V-Nova Perseus video [s/codec/noise modeller] announced

[LigH]

It is always again amazing when groundbreaking technologies are announced around the 1st of April; this time, a new consortium named V-Nova (http://www.v-nova.com/en/index.html) claims to be twice as efficient as HEVC while fast enough for realtime HD compression. Does this have any relation to NGIV, or is it just another hoax?

[madanmohan]

It is always again amazing when groundbreaking technologies are announced around the 1st of April; this time, a new consortium named V-Nova (http://www.v-nova.com/en/index.html) claims to be twice as efficient as HEVC while fast enough for realtime HD compression. Does this have any relation to NGIV, or is it just another hoax?

Announcement came on April 1st but I hope its real. :)

[kypec]

Announcement came on April 1st but I hope its real. :)
Yeah, the announcement is real, that's for sure. Unfortunately, the claims and promises about all superiority of their codec in that announcement are not. :D

[madanmohan]

Is there a possibility to get funding for independent research projects like NGIV?

Yeah, the announcement is real, that's for sure. Unfortunately, the claims and promises about all superiority of their codec in that announcement are not. :D

oh, that's unfortunate.

[Skarstorm]

In there Full Press Release (PDF) (http://www.v-nova.com/en/press/2015-04-01/V-Nova-Launch-Press-Release(Wed-01-04-15).pdf) they have endorsements of Broadcom and EBU.
Well this makes me realy look forwards to the NAB Show 2015 (Las Vegas, April 13-16),
for now we can only wait untill the other shoe drops.

[benwaggoner]

In there Full Press Release (PDF) (http://www.v-nova.com/en/press/2015-04-01/V-Nova-Launch-Press-Release(Wed-01-04-15).pdf) they have endorsements of Broadcom and EBU.
Well this makes me realy look forwards to the NAB Show 2015 (Las Vegas, April 13-16),
for now we can only wait untill the other shoe drops.
Not to get things on-topic, but I just got back from five days at NAB. I didn't see or hear anything about Nova there, nor NGIV (although that gets a pass based on not having made extraordinary claims).

Did anyone learn or see anything? The codec buzz and demos at the show were all HEVC for new stuff. Very little MPEG-2 compared to past years. Tons of AVC, of course, but that wasn't what was getting really highlighted.

[Skarstorm]

Well @ the VideoHelp forum (http://forum.videohelp.com/threads/371115-V-Nova-a-new-video-codec-50-more-efficient-than-hevc?p=2385674&viewfull=1#post2385674) there is photographic evidence of their presence.

[benwaggoner]

Well @ the VideoHelp forum (http://forum.videohelp.com/threads/371115-V-Nova-a-new-video-codec-50-more-efficient-than-hevc?p=2385674&viewfull=1#post2385674) there is photographic evidence of their presence.


Thanks. I know some of those guys, including one I've worked with quite closely and fruitfully with in recent years. I'm taking this a lot more seriously now. The use of non-standard codecs in backhaul certainly has happened before. And VisualOn already supports software decoders, so adding a new one could make sense for them.


-Ben Waggoner (via TapaTalk)

[LigH]

Remember DivX testing their new codecs with the community here. Even though Atéme could not compete with x264 after all, discussing flaws and possible improvements here improved their code to some extent. I hope the V-Nova developers won't be afraid to share a codec for discussion and improvements as well, may it be as basic as Daala's, possibly with a date killswitch... Here at doom9 they will find people with technical background, able to detect flaws and to provide suggestions.

[kuchikirukia]

Can there be any more buzzwords?

[benwaggoner]

Can there be any more buzzwords?
Always. The press release didn't even use "disruptive" :).

No one ever ran with my slogan for VC-1:
"Adaptive Deadzone - it's not a threat, it's a promise."

[CruNcher]

“We don’t touch lots of the standards. We only touch one very tiny thing, which is the elemental stream and some of the transforms. The standardization worries are misplaced. It works, it is inexpensive, so where is the problem and why is there a need to rubber-stamp PERSEUS with a standardization body?"

He also emphasizes that in the real world, PERSEUS is not an alternative codec to H.264 or HEVC but will be used with encoders that also have these compression capabilities. So the expectation is that it will be offered to platform operators by encoder vendors who license PERSEUS as an option. Meardi emphasizes that this new codec is not going to create a fork in the road in the current upgrade process, which until now has been a straight and clear route from H.264 to HEVC encoding.

Sounds like requantization in some different HVS way ?

[vivan]

Sounds like psy.

[Blue_MiSfit]

I haven't seen anything real yet.

[CruNcher]

On their team is at least one of the Engineers of the Super-FX GPU (Argonaut Games London)

Luca Rossato is the Chief scientist worked on MPEG mainly for Telecom Italia (he is credited with 25 years of MPEG experience)

Guido Meardi seems the Mckinsey Super Econimics Brain hes also mentioned as Inventor in the Patents application though, he doesn't speak like a total IT noob about Perseus ;)

Before this he was mainly active in Automated Lean Production optimization

Yeah it sounds something they could have contrubuted to the standardization process of H.265 but they didn't VC-1 all over again only that it's not a whole codec but more a subset it seems.

[roo1234]

I read that Airbus has licensed it to transmit HD video over 3G...

[LigH]

I wouldn't wish anyone to discover being a hype victim...

[Gravitator]

I wouldn't wish anyone to discover being a hype victim...
Do not make yourself a victim prematurely...
Especially since we see their faces. ;)

[zambelli]

I saw their public demos at NAB and they were impressive, but they've been very secretive about giving private demos or hands-on access to codec tools. Their leadership and engineering roster seems legit, but the way they're going on about it definitely screams "vaporware".

[Gravitator]

I saw their public demos at NAB and they were impressive

Hi! What impressed?

[zambelli]

High quality at extremely low bitrates. They were showing SD at a few hundred kbps, and 720p at 500 kbps, both looking very good. Of course it's one thing to show a demo of carefully selected and compressed content, and a completely different thing to demonstrate it consistently in the field... But even if their codec Perseus is half as good as they claim, it'd still be a strong competitor to HEVC and VP9. Here are their claims: http://www.studiodaily.com/2015/04/v-nova-claims-perseus-codec-twice-efficient-h-264-hevc/

And if their codec really is as good as they claim... Wow, that would completely blow HEVC/VP9 out of the water. I won't believe it though until somebody unbiased has had a chance to play with an actual Perseus encoder.

[CruNcher]

Not sure they made relatively clear this is not something like a full blown codec but builds into existing codecs actually MPEG

http://forum.doom9.org/showpost.php?p=1730162

The innovative and unique approaches described herein allow separation of the signal into two different sets of information (such as a core signal and transient layer signal), and then encoding and decoding the two sets of information independently. In this way, very different encoding/decoding approaches can be applied to the core signal, which requires a reconstruction scheme with strict adherence to the original signal, and to the transient layer signal, which typically requires a reconstruction scheme with just the same stochastic properties, but not necessarily identical to the transient layer contained in the original signal.

In addition, in case of network congestion, embodiments described herein allow reduction of bitrates required to transmit a rendition of the signal by independently modulating the necessary bitrates for the core signal and for the transient layer signal. For example, higher priority can be given for transmitting the core signal, and thus allowing smoother quality degradation compared to approaches where encoding of signal does not involve distinguishing the core signal component from the transient layer signal.

One embodiment herein includes a method for adjusting a first decoded rendition of a signal by combining it with information reconstructed based at least in part on received parameters corresponding to statistical properties of the signal, thus increasing the perceived quality of said decoded rendition.


Sounds partly like FGM

Something that gathers specific Signal information and sends it to the decoder for reconstruction Core Signal surely means the Main Bitstream and then something somewhere saved in some layer (either bitstream or on Container level) that is read by their Adapted Decoder for reconstruction.

here they get a little bit more specific ;)

More specifically, in a non-limiting example embodiment, a signal processor configured as a decoder receives two sets of reconstruction data. The first set of reconstruction data is decoded according to a first decoding method, reconstructing a first rendition of the signal ("core signal layer"). The second set of reconstruction data is decoded according to a second decoding method (different from the first decoding method), reconstructing a second rendition of the signal ("transient layer"). The decoder then combines the two renditions, obtaining a final rendition of the signal. In some non-limiting embodiments, said second set of reconstruction data comprises at least one parameter corresponding to statistical properties of the signal. In other non- limiting embodiments, said second rendition of the signal is reconstructed also based at least in part on the quantization parameters used to decode said first rendition of the signal, effectively randomizing quantization errors via dithering.

In other non-limiting embodiments, the decoder reconstructs said second rendition of the signal according to a method that simulates the generation of values according to random generation of numbers, while leveraging information that is known at both encoder side and decoder side (i.e., allowing the encoder to precisely simulate the results of the reconstruction of the second rendition of the signal at the decoder side). In a further non-limiting embodiment, the decoder simulates the random generation of numbers based on one or more reference tables of numbers, selecting a starting position in the table according to parameters that are available at both the encoding and the decoding side.

According to embodiments herein, depending on parameters such as the availability of bandwidth, or the target fidelity of the transmitted rendition, a signal processor configured as a streaming server can choose whether to transmit the transient layer with precise fidelity (at the cost of a higher amount of transmitted information) or just with statistical fidelity (i.e., reducing the adherence of the reconstruction with the original signal, but also reducing the amount of information to be transmitted). Thus, if bandwidth is available, embodiments herein can include transmitting the original signal including the original noise for restitution. Transmitting a signal including the original noise can require substantial bandwidth, even though the exact noise is not critical to the restitution. When less bandwidth is available to transmit data to a target recipient, the substitute noise information (i.e., having statistical characteristics of the original noise) can be transmitted to the target recipient in lieu of transmitting the original noise that requires substantial extra bandwidth.

In one non-limiting example embodiment, statistical characteristics of noise identified in a signal can be identified and encoded as substitute noise according to a tier-based hierarchical encoded method.

I would be really surprised if this holds stand against a Patent Attack by Thomson :D

In accordance with another embodiment, the encoder 110 can receive signal. The encoder 110 can be configured to parse the signal 100 into multiple components including a noise signal component (such as signal 130-2) and a non-noise signal component (such as core signal 130-1). The encoder encodes the noise signal component independently of encoding the non-noise signal component and stores the encoded non-noise signal component and the encoded noise signal component in one or more repositories.

Definitely Noise (Recording) analyzing and reconstruction on the Decoder Side which also means a efficient Denoiser is implemented ;)
So Recording the Noise saving statistical data about it Denoising and when needed for the input reconstructing it on the Decoder Side voila Perseus ;)
Or voila FGM ;)

And using Quantization Analysis @ the Decoding Input stage ;)

Input->Cleaning Process and Noise Recording based on Input Quantization Data->Dithering->Saving of statistical Noise Data->Sending->Reconstructing if needed by the Decoder

Maybe the Denoising and Reconstructing is GPU accelerated thus why they have a GPU Expert on board ;)

Found the Figure :)

https://patentscope.wipo.int/search/docservice_fpimage/WOIB2013050486@@@false@@@en;jsessionid=CDE23A09398C222864124C71058B7317.wapp2nC

So pretty much what some do since years here with their inputs except the Noise reconstruction from statistical recorded data gathered @ the Denoising stage ;)

What Thomson actually proposed with FGM for H.264

[CruNcher]

Perseus Implementation Tests (720p) into X264 @ V-Nova HQ

http://i1.sendpic.org/t/vl/vlipU14o4M2FrTDVdn5dHAyeYsg.jpg (http://sendpic.org/view/1/i/eMqjB5SaFcYlUngZ0t5MrsMrtG7.png)

http://i1.sendpic.org/t/bN/bNyBAeSSnj8YNul6Dp5fFJUQiJ1.jpg (http://sendpic.org/view/1/i/5cTAApiZM4TY1M4mF8cGqpy94Rp.png)

http://i2.sendpic.org/t/ww/wwd4TnGcTIbu3iSg8VRzOsQBuwf.jpg (http://sendpic.org/view/2/i/zEOUOBSGReRziWA1TD8iSQ0e52Z.png)

[smegolas]

http://snackbox.org/~snacky/ds_quote.txt

[CruNcher]

Surely EyeIO doesn't have filed so many patents about improving MPEG parts upto H.265 then V-Nova did ;)

[LigH]

So they created a custom x264 build using Perseus as noise modeller, like lowpass filtering the video to be compressed as AVC, and the output interleaving AVC and the highpass metadata?

I wish you had shared a link to your source. So I see only very interpretable screenshots, not knowing any certain context.

[Atak_Snajpera]

is this suppose to work like mp3pro but for video?

[LigH]

According to these quotes in the VideoHelp thread (http://forum.videohelp.com/threads/371115-V-Nova-a-new-video-codec-50-more-efficient-than-hevc?p=2411906&viewfull=1#post2411906), we would assume so: similar to SBR used in mp3PRO and HE-AAC.

Please try hard to ignore most of the trolling there; Stears555 is already on several ignore lists.

[Kurtnoise]

SBR is not a lowpass filter...

[LigH]

Not exactly, true; rather indirectly: Halving the sample frequency of audio divides the spectrum in a low frequency part which is encoded conventionally, and a high frequency part which is modelled in relation to the low frequency part.

What would that mean for videos? Encoding half the resolution = a fourth of the area conventionally, and modelling the details which were lost by the downsampling (which will have calculated an average, which is similar to blurring, the low-pass of image processing). That reminds me a lot on "Wavelet decomposition", using just one decomposition step.

http://i.stack.imgur.com/gCbV6.gif

We should move this discussion from here to there (http://forum.doom9.org/showthread.php?t=172044), it's disturbing the "x264 development". :o

[birdie]

Seems to be real:

V-Nova’s PERSEUS codec passes IRT’s initial video quality-checks

V-Nova, originator of the high-performance PERSEUS video codec, says it has passed initial video quality tests by Germany’s IRT (Institut für Rundfunktechnik) for HD TV contribution applications.

The IRT – the respected principal research arm of German-language public service broadcasters – evaluated PERSEUS against JPEG2000 contribution and production codecs benchmarks at comparable bit-rates.

The results, based on a mixture of objective and subjective evaluations, confirm that PERSEUS offers a substantial improvement compared to JPEG2000 and a quality that challenges H.264-intra-based production codecs such as AVC-Intra (these carry out video compression processing within individual frames, rather than ‘inter’ schemes, where the processing is completed over multiple frames).

The tests were performed on HD interlaced content at 1080i/25.

IRT said that PERSEUS provided “commercially valuable increases in efficiency over contribution compression codes AVC-Intra and JPEG2000 at the most widely adopted HD format today, 1080i.” It added that, due to the hierarchical nature of the PERSEUS codec, the benefits would be even larger at higher resolutions such as UHD/4K and higher frame rates such as 1080p50/60.

Eric Achtmann, V-Nova Executive Chairman and co-Founder, said that proving the technology with the most respected broadcasting organisations was “fundamental in moving PERSEUS from its award-winning NAB 2015 debut to an established commercial “reality” at IBC.”

[birdie]

From the look of it ( https://www.youtube.com/watch?v=vbEQx-dJG0k&t=1m8s ) these guys are heavily reliant on x264 ;-)

A nice screenshot (https://www.imageupload.co.uk/images/2015/10/06/perseus.jpg) ;-)

[dapperdan]

I don't know if this is just a translation problem, but these two sentences seem to have opposite meanings:

1. PERSEUS offers a substantial improvement compared to JPEG2000 and a quality that challenges H.264-intra-based production codecs such as AVC-Intra

2. increases in efficiency over contribution compression codes AVC-Intra

I would have read "quality that challenges" as meaning, "almost but not quite as good as", yet the second sentence claims it's better. Does x264 have an intra only mode? (edit: wikipedia claims x264 has an AVC-Intra mode) And does it beat whatever AVC-Intra encoder was used in this test?

[kolak]

Is this from some PR text? Than forget about it. I would rather hear what IRT really thinks :)

[CruNcher]

From the look of it ( https://www.youtube.com/watch?v=vbEQx-dJG0k&t=1m8s ) these guys are heavily reliant on x264 ;-)

A nice screenshot (https://www.imageupload.co.uk/images/2015/10/06/perseus.jpg) ;-)

Nope they rather chose the most efficient Open Source Encoder out there to test their MPEG Agnostic implementation improvements.

Of course also to deliver the updates where x264 is in use on the encoder/decoder side.

You could pretty much implement this in any MPEG encoder that is their goal, fast easy deployment in the current MPEG ecosystem without extreme hardware change.

Non interoperable Decoders would just ignore the improvement layer.


We can just wonder why this wasn't part of the actual Standardization Process in H.265 and done by another consortium around MPEG members in the hidden.




For the HEVC Advanced Licensing you now get a nice boost up the overall calculation is though still important but with V-Nova Perseus as a enhancer in Bandwidth it becomes something you should recalculate (especially as all this stuff was done under Mckinsey surveillance).

Not sure how hard the hit will be for AFOMC but it's something to calculate pushing development cycles so far away because thinking HEVC Advanced wont be accepted shines in a new light now.

Especially when we should see a merging of Perseus into HEVC Advanced Licensing.

For me this is nothing that happens without a reason but very carefully planned out and keeping this secret is hitting hard.

Though since mp3PRO/AAC-HE FGM and SVC it could have been expected to happen from someone, that implements it more efficient and brings everything together in a efficient hierarchical way (with enough resources @ hand), this is what the V-Nova part worked on in the hidden for almost 5 years.



Any Codec now has to measure it's performance vs the for now unofficial HEVC-HE we seeing rising up here

I hope it's not to late :(

But seeing how fast this is spreading :(

i totally underestimated their capabilities :(

[LigH]

We should move this discussion from here to there (http://forum.doom9.org/showthread.php?t=172044), it's disturbing the "x264 development". :o

:o We are getting more and more off-topic...

[CruNcher]

We are already there ;)

[LigH]

Thank you, moderator, for moving. :)

Now I wonder, has anyone of you a little insight in noise modelling and frequency spectrum parametrization techniques, and some well grounded knowledge how efficient parametric detail restoration can be in visual techniques? Shall we expect it as convenient as known from audio formats? If my analogy of the decomposition is correct, then the 2D nature of video may even have an advantage that the lower frequency base has only a fourth of the area, so 3/4 of high frequencies are a lot of overhead to reduce by parametrization. But I am aware that the perception differs between audio and video signals.

[CruNcher]

They combining SBR + FGM that should be pretty efficient Visually, no real surprise that Thomson was one of their first users there ;)

It's a complete revival of it but as a base for it not like in H.264 a extra profile that practically no one is really using

[ndjamena]

If I'm reading this correctly, Perseus is just an advance form of ascii art that's applied on top of a low bitrate video stream...

Would anyone like to correct me?

[Motenai Yoda]

How it is different from the old film grain tecnology?
https://hopa.memberclicks.net/assets/documents/thomson_fgt_hpa2006.PDF

also why don't take the residual noise and model for the whole frame/gop a related high comprimible noise pattern and add it right after mdct and mv steps?

[foxyshadis]

It's different in the sense that someone is actually implementing it. For all anyone knows, maybe it is just FGM, but FGM is unnecessarily complex for most use-cases.

[dapperdan]

A sort-of review of the codec/noise modeller:

http://www.streamingmedia.com/Articles/ReadArticle.aspx?ArticleID=111816&PageNum=1

They maintain an upbeat tone, but I think the data suggests that it's basically marketing BS? Certainly not living up to the early marketing. There's *something* there, but it seems ultra-niche and it's not clear to me that it's even much of a win in that niche.

[kieranrk]

Here is a technical analysis:

http://codecs.multimedia.cx/?p=1136

[Gravitator]

https://www.v-nova.com/wp-content/uploads/v-nova-perseus-plus-advanced-video-compression-technology.jpg https://www.v-nova.com/wp-content/uploads/v-nova-perseus-plus-vs-x264.jpg
https://www.v-nova.com/perseus-video-compression-technology/ (https://www.v-nova.com/perseus-video-compression-technology/)

[LigH]

https://1.bp.blogspot.com/-zjpvmP_kaaI/VGUpYPyLQiI/AAAAAAAAcFA/dMbRC4eKRJw/s1600/right%2Bback.png

Stills are useless. Marketing optimized stills are even more useless. Give us the tool and let us test.

[sneaker_ger]

We've seen a similar picture of x265 vs x264 in the past. Happens with low bits/pixel. If you downscale the x264 encode will be much better (than low bpp x264). And if I'm not mistaken that's a core component of Perseus, i.e. downscale, encode via e.g. x264 and on playback decode H.264, upscale plus some postprocessing (with info from extra layer?).

[LigH]

Yes, quite probably similar to mp3Pro or HE-AAC (low pass + spectral band modelling), just in the visual domain.

[Motenai Yoda]

looks more like how progressive jpeg works, a low quality/resolution layer at the base, and one or more refinement layers at higher quality/resolution witch supply residual details (difference) into a space domain (maybe using the same mv analysis of the underlying stream, rather than an sbr or fgt.
Also this way they can stream out only the requested layers or only the base layer, and change how this base is encoded.

[LigH]

True so far; but the HQ layer may store easier-to-compress patterns instead of the true detail; that's the reason to call it "modelling".