Google VP9 "Next Generation Open Video" information posted

[LigH]

In addition, as already stated: In YUV 4:2:2 (YUY2 / UYVY), dividing the bits of used data per frame by the number of pixels per frame, every pixel would have only an average bit amount (with hardly any "meaning per pixel") because the two chrominance components share the blue/yellow and the red/green difference from grey among two pixels. With 8 bit per component, the result would be an average 8 + 8:2 + 8:2 = 16 bit; but these 16 bit are not addressable per pixel, only per 2 pixel group. Even worse for YUV 4:2:0 (YV12 / i420), there are shared chroma difference values among a 2×2 pixel square to an average 8 + 8:4 + 8:4 = 12 bit per pixel. Such statistics have too little meaning if your scope is a single pixel, they are even misleading.

[Jamaika]

Some improvements VP10 (currently 12-20% bitrate reduction):

•high bit-depth internal: 10/12 bits (also artificially to improve prediction quality: 2-4% gain for 8-bit content),
•partition structure: 8-way partition (4 in VP9), 128x128 unit size (64x64 in VP9),
•new intra prediction, “recursive 3-/4-tap filtering framework from the left and top edges”, more directional prediction modes,
•Inter-prediction – up to 6 reference frames (3 in VP9),
•InterIntra – “A compound prediction mode that combines an inter- and intra-predictor”,
•OWMC (overlapped wedge motion compensation) - “allows blocks to have oblique partitions”, using “wedge codebook”: for dividing a block into parts using separate predictors (inter/intra),
•OBMC (overlapped block motion compensation) – “blend multiple predictors using motion information from neighboring causal blocks”,
•Global motion: pan/tilt/zoom/shake, gaming,
•“Super transform: transforms that cover multiple prediction blocks, a combination of prediction and transform”,
•Extended transforms: 16 transforms ((DCT/ADST/FlipADST/DST)^2) for 4x4/8x8/16x16, explored 32x32 and 64x64 wavelet/DCT hybrids, transform skip mode, recursive transform units,
•Entropy coding – only Asymmetric Numeral Systems mentioned as much faster for decoding.

[zerowalker]

A bit confused about this stuff.

Google and others work on an Open Video Encoder thingy, but they still make the VP10, then there is Daala which is also aiming for the same thing?
Why are they allied if they are competing, am i missing something here?

[BadFrame]

Originally Posted by zerowalker

Why are they allied if they are competing, am i missing something here?

The AOM (Alliance of Open Media) codec is based upon VP10 (which is still in heavy development as a separate branch), but they will use any useful technology which their members have access to, for example if you look at the AOM branch, there are Daala features being added:

https://chromium-review.googlesource...oject:webm/aom

I don't know if VP10 will actually materialise as a separate stable codec release or if it will be decprecated as it is the base of the AOM codec (depends on how long it will take to develop I guess).

[mandarinka]

Originally Posted by Jamaika

Some improvements VP10 (currently 12-20% bitrate reduction):

•high bit-depth internal: 10/12 bits (also artificially to improve prediction quality: 2-4% gain for 8-bit content),
•partition structure: 8-way partition (4 in VP9), 128x128 unit size (64x64 in VP9),
•new intra prediction, “recursive 3-/4-tap filtering framework from the left and top edges”, more directional prediction modes,
•Inter-prediction – up to 6 reference frames (3 in VP9),
•InterIntra – “A compound prediction mode that combines an inter- and intra-predictor”,
•OWMC (overlapped wedge motion compensation) - “allows blocks to have oblique partitions”, using “wedge codebook”: for dividing a block into parts using separate predictors (inter/intra),
•OBMC (overlapped block motion compensation) – “blend multiple predictors using motion information from neighboring causal blocks”,
•Global motion: pan/tilt/zoom/shake, gaming,
•“Super transform: transforms that cover multiple prediction blocks, a combination of prediction and transform”,
•Extended transforms: 16 transforms ((DCT/ADST/FlipADST/DST)^2) for 4x4/8x8/16x16, explored 32x32 and 64x64 wavelet/DCT hybrids, transform skip mode, recursive transform units,
•Entropy coding – only Asymmetric Numeral Systems mentioned as much faster for decoding.

Note that some of those tools might be sacrificed later due to performance concerns ( ) or disappointing results. If you look at Daala, it dropped a large parts of originally planned features, with lapping being evaluated for possible removal too IIRC - they are not sure its complexity brings anything on top of simple loop filtering. Global Motion was famously a feature of Mpeg4 ASP, but the consensus was that it was useless, back then.

Originally Posted by Jamaika

•Inter-prediction – up to 6 reference frames (3 in VP9)

Aww, I hoped they would be more ambitious, this is still just catching up to HEVC (8) and AVC (16).

Edit:

Originally Posted by Jamaika

partition structure: 8-way partition (4 in VP9)

This seems to be just another name for rectangular partitions that AVC/HEVC already has, based on what they show in the video (atlhough it might not support partitions like 4x16 or 8x32?). I guess this is to get around patents, like that convoluted way to copy over b-frames in VP9.

Those oblique and wedge partitions are ambitions though, I wonder if that makes into the final codec (fingers crossed).

"Rotation/zoom motion"
\o/

"Intra-block motion" should also be useful.

Also, anyone who checked the video: is it just me or does the thing they call "OBMC" not really mean OBMC? Isn't that just combining/merging of predictors that conventional codecs do? The blocks on the picture in presentation don't look like being overlapped.

[nevcairiel]

Originally Posted by mandarinka

Aww, I hoped they would be more ambitious, this is still just catching up to HEVC (8) and AVC (16).

The fact that HEVC moved down from 16 to 8 should already tell you something - its just not worth it to have such a high number of refs.

[mandarinka]

Generally perhaps, but don't you know all that matters in encoding is anime?

[mzso]

Originally Posted by mandarinka

Note that some of those tools might be sacrificed later due to performance concerns ( ) or disappointing results. If you look at Daala, it dropped a large parts of originally planned features, with lapping being evaluated for possible removal too IIRC - they are not sure its complexity brings anything on top of simple loop filtering. Global Motion was famously a feature of Mpeg4 ASP, but the consensus was that it was useless, back then.

Well, if they remove the lapping transformation, then they removed the only notable thing about the codec. The thing the codec is known for.
They might as well just dump the project and start enhancing existing formats or contribute to VP10. (Unless they put something more radical in there)

[mandarinka]

Originally Posted by mzso

They might as well just dump the project and start enhancing existing formats or contribute to VP10. (Unless they put something more radical in there)

It's almost what they do now, actually.
Frequency domain intra prediction is dead, some other thing too but I can't recall what it is. They originally thought lapping will spare them from the need to have loopfilter. But without loopfilter, they have terrible ringing because they don't have intra because lapping. I think Daala currently has at least two loopfilters (complex deringing one, then some smoothing(?) thing from Thor), so that beggs the question if the lapping was worth it, ever? Losing intra prediction at least to me sounds like a potential bomb.

The big remains from Daala specialities is their entropy coding and PVQ. Those are still used, but they are also experimenting with porting them to VP10. AOM might mean Daala won't actually ship.

[foxyshadis]

Originally Posted by nevcairiel

The fact that HEVC moved down from 16 to 8 should already tell you something - its just not worth it to have such a high number of refs.

HEVC certainly supports 16 refs, at least for B-frames; x265 just limits it to 8. If x265 ever gets long-term references, or even a reference structure beyond pure hierarchical, those extra refs might come in very handy, but for now, yeah, what's the point?

[nevcairiel]

Originally Posted by foxyshadis

HEVC certainly supports 16 refs, at least for B-frames; x265 just limits it to 8. If x265 ever gets long-term references, or even a reference structure beyond pure hierarchical, those extra refs might come in very handy, but for now, yeah, what's the point?

While the HEVC bitstream can have 16 references, every individual picture can only access 8 of them, at least in all defined profiles/tiers/levels. But you get a bit of extra flexibility by keeping more around for another frame, I suppose.

[dapperdan]

https://blogs.windows.com/msedgedev/...icrosoft-edge/

VP9 support returns to Windows 10

[Motenai Yoda]

Originally Posted by dapperdan

https://blogs.windows.com/msedgedev/...icrosoft-edge/

VP9 support returns to Windows 10

I don't get why disabling vp9 even on desktop...

[Nintendo Maniac 64]

Originally Posted by Motenai Yoda

I don't get why disabling vp9 even on desktop...

Because even on the desktop, VP9 decoding is quite intensive - my 2.4GHz Core 2 Duo can only do 1080p 60fps VP9, and that's even in MPC-HC 64bit with D3D Fullscreen and LAVfilters 0.68 (this config gives the lowest CPU utilization possible AFAIK).

Now consider that you will get worse performance in a browser and that Skylake only has maybe twice the IPC of a Core 2 Duo.

Heck, even on my Pentium G3258 at its stock 3.2GHz, using the same lowest-CPU-usage-possible software configuration as the Core 2 Duo PC, it can't handle 3840x2160 VP9 @ 48fps (and not barely either; it also can't do 40fps); it could handle 3840x2160 VP9 @ 30fps though.

EDIT: With my Pentium G3258 @ 4.6GHz, again using the same MPC-HC configuration, I can do 3840x2160 VP9 @ 50fps, but 3840x2160 VP9 @ 60fps is too much.

[BadFrame]

The world’s best VP9 encoder: Eve

Promises to give 5-10% better compression ratio than the official libvpx, and also be 10-20% faster.

Compared to x264 it offers 15-20% better compresion rates, but is ~5x slower.

https://blogs.gnome.org/rbultje/2016...encoder-eve-2/

No word yet as of availability.

[hajj_3]

Originally Posted by BadFrame

The world’s best VP9 encoder: Eve

Promises to give 5-10% better compression ratio than the official libvpx, and also be 10-20% faster.

Compared to x264 it offers 15-20% better compresion rates, but is ~5x slower.

Looks interesting, i shall keep an eye out on this thread for independent comparisons of Eve vs x264 vs x265.

[Selur]

No word yet as of availability.

from the blog entries:

rbultje says:
May 2, 2016 at 3:24 pm

@Braulio: multi-threading is planned, but not yet finished.

@krs: it is currently not available as opensource. We are hoping to test it with customers who are willing to pay for it, so we can build a business around it that can sustain its development and pay the rent etc.

so till the eve encoder is available for public use vp10 and it's successors might be available,..

[BadFrame]

Originally Posted by Selur

from the blog entries:
so till the eve encoder is available for public use vp10 and it's successors might be available,..

Here's hoping they'll find the same business model as x264/x265 viable, which is a fully open source GPL version, with the option for commercial ventures to purchase a license which allow proprietary use.

[Beelzebubu]

Originally Posted by BadFrame

Here's hoping they'll find the same business model as x264/x265 viable, which is a fully open source GPL version, with the option for commercial ventures to purchase a license which allow proprietary use.

I may be stupid, but I just don't see it working. The key issue is that GPL covers redistribution, not use, and therefore most streaming video companies (which is really the primary market where VP9 stands a solid chance of becoming a significant player) will have no incentive to pay at all. After all, they don't redistribute.

Or am I missing something?

[nevcairiel]

Originally Posted by Beelzebubu

I may be stupid, but I just don't see it working. The key issue is that GPL covers redistribution, not use, and therefore most streaming video companies (which is really the primary market where VP9 stands a solid chance of becoming a significant player) will have no incentive to pay at all. After all, they don't redistribute.

Or am I missing something?

This is true, it really only concerns people wanting to ship it as part of their software.