Xiph & Mozilla's Daala codec

[dapperdan]

Just thought i'd start a generic thread since there doesn't seem to be one yet.

A bit late with this news but apparently there was a get-together last week to "start moving our Daala project from a research platform to a working codec prototype"

http://www.ietf.org/mail-archive/web.../msg00141.html

Some of the work/discussion can be found here:

https://daala.etherpad.mozilla.org/coding-party

[benwaggoner]

Can you give a summary of how and when Daala might be something interesting or useful?

[mandarinka]

It is still in a very alpha stage and the format is not frozen by far. Apparently though you can already test the proto-encoder and decoder. It already does inter prediction, but no scenecut detection and various other features (like rate control) aren't in yet. But I really don't have much insight into it, so you'll have to wait for somebody more relevant to describe the state...

[benwaggoner]

Ah. Sounds like something that wouldn't become relevant until 2015+ as a HEVC competitor then.

[mandarinka]

Note that my "very alpha" comment and the words about limited features having been implemented - all that is about the example encoder/decoder.

AFAIK they have the format's coding schemes laid out for a long time now, so while it has not been yet finalised and still being investigated, it is not like the format design is at the beginning.

[benwaggoner]

Originally Posted by mandarinka

Note that my "very alpha" comment and the words about limited features having been implemented - all that is about the example encoder/decoder.

AFAIK they have the format's coding schemes laid out for a long time now, so while it has not been yet finalised and still being investigated, it is not like the format design is at the beginning.

Sure. But few features survive intact between draft design and implementation .

For myself personally, I start really paying attention to codecs when I can actually make and playback a bitstream that doesn't something notable and new. It sounds like Daala is at least a couple years out from being able to do a real end-to-end test of a reasonably complete implementation of a reasonably final bitstream spec.

[mandarinka]

Xiph foundation has published a presentation about the codec (or rather it seems to be the first in a line of writeups that will follow...). It's be mainly about the lapped transform that is used in the codec.

http://people.xiph.org/~xiphmont/demo/daala/demo1.shtml

[mandarinka]

A third one is out too:

http://people.xiph.org/~xiphmont/demo/daala/demo3.shtml

[benwaggoner]

Okay. Now I'm thinking that Daala might actually turn into something brilliant. I've been thinking about the frequency domain prediction approach for a while now, and it's great to see some real evidence. TrueMotion prediction turning into a texture replicator after machine learning suggests some very powerful techniques that could be refined out of that.

I need to lie down and meditate on the potential implications of a lapped TF for video...

[dapperdan]

"Mozilla is organizing a coding party for Sept 30 - Oct 3 at the Mozilla headquarters in Mountain View to continue moving the Daala project further forward as a working codec prototype. Anyone and everyone are welcome to come hack on things."

http://www.ietf.org/mail-archive/web.../msg00143.html

Also, Monty's blog introducing the 3rd write-up that was linked above, suggests the second stage of the TF was discovered while documenting the first, which is interesting.

http://xiphmont.livejournal.com/60856.html

[MoSal]

A 4th demo was published last week

[khagaroth]

Also, Monty switched jobs and migrated from Red Hat to Mozilla to focus on Daala. .

[dapperdan]

A video of a talk by one of the Daala team, on the topic of Daala and Opus:

http://people.xiph.org/~xiphmont/vid..._and_Daala.ogv

Fairly high level, but up-to-date with their progress.

[mandarinka]

Hmm, videos require one to watch through them, unpleasant when you want to quickly extract the information

I wish people would return to *writing* reports... /well, I guess for that I need to wait for more Xiph demos./

[nakTT]

Originally Posted by dapperdan

A video of a talk by one of the Daala team, on the topic of Daala and Opus:

http://people.xiph.org/~xiphmont/vid..._and_Daala.ogv

Fairly high level, but up-to-date with their progress.

Hope you could do us a favour by sharing the interesting point raised in the video. Thank you in advance.

[dapperdan]

Slides from the talk video above:

https://people.xiph.org/~greg/gstreamer-daala-opus.pdf

[nakTT]

Originally Posted by dapperdan

Slides from the talk video above:

https://people.xiph.org/~greg/gstreamer-daala-opus.pdf

Thanks for the link.

[dapperdan]

And an article in LWN on the talk:

http://lwn.net/Articles/571978/

[qyot27]

Originally Posted by dapperdan

And an article in LWN on the talk:

http://lwn.net/Articles/571978/

I have to admit that the Theora-related "there's only so far you can get by putting rockets on a pig" comment made me chuckle.

I ran another compilation and really basic encoding test with Daala, and it has gotten substantially better since I last tested it several months ago: I can confirm that the claim in the article of artifacts looking grainy or noisy rather than blurry is true. When the ratecontrol and other accuracy and psy features are introduced, then this will be much more interesting - as it stands right now, you can have a 6000kbps stream generated by -v 10 and it still shows the grainy artifacts (although there is a very noticeable improvement going from -v 20, resulting in 3450kbps, to -v 10 @ 6000kbps).

[zerowalker]

How does it compare to VP9 and X265?
I don´t mean in real quality but in the state it is in currently.

From my understanding this is the latest one and has least development to it.