will reach beta soon. We think you can't have too many free codecs, but that the Open Source community also needs to continue to develop codecs with increasingly better performance. We also think there could be a good deal in Theora that we could use in Dirac, and we'd like to work with developers who've been closely involved in Theora. We intend to pack the Dirac elementary stream into MXF, which has lots of useful features. That doesn't preclude it packing into Ogg as well, and it's probably a good idea to have a variety of packing formats. For this the elementary stream needs to be very well defined. See the algorithm documentation13. Why did you choose the techniques that you're using?
We chose wavelets for a number of reasons. They perform very well in still image compression, and can be said to be state of the art there. They also provide a degree of scalability, so a codec based on wavelets can perform well across a range of video standards. By scalability we don't mean embedded bitstreams or the ability to extract lower-resolution video from higher-resolution bitstreams - Dirac doesn't do this as it's very complex - only the flexibility to apply the same tools to a range of resolutions, perhaps with different parameters. Wavelet coding is also a well-studied and well-understood field. Overlapped block motion compensation is used to reduce blockiness and ease the job of the wavelet transform in coding motion compensated differences. The wavelet transform may not be the best tool to do this, and we'd welcome other contributions, but it does have the benefit of simplicity in that the tools for all frame types are the same. Dirac supports more or less any block sizes for motion compensation and this again helps in scaling the algorithms as larger blocks can be used for higher resolution pictures and smaller ones for low resolution pictures.14. What's with your picture file format?
It's a temporary fix. There is a dearth of good file formats for video sequences, most of them being multi-file formats, with little or no metadata. We're going to release some tools for converting formats soon. The encoder and decoder applications are just example programs: what really matters are the libraries, which just need classes capable of reading and writing picture data. Ultimately we'll subclass these picture IO classes to cover every file format we need to, and we'll also wrap uncompressed video in MXF as well, for which there are player utilities. You can't yet play a compressed bitstream, but hopefully it won't take too long to get a DirectShow filter, for example, sorted out. Bear in mind that any code written to play compressed video is liable to change until the wrapping format and demuxing operations, at least, is determined. You need video in the right format, which is raw planar YUV with a separate header file. There's a program to make the header. Then you need to start the encoder running. Controlling the encoder is something of a black art, but there are some instructions in the distribution. By telling people about Dirac. By contributing to the code: contact us to find out more. . A very great deal. The Dirac algorithm is still evolving, and new tools (for interlaced video, for example) and enhancements to the current tools are needed. The code needs optimisation for speed, in order to be able to code and (especially) decode in real time. The bitstream formats (elementary stream and packing) need defining. The decoder libraries need to be integrated with a variety of players. For a full(er) list, see the TODO list in the distribution. At the moment, we have no idea.