xiphmont: (Default)
[personal profile] xiphmont

Intra paint is not a technique that's part of the original Daala plan and, as of right now, we're not currently using it in Daala. Jean-Marc envisioned it as a simpler, potentially more effective replacement for intra-prediction. That didn't quite work out-- but it has useful and visually pleasing qualities that, of all things, make it an interesting postprocessing filter, especially for deringing.

Several people have said 'that should be an Instagram filter!' I'm sure Facebook could shake a few million loose for us to make that happen ;-)

Date: 2014-09-24 08:18 pm (UTC)
ext_659502: (Default)
From: [identity profile] some41.livejournal.com
I always wondered if a similar approach can work for audio: extract and encode transients separately, and use long block mdct for noise and tonal content.

Date: 2014-09-24 08:55 pm (UTC)
From: (Anonymous)
I thought that was the idea behind the "Ghost" codec research.

Date: 2014-09-25 12:59 am (UTC)
ext_659502: (Default)
From: [identity profile] some41.livejournal.com
if I remember correctly, Ghost does the opposite -- it extracts tonal signal.

Date: 2014-09-25 09:44 am (UTC)
From: (Anonymous)
> Ghost does the opposite -- it extracts tonal signal.

Which seems the right way to me. Noise doesn't need accuracy, unlike tonal signals...

Date: 2014-09-25 01:22 am (UTC)
From: [identity profile] jmspeex.livejournal.com
It's been tried several times, but I'm not aware of any practical codec where coding transients separately really makes a difference. It's really hard because there's no clear boundary between transient and non-transient and coding transients separately means more redundancy. In the end, the current solution of using shorter transforms works well enough for now.

Variable duration frames in Opus

Date: 2014-09-25 09:38 am (UTC)
From: (Anonymous)
What happened with variable duration frames in Opus? Did they help quality any?

Re: Variable duration frames in Opus

Date: 2014-09-25 07:56 pm (UTC)
From: [identity profile] jmspeex.livejournal.com
Not at the moment. Right now it's hurting and I haven't yet understood why (not a priority since the encoder is already pretty good).

Date: 2014-09-24 11:10 pm (UTC)
From: [identity profile] funroger.livejournal.com
Looks outstanding. At least there is no high-frequency noise.

Any hope for frequency domain intra?

Date: 2014-09-26 03:34 am (UTC)
From: (Anonymous)
So is this "we don't see the way forward for freq domain intra right now and so we're pursuing other leads while still hoping something may work out on that front"? Or is it "freq domain intra turns out to be provably impossible, no matter how many bright minds working round the clock you throw at it"?


Date: 2014-09-26 09:14 am (UTC)
From: (Anonymous)
"And because of lapping and block size alignment issues, we cannot just turn it on and off when we want. Yet another neat idea that turns out to fail due to practical considerations."

So blocks kill this. I was wondering...: Can't we move past blocks? Couldn't the images be encoded progressively and adaptively to the content (and gain some advantage)? Rectangle blocks aren't really a good representation of reality. (Unless you photograph a perfectly aligned checkboard pattern)

Re: Zs.

Date: 2014-09-27 02:00 pm (UTC)
From: (Anonymous)
> Can't we move past blocks?
No problem. Design a _fast_ codec with such properties you listed.
>Rectangle blocks aren't really a good representation of reality. (Unless you photograph a perfectly aligned checkboard pattern)
Daala can't encode rectangular blocks well, no matter how they're aligned.

Re: Zs.

Date: 2014-10-02 02:17 pm (UTC)
From: (Anonymous)
Nice trolling. But apparently you didn't realize that it was a question...

Date: 2014-11-23 04:38 pm (UTC)
From: (Anonymous)
Was this idea discarded?

Date: 2016-02-17 06:52 pm (UTC)
From: (Anonymous)

Date: 2016-02-17 09:20 pm (UTC)
From: [identity profile] xiphmont.livejournal.com
It was discarded as a coding tool, but it has been repurposed as a de-ringing post processing filter. This is one of the first pieces of Daala that is likely to make it into the new AOM codec.


xiphmont: (Default)

Most Popular Tags