Google VP10 video codec

[Jarek]

Finally some details about the progress of development of VP10:
https://www.youtube.com/watch?v=LzKP7JJjXtY

Some improvements over VP9 (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.

[Jyrki Alakuijala]

10/12 bits (also artificially to improve prediction quality: 2-4% gain for 8-bit content),

I wonder how many other systems can benefit from a similar optimization.

[thorfdbg]

I wonder how many other systems can benefit from a similar optimization.

JPEG does, in a sense. If you look at the JPEG XT reference implementation - which is also a JPEG implementation by nature - then one of things it does is to "artificially expand" the bit depths by four bits. Or rather, to compute in fixpoint. That is, the data that comes out of the YCbCr transformation and goes into the DCT is pre-shifted by four bits. Similarly, the DCT is pre-shifted accordingly, and the quantizer is adapted similarly to remove the additional bits. This gives a small PSNR advantage, especially for low compression ratios or high bit depths.

[xcrh]

Finally some details? Hmm, VP10 commits in libvpx were here for more than year or so and git version of libvpx could be built with VP10 for a while (it have to be explicitly requested in configure though), so "finally some details" sounds rather strange. I've even got some VP10-encoded movies for like a few months. Yet there is little issue: VP10 isn't declared stable yet, it changes, so it does not builds by default and virtually no programs on the planet readily supports decoding VP10 at this point.

So IMHO real question should sound like this: has VP10 been stabilized? Is this list of features final? Is bitstream format frozen or not at this point? I also wonder what they are up to as Open Media Alliance member, it got some other fancy codecs like Daala on board, and it surely got some tricks to offer, like their "lapsed tansform", which is supposed to produce far less annoying compression artifacts.

Oh, btw congrats on ANS getting to it, I guess it can easily be like a first video codec on the planet using it and that's where it should rock, video needs decode speed any day (in google's shoes I would try hard to hire Jarek, btw XD).

[Gonzalo]

How does it compare to HEVC - h.265?

[clare]

Hmm, VP10 commits in libvpx were here for more than year or so and git version of libvpx could be built with VP10 for a while

Coding for VP10 is happening in the branch nextgenv2, not master.
It's very active:
https://chromium.googlesource.com/we...+log/nextgenv2

[Jarek]

Some updates:
https://en.wikipedia.org/wiki/VP9#VP10
https://en.wikipedia.org/wiki/Alliance_for_Open_Media
http://www.streamingmedia.com/Articl...l/-110383.aspx
http://www.streamingmedia.com/Articl...l/-111497.aspx
http://www.eetimes.com/document.asp?doc_id=1330023
AOM was founded 1st September 2015 by Amazon, Cisco, Google, Intel, Microsoft, Mozilla and Netflix, this year AMD, ARM, NVIDIA, Adobe, Ateme, Ittiam and Vidyo had joined.

The last link says that it currently has ~23% reduction comparing to VP9 and:
"We are still very much hoping to get better compression on the order of 40% at reasonable complexity by the time we are ready to release AV1 [a royalty-free codec using VP-10]," said Mukherjee in a follow up email.

So it seems VP9 was the last one: VP10 will not be released alone, only as AV1 before March 2017 - with additions from Mozilla Daala (lapped transform?) and Cisco Thor (transformations of large blocks?).

[dado023]

its not VP10, but still interesting results....taken from phoronix:

https://i.imgur.com/UwPvSFJ.png