This site exists to fulfil a frequently heard reqest: To be able to compare the performance of codecs, consistently, openly, and usefully.
What’s Wrong with Measurements?
Most measurements of codecs occur in some specific context. Very often that context is set by constraints that appear artificial or unrealistic in the context where codecs are really used: In the real world - and most especially, on the Internet, in the delivery of video based services.
This site aims to generate video quality metrics under multiple scenarios, all of which bear some relation to Real Life. We don’t expect to be able to mirror real life perfectly - that would probably make it impossible to produce reproducible results, another goal, but we aim to find ways to score the performance of codecs in such a way that if two codecs differ significantly in score under a given scenario, the one that scores best is also likely to work best in practice in real life scenarios that resemble that scenario.
This site aims to be open in three very important ways:
Open source. Scripts, Web pages, video sources, the lot. Anyone should be able to, with a few simple commands, build their own copy of the website and its backing system, and reproduce our results.
Open contribution. Anyone who thinks they can make things “a little bit better” for a codec in a scenario should be free to make the proposal, submit a diff, and have it incorporated in the site. Anyone with a “better mousetrap” in the form of a new codec, a new metric or a new scenario emulator should similarly be able to contribute their code (provided it’s freely distributable under the terms we adhere to), and show how their codec, metric or scenario performs with all the others.
Open results. While this is an obvious consequence of being open source, it is still an important contrast to other efforts: What we measure will be public. There will be no secrets.
Testing a Whole Codec vs. Testing the Encoding in Isolation
Testing codecs has very often been an acrimonious matter, because some parties wish to test just a specific component of the encoding process, while others feel that ignoring properties that matter in real life situations is making comparisons unfair and unreasonable.
One simile (props to
FIXME from Ericsson for this): “If you want to test the
performance of an engine, you don’t want to put it in a car - you isolate it
on a measurement bench to remove the variability from your measurements”. True
as far as it goes - but conversely, if you are comparing cars, removing their
engines and testing them in isolation gives no information about their
handling characteristics on the road; for that, brakes, steering and
transmission all matter.
We have chosen to test codecs in, as far as possible, the way they will be used. This means that we test real implementations, that people use in production, and test them including all tools that matter, including rate control, filters, preprocessing, postprocessing and so on - if it improves the performance, and it’s clearly identified (and available in opensource), let it be used!
Scenarios We Aim to Cover
The scenarios we aim to cover in the initial set are the following:
Traditional “Fixed quality, no adaptation”. This is not an important version in Real Life, but is important because it allows some means of comparison with measurements done under this testing methodology at, for instance, MPEG.
Pre-encoded video targeted at fixed bandwidth delivery. This is a scenario where we can take the time we need for encoding, and use tools like multi-pass encoding - but the result needs to be measured on the experience it achieves if delivered over a fixed bandwidth channel; if peaks in bandwidth are big enough to overflow jitter buffers, for instance, the codec needs to be penalized.
Interactive video, targeted at fixed bandwidth delivery. In this scenario, encoding needs to be done one (or a few) frames at a time, with no lookahead into future frames (they haven’t been produced yet), and no or very limited buffering at the receiver side. Codecs that overrun the bitrate allocation at any point in time need to be sharply penalized for doing so. In this scenario, it’s also relevant to look at speed of encoding and decoding; encoding that cannot keep up with real time on a machine running these tests is unlikely to be very useful in real life if one depends on a software codec.
Other scenarios may be added once a consensus emerges on what they should be, and once the tools are available for measuring the metrics that are important in those scenarios.
Issues This Site Does Not Cover
All that this site is concerned with is that it produces reproducible results. It is absolutely unconcerned with many details that matter for real deployments, such as commercial licenses, patent restrictions, flexibility of software, tunability for special scenarios and so on.
These issues can only be evaluated in the context of a specific usage scenario and business model; it is not possible for a project like this to give useful guidance in those matters.