Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Google surpasses TMMI and Dimension with their own upscaler.
https://blog.google/products/google-plus/saving-you-bandwidth-through-machine-learning/
It is called RAISR, is free and open and Google will be using it to save bandwidth. You may already be viewing their upscaled images in your browser. They found the business model and they have the product. No need for anyone else to bother.
There is also a video codec done by another company which Twitter paid 150 million for, but it uses the same technique.
Anyone in the video industry would ask the same questions.
I am not on the defendant's side. For the record, I think their patents are woefully obsolete/useless and hope they both get completely cancelled. That such money has been wasted by both parties arguing over them is nothing short of incredible.
Oh, speaking of patents, I have another question: Does the new codec infringe or possibly infringe on existing patents? It might take a while for everyone involved to sort that out.
Hopefully some questions will be answered, such as:
Does the new algorithm require a longer encoding time? How about decode time? If it still needs a server farm, have the server farm benchmarks changed?
Does it use more memory or other significant resources? What are the minimum platform requirements?
Is it based on fractals, DCT, wavelets, or some combination, or is it something entirely different?
What was Raytheon using? How does the new codec compare against H264 and HEVC?
Does it support different bit depths?
What was the nature of the video test? Because if it was from a stationary video camera filming a largely unchanging scene, any codec could achieve huge compression. Was the video grayscale or color? What was the frame rate and resolution? Does the new algorithm only work on specific types of videos? How well does it handle footage that is changing frequently?
What is meant by "good" quality? What were the PSNR (peak signal-to-noise ratio) measurements?
What is the compression ratio for a single-frame movie (i.e. still image)? Will a still image format also be available?
Does the codec offer fine-grained quality/bandwidth tradeoffs?
If the test results were available back in November, why were they kept secret until now? And why is nothing being said now other than the compression ratio? Is more work needed? If so, how much more?
Why has the mainstream media not picked up the story?
Will there be any third-party validation? Will the public be able to download anything to try it out? Will there be a codec module for ffmpeg?
Why was the new codec mentioned last in the shareholder letter? Should it not have been the lead item?
Will the codec or the company be sold to a third party? If so, then what was all that talk about the long term?
Competitive landscape? Not so sure, everyone is pretty much using one of the H26x codecs these days.
HEVC on mobile
Imagine a video processor that can decode and encode HEVC up to 4K UHD 120 frames per second, and small and efficient enough to go into mobile devices.
ARM is working on it to go with their Cortex CPUs; it is the Mali-V550.
HEVC is off to a great 2015.
Interesting new scaling method
Someone came out with Scalable Function Graphics which describes an image as an equation.
The included pictures on the page compare against bicubic and fractal. Here is one of them.
No optical upscaling for any method, of course, just taking the existing pixels and sharpening them. The reflection in the eyeball is still just a large blob, the iris does not look detailed, the white of the eye has no veins, etc.
HEVC starts going mainstream
nVidia launched its GeForce GTX 960 video card, which on top of being a nice gaming card for only $200, offers up to 4K HEVC decode. The costlier GTX 980 had HEVC earlier, but now mainstream gamers and video watchers can enjoy.
And for those who do not want to add a video card to their system, Intel updated its graphics drivers to add accelerated HEVC decode support to its CPUs. Windows 7 thru 8.1 are supported, in 32- and 64-bit.
Resolution independance?
Fractal interpolation is only a sharpening algorithm. It does not add meaningful detail. If it did, then movie producers would shoot tiny 240 x 135 video and then zoom it 8x or 16x (among other absurdities).
Perfect Resize is one of the few commercial applications that ever successfully shipped using this method. You can download a demo (or buy a licence for a mere $80) and see for yourself what is meant by resolution independance. Compare an upscaled image to an original of the same resolution and you will see the difference. It is suitable for printing situations where one does not stand too close to a large poster or to a billboard, but not for video scaling.
Finally, here is a paper written by none other than Michael Barnsley himself.
Figure 2 is an extremely low-resolution photo of a hot air balloon (it looks to be a 64 x 64 image) which Barnsley then claims to have upscaled and zoomed into figures 3, 4, and 5.
He is quite the "hot air" salesman, and the idea of scaling up to any resolution without losing detail is absolute nonsense.
Speaking of CES...
From a Wired article two days ago:
"The good news is you may not need to upgrade your service to get smooth 4K streaming. Beamr, a post-encoding video-optimization service, is working with movie studios and streaming providers to chop 4K video bitrates by 40 percent. Beamr claims it can whittle 4K video down to a 9.5Mbps stream, making it easier for networks to handle the load. It works with HEVC-encoded video, and the company demoed its bitrate-shaving work at CES. The technology will be used in a new 4K partnership with M-Go."
They were at NAB 2014 too. From the FAQ on the Beamr website:
"HEVC has sophisticated mathematical tools for video encoding, but does not utilize perceptual quality measures in the encoding process. Therefore, our core technology – the quality measure and video encoder controller -- can be applied, with minor adaptations, to HEVC encoding as well. We have demonstrated reduction of 4K HEVC bitrates by 40% during NAB 2014, and plan to integrate HEVC support into the Beamr Video software by Q1 2015."
Beamr does not require people to use a different format. It simply fine tunes the parameters used by the HEVC encoder to squeeze certain areas of the image more. Since HEVC was already very good, now it is even better. It works for H264 too, so 1080p content can be made available to more lower-bandwidth users without changing the existing H264 infrastructure.
It is an interesting reverse approach. Instead of upscaling low-resolution content and trying to add detail, Beamr starts with high-resolution content and removes detail where it will not be missed. It is clever, and it works. The company received $9.5 million in funding in April.
HEVC is up to 40% to 50% smaller than H264, and Beamr compresses HEVC up to 40% more, so that makes it 64% to 70% more compressed than H264.
It might explain why TMM's stock started going south after NAB. With H264/HEVC compression that good, and no need to change infrastructure, who needs TRUDEF?
I concur that HEVC is now too dominant to dislodge in the broader market.
Set-top box distributors are still heavily invested in MPEG-2, and see an easy opportunity to skip H.264 and jump directly to H.265 (HEVC). HEVC-capable set-top boxes are already shipping, and can provide 4K quality at higher bandwidth or 1080p HD quality at lower bandwidth than that of H.264, thus enabling many customers with slower connections to enjoy HD content. Sigma Designs introduced their HEVC set-top box chipset in September 2014. In May, Entropic released their EN7310 SoC supporting HEVC, and forecasts 35 million American homes using their system in 2015.
Akamai is looking forward to using HEVC to significantly optimize their networks. They are not worried by the prolonged uptake of 4K because their own market research shows that users are much more interested in reducing 1080p bandwidth, a job at which HEVC excels. A lot of existing HD content will be reencoded with HEVC, and with all the hardware acceleration and optimized software available, that will not take long.
Blu-Ray Disc Association announced it will use HEVC in 2015, supporting 10-bit color and up to 4K 60 fps quality.
In 2013, MPEG-DASH became available for HEVC. DASH is short for Dynamic Adaptive Streaming over HTTP, which lets a browser or other client application switch on-the-fly to a different resolution of a video stream if bandwidth availability changes, keeping playback smooth on rough networks.
Elemental, a key HEVC infrastructure developer, secured $14.5 million in series D financing from Telstra, Australia's largest telco, and from Sky, Europe's largest entertainment company. Telstra is looking to use HEVC to enhance video delivery on its mobile networks.
In March 2013, NTT Docomo of Japan began licencing their HEVC decoder to mobile device vendors. Samsung implemented HEVC hardware into their Galaxy S4 phone. Apple has HEVC decode (and 720p encode) in the iPhone 6. Mobile vendors are excited because video is the largest bandwidth use and HEVC will let them expand easier.
On December 19, 2014, Google released the official version of Android 5.0 "Lollipop" which includes APIs for HEVC. Three smartphone brands (HTC, Moto, and Nexus) already ship with it, and an update is available for the Samsung Galaxy S5, which can both encode and decode up to 4K 30 fps. Google has its own VP9 codec but went with HEVC anyway (if you cannot beat them, join them).
This is probably the tipping point. Android outsells all other operating systems combined with over a billion active users, Windows 10 directly supports HEVC, and Apple will be forced to increase HEVC support in iOS and in OS X to keep up. With the disc player, set-top box, desktop, and mobile industries this heavily on board, other codecs will be resigned to niches. With all the pieces now in place, HEVC is solidly on track to expand and succeed H.264.
The competition seems to be moving onward and upward.
The nice people at DivX have HEVC encoding of HD video in realtime
using a software-only approach on an Intel Core i5 CPU.
Amazon's Prime video streaming service now supports 4K using HEVC.
Fabrice Bellard created BPG, a new still image format based on HEVC that compresses better than anything else. Here it is against JPEG, and here is an interactive comparison viewer. He has an open source library available (encode/decode) and even a Javascript version of the decoder.
Samsung ignores TRUDEF
In October, Samsung released their NX1 camera. It can do 24 fps Cinema 4K and 30 fps UHD, encoding HEVC, all in the compact portable body of a DSLR.
The small size and power efficiency of such a h265-capable GPU is noteworthy -- realtime 4K capture, no server farm. One could use a small array of these to capture 8K video or more, and easily fit them inside a drone.
Microsoft announced that Windows 10 will natively support HEVC (in Windows Media Player, etc.).
As endorsements for a codec go, it does not get much better than that.
Speaking of security video applications, the nice people at
Ambarella have released a 4K security camera that encodes HEVC, using a SoC (system on a chip).
You know, thinking about these big companies, there is something odd.
Even though Raytheon is working with TMM, no one else has bothered to buy TMM or Dimension. Apple, Google, Microsoft... it would cost them little to do so. Google is desperate for a codec to compete against HEVC but has not bothered with TRUDEF.
Also, none of them care that a competitor might swoop in and pick up the technology. They do not even care that Raytheon might gain an advantage.
Microsoft and Google? Okay.
Microsoft has been supporting HEVC in the Windows DirectX SDK since August 2013. Guess they did not feel that TRUDEF was worth including or investing time in.
Google has a huge server cluster, yes. Part of it is used to host YouTube. Which they are busy upgrading to use their own codec, VP9. And instead of bothering with TRUDEF, they are hard at work on their next codec, VP10.
Anyone setting up a server farm has the option to equip each machine with hardware accelerated HEVC, which makes each server the equivalent of many TRUDEF servers. And a single machine can encode fast enough to support realtime broadcasting, something no amount of machines can do for TRUDEF. And all of these capabilities are available now. They are not being researched, they are being delivered.
No business would spend significantly more money to run multiple servers to do the same job as a single PC. It does not make sense.
Also, virtual servers would not work for such a compute-intensive process as video encoding. Two virtual servers would run half as fast, defeating the purpose.
Speaking of large businesses and other enterprise clients, Intel last month released their Media Server Studio 2015, which supports optimized HEVC encode/decode for their Xeon CPUs.
Intel, Apple, nVidia... three big vendors that passed over TRUDEF and are shipping HEVC product today.
nVidia recently released their GTX 980 video card, based on the Maxwell 2.0 architecture which includes a hardware accelerated HEVC decoder/encoder.
Who needs a server farm when a single video card can encode H.265 in realtime.
Apple is using H265 in its Facetime feature on the new iPhone 6, cutting in half the cellular data bandwidth required. They were able to include a realtime encoder.
So how are things? Oh, not so good. Someone said it would be at a dollar by Labor Day. Oh well. I guess I will check back in another few months. Maybe Santa will be kind.
The problem here is that all the low-hanging fruit has been picked. People (very smart people) have tried for years to make fractal image compression practical. Now I am not saying that Raytheon might not come up with something (maybe something niche), but it is probably going to take a long time. They are likely poring over every bit of research that their peers have done over the last twenty-five years, and confirming what they all learned: images do not have enough self-similarity.
For upscaling, there are many good, fast (and free) algorithms already available. If they want to upscale video, it would be better to use them and just compress the results with H264/265. Raytheon probably already does this, they are no strangers to signal processing.
Why do fractals appeal so much? Because of our own human bias. People look at an image, and because with their eyes they can quickly identify all the parts of an image that are similar, they think the problem is solved. Show an artist a landscape picture, and he immediately identifies sky, grass, and trees. But ask him to paint the same picture from memory, and the blades of grass will be different, the pattern of clouds not really the same, the tree foliage will be sparser or denser or the lighting will be different. In exchange for his mind's ability to hugely compress what he saw, he did not encode any details and therefore must synthesize entirely new ones. He forgets to draw a bird's nest, or draws it in a different tree.
This is not what people want a codec to do. They want all the details preserved. They want their input to match the output (especially for surveillance applications). And this preservation of detail is what frustrates the fractal block-matching system. Even a tiny block has an incredible number of possible color arrangements (e.g. a 2 x 2 grayscale block has four billion combinations). Block matching is great when blocks match exactly, but when they do not (and they almost always do not), one must discard detail or expend valuable space to encode it. This is a limiting factor that was known very early on, and Raytheon is doubtless aware of it, because it is the first problem one runs into when doing block matching.
Science is good. :)
Anyway, nothing to do anymore but wait. Have a good summer everyone, and see you in September.
Unless Iterated was hiring very young, I could not have been there. Anyway, it was my boss who had the software. His boss was evaluating a licencing deal several years back and it did not go anywhere. As for stolen, what could we do with a binary executable? Iterated did not provide source code.
Have you worked with VDK 1-1.4
Yes I was using 1.4.
And to Jay, the graphic comparison is only to show the intrinsic limitations of upscaling. There are some good methods —Trudef may even be one of them — but they should not be considered a way to improve compression. Sure, one has one-fourth the data to encode but the quality is nowhere as good as the original image.
For reference, I made a comparison of upscaled versus original high-resolution content:
http://paulschlessinger.wordpress.com/2014/06/04/upscaling-versus-high-res-content
One of the fundamental problems with the VDK is that it performs two image processing passes, one to upscale and the other to compress. Both passes use block copying, which introduces errors, and the errors from the upscaler are compounded by the compressor.
This might be why Dimension went for upscaling only, they figured that dropping the compression feature would be an easy way to eliminate a large error source. In TMM's case, they need to do both passes together with a single algorithm (or use an upscaler of better quality), greatly increase the bandwidth utilization, and have it still work fast enough for realtime decoding.
It is a hard job. Block copying is decode-friendly, but quality is poor unless blocks are split. But block splitting hurts the compression ratio. It is damned if you do and damned if you do not. There is just not enough compression at an acceptable quality level. Whatever the work with Raytheon accomplishes, I suspect it will include a sizeable departure from the block copying methodology, so we could be waiting a while.
Improved simple upscaler to give a better idea of what block copying upscaling looks like.
http://paulschlessinger.wordpress.com/2014/06/03/a-simple-upscaler-improved
Again with the VDK 2.0 nonsense. I never used that version. How could I have, since my boss had only a normal PC.
And if VDK 1.4 was better, then why would Iterated have stopped using it too? It was also obsolete.
TMM and Raytheon need to work together to make Trudef work together with Raytheon's software
That is called integration, not research.
Did you ever think that just maybe Raytheon has already determined whether it "works or not" before allowing TMMI to make a news release about their involvement.
Well, that is an interesting question, and yes, I did think about it.
If TMM had Trudef working enough to interest Raytheon, then why bother soliciting help? Just commercialize it instead. In fact, if it were good enough, why would any more research be needed at all? Why not give public demonstrations that would amaze everyone and finally silence the critics?
No. The only logical answer is that TMM could not make it work.
The new management probably did not realize that the VDK was worthless, but enough time has passed that, through their own anaylsis, they have become aware. So now they have turned to Raytheon to see if they can make anything out of it.
On a side note, Dr. Sloan has much to answer for. He is both an original inventor and on the board of advisors, but apparently did not contribute enough to make Raytheon's involvement unnecessary. Why has he let TMM waste so much time? If the technology can be made to work, surely he would have known how.
Raytheon is probably just covering their bases. They maintain research in imaging and to do a complete study they realized they could include Trudef for a small (to them) cost.
They will assess and analyze fractal imaging top to bottom, and then reaffirm what everyone else has already known, which is that the negatives outweigh the positives. In research, it is valuable even to firmly identify dead ends.
On a positive note, maybe Raytheon will get to the truth about Trudef and tell everyone whether it works or not.
That being said, TMM has gone "all in." If Raytheon does not find it useful, then it is game over.
So TMM had Trudef running on Windows, Android, and Surface, and even built a server farm. They must have been confident it could go to market soon. But now they are back in research mode and cannot even do the research by themselves. And the applications they talk of are not the broader codec market.
They might carve out a small niche, but Raytheon will also own part of it. So TMM winds up with part of a part. I am not surprised investors are depressed. This is not how it was supposed to unfold.
An upscaling test
I wrote a simple upscaler using a method similar to the one in the Dimension patent and posted the result here:
http://paulschlessinger.wordpress.com/2014/05/24/a-simple-upscaler
It does not do artifact filtering, but helps give an idea of what the edge sharpening would look like.
You give me too much credit. Those promoting the stock outnumber me and are voicing their opinions loudly and frequently. That the stock fails to rise, says more about them.
Well, on that you and I are in agreement. TMM has genuine trouble with their press releases.
I am only offering a reasonable analysis using the available data. Given the recent quick reversal of yesterday's rise, I am not the only one here with a conservative outlook.
As for MPEG LA, they are doing very well without my help.
Why do you keep mentioning VDK 2.0? I did not say I had used that version.
Blog update:
"Resolution independance" and "Unsuitability of fractal algorithms for surveillance applications"
paulschlessinger.wordpress.com
Has Dimension shifted to just upscaling?
It sounds like it from their press release about NAB.