Alliance for Open Media codecs

[Clare]

Originally Posted by wiak

hello there, your zenpai here! i got a 1700X

I'm looking to buy one too (or maybe wat for the Zen+ in February), what kind of fps do you get for AV1 or x265?

[bstrobl]

Streaming Tech Sweden AV1/HEVC comparison by Jan Ozer: https://www.youtube.com/watch?v=KzqOIddTNPQ

Hopefully the AV1 update by Jai Krishna will be uploaded soon. For those who are impatient, the slides are here: https://speakerdeck.com/stswe/av1-up...an-from-google

[hajj_3]

Originally Posted by bstrobl

Streaming Tech Sweden AV1/HEVC comparison by Jan Ozer: https://www.youtube.com/watch?v=KzqOIddTNPQ

Hopefully the AV1 update by Jai Krishna will be uploaded soon. For those who are impatient, the slides are here: https://speakerdeck.com/stswe/av1-up...an-from-google

only ~25% better compression than vp9 so far. They were claiming 30% by bitstream freeze, lets hope they make it.

They have delayed the bitstream freeze to January now instead of end of year, not a big deal.

[VincAlastor]

https://demo.bitmovin.com/public/firefox/av1/

how can i grab this demo stream? I wanna analyse and test this stream outside of firefox.

[sneaker_ger]

youtube-dl https://bitmovin-demos.commondatastorage.googleapis.com/mozillaAV1_Oct2017/e87fb2378f01103d5d6e477a4ef6892dc714e614_v1/stream.mpd

[bstrobl]

Mozilla Blog: https://hacks.mozilla.org/2017/11/da...-of-av1-video/

[mzso]

Hmmm. It looks like they removed the 1080 videos from the demo page.

[benwaggoner]

Originally Posted by bstrobl

Hopefully the AV1 update by Jai Krishna will be uploaded soon. For those who are impatient, the slides are here: https://speakerdeck.com/stswe/av1-up...an-from-google

It makes me somewhat nervous that the percentage gains go down the more subjectively correlated the metric used is.

-26.51% for VOD PSNR-Y but only -23.57% for MS-SSIM. And no metrics that include any temporal component. I'd love to see at least VMAF scores, which is the metric most predictive of DMOS that we have today.

The VPx series was always heavily tuned to optimize for PSNR, so its theoretical PSNR always overstated the series' actual perceptual quality. At least since VP6, which had a lot of advanced postprocessing technology that did a good job of hiding artifacts in a lot of ways.

[IgorC]

Originally Posted by VincAlastor

https://demo.bitmovin.com/public/firefox/av1/

how can i grab this demo stream? I wanna analyse and test this stream outside of firefox.

Originally Posted by mzso

Hmmm. It looks like they removed the 1080 videos from the demo page.

Yep.

Also audio is Opus @ 32 kbps. Opus is the best available audio codec right now though 32 kbps is low.

Opus @ 48 kbps has much better quality which makes more sense for 720p streams (and 64-96+ kbps for 1080p)

[VincAlastor]

Originally Posted by IgorC

Yep.

Also audio is Opus @ 32 kbps. Opus is the best available audio codec right now though 32 kbps is low.

Opus @ 48 kbps has much better quality which makes more sense for 720p streams (and 64-96+ kbps for 1080p)

But it's amazing what audio quality is possible with 32 kbit/s! Thank you IgorC and all other opus audio developers!

I can't wait for upcoming opus 1.3 alpha/beta release!

[bstrobl]

Originally Posted by benwaggoner

It makes me somewhat nervous that the percentage gains go down the more subjectively correlated the metric used is.

-26.51% for VOD PSNR-Y but only -23.57% for MS-SSIM. And no metrics that include any temporal component. I'd love to see at least VMAF scores, which is the metric most predictive of DMOS that we have today.

The VPx series was always heavily tuned to optimize for PSNR, so its theoretical PSNR always overstated the series' actual perceptual quality. At least since VP6, which had a lot of advanced postprocessing technology that did a good job of hiding artifacts in a lot of ways.

I suppose the good news here is that AV1 has had no form of optimisation beyond the basic stuff done to it yet. Might be able to squeeze out a couple percent more with improved encoders. Right now encoding time is the biggest problem, especially for RTC.

[benwaggoner]

Originally Posted by bstrobl

I suppose the good news here is that AV1 has had no form of optimisation beyond the basic stuff done to it yet. Might be able to squeeze out a couple percent more with improved encoders. Right now encoding time is the biggest problem, especially for RTC.

I'm sure there are massive potential gains that can be squeezed out by real-world encoders. Something like x264 and x265 can run rings around the reference encoders with real-world long form content using real-world parameters, and still be orders of magnitude faster.

It just takes a whole lot of development that's out of scope of defining a bitstream to get those gains. Codec development focused on PSNR @ Constant QP also runs the risk of over optimizing for a design where QP correlates highly with PSNR.

[bstrobl]

Originally Posted by benwaggoner

I'm sure there are massive potential gains that can be squeezed out by real-world encoders. Something like x264 and x265 can run rings around the reference encoders with real-world long form content using real-world parameters, and still be orders of magnitude faster.

It just takes a whole lot of development that's out of scope of defining a bitstream to get those gains. Codec development focused on PSNR @ Constant QP also runs the risk of over optimizing for a design where QP correlates highly with PSNR.

Doesn't HEVC do the same thing to an extend? Besides, the VP9 base will still be prominent in AV1 due to limited development time and the desire for reusability of components by chipset manufacturers.

The fact that AV1 does well when compared to x265 with a five year head start is neat, but I won't deny the fact that this was a somewhat rushed development.

My guess is most desirable features and improvements will only come with AV2, when the Alliance have most of the development kinks worked out.

[iwod]

Originally Posted by bstrobl

Doesn't HEVC do the same thing to an extend? Besides, the VP9 base will still be prominent in AV1 due to limited development time and the desire for reusability of components by chipset manufacturers.

The fact that AV1 does well when compared to x265 with a five year head start is neat, but I won't deny the fact that this was a somewhat rushed development.

My guess is most desirable features and improvements will only come with AV2, when the Alliance have most of the development kinks worked out.

I am hoping for the best on AV2. But given how they have communicate, and expecting to have AV2 and AV3 done in next 5 years just shows they have absolutely no idea how real world progress and works.

Then there is the law of diminishing return, a lot of the improvement and progress made in x264 and x265 took years to fine tune. It wasn't long ago x264 was still beating x265 at high bitrate encoding.

[mzso]

Originally Posted by iwod

I am hoping for the best on AV2. But given how they have communicate, and expecting to have AV2 and AV3 done in next 5 years just shows they have absolutely no idea how real world progress and works.

I don't have high hopes. They'll just probably keep tweaking the same basic concept (dct on blocks) as it's done for decades now.

Originally Posted by iwod

Then there is the law of diminishing return, a lot of the improvement and progress made in x264 and x265 took years to fine tune. It wasn't long ago x264 was still beating x265 at high bitrate encoding.

That's why it would make sense to head to a radical new direction. But they won't.
I doubt that there'll be any interest in adopting AV2 if it ever gets finished. By then AV1 will be pretty well optimized and AV2 won't offer any improvement of particular significance.

[bstrobl]

Originally Posted by mzso

That's why it would make sense to head to a radical new direction. But they won't.
I doubt that there'll be any interest in adopting AV2 if it ever gets finished. By then AV1 will be pretty well optimized and AV2 won't offer any improvement of particular significance.

Nothing wrong with DCT on blocks, worked well for a long time and still does .
It's a bit like squeezing out everything from current microprocessors and other modern technology, eventually you simply reach what is actually possible without going off into ridiculous complexity/cost. Daala is a good codec with a different design, but doesn't have a huge benefit over HEVC/AV1. Same with wavelets.

I doubt the Alliance will release a new codec if they can't make it at least 30% better than AV1. If there is no AV2, we will at least have a solid codec that is royalty free to depend on. Nothing wrong with a long lasting standard when you don't have to replace hardware/software and re-encode everything.

[mzso]

Originally Posted by bstrobl

Nothing wrong with DCT on blocks, worked well for a long time and still does .
It's a bit like squeezing out everything from current microprocessors and other modern technology, eventually you simply reach what is actually possible without going off into ridiculous complexity/cost.

Yeah, it feels like most of the complexity of present codecs is spent on catering to blocks and deblocking.

Originally Posted by bstrobl

Daala is a good codec with a different design, but doesn't have a huge benefit over HEVC/AV1. Same with wavelets.

I think alternative approaches are sorely under-researched. Not much effort was put into wavelets, even less into Daala's overlapped transformation before it was dopped.

[benwaggoner]

Originally Posted by mzso

Yeah, it feels like most of the complexity of present codecs is spent on catering to blocks and deblocking.

Well, they ARE block-based codecs! New codecs are adding LOTS of knobs around block sizes and deblocking modes.

I think alternative approaches are sorely under-researched. Not much effort was put into wavelets, even less into Daala's overlapped transformation before it was dopped.

That is one of the challenges with any radically new technology. We have decades and millions of engineer-hours of experience optimizing around DCT-like block-based encoding. Even if a fundamentally new transform had a lot more potential, its initial implementations would have to compete against all the refinement of the current architectures. And it's very hard to estimate how well a basic technology will pan out. Wavelet-based video codecs were "obviously" the future, until actual real-world implementations went splat.

If we started with J2K instead of JPEG, and had all the motion estimation work done based around wavelets, hundreds of PhDs and developers could have found the magic way. And then someone with a crazy block-based codec idea would have run out of funding at just full-pel motion search, and we'd say "yeah, blocks were kind of interesting, but just never proved to be competitive."

[bstrobl]

Took a while but here is the Google update on AV1:

https://www.youtube.com/watch?v=sBAH...aqcvb0&index=7

[hajj_3]

Originally Posted by bstrobl

Took a while but here is the Google update on AV1:

https://www.youtube.com/watch?v=sBAH...aqcvb0&index=7

The video says that they are currently up to 29% improvement over VP9 and aim to be at ~35% by the bitfreeze in January. That is a welcomed improvement as they say a month or so ago that they were aiming for 30% improvement.