Cinegear 2019 News and Rumors

[jcs]

For right now, i agree. That is 6k raw, and not actual real 6K detail, i assume? As what we have been calling 4k the last few years is probably not 4K

Right, to get real 4096 we need 4096/.66 = ~6206 Bayer pixels, and for 3840 we need 3840/.66 = ~5818 Bayer pixels with proper filtered downsampling vs. line/pixel skipping. This is to achieve real post-deBayer detail along with minimal aliasing (we also need a good optical low pass filter on the sensor to cut frequencies above the sensor resolution per Nyquist).

This is why the latest ~6K cameras producing 4K look so much more detailed, without aliasing, vs. say the C300 II and C200 (and any other camera with a 4K sensor producing 4K output), which look much less detailed/softer and also alias on fine fabric etc. With a properly tuned OLPF, the 4096 sensor cameras can output 2.7K without aliasing. Comparing e.g the 6K Sony A7 III and Fuji X-T3 to the 4K sensor cameras shows this quite well.

 

[Cary Knoop]

Right, to get real 4096 we need 4096/.66 = ~6206 Bayer pixels, and for 3840 we need 3840/.66 = ~5818 Bayer pixels with proper filtered downsampling vs. line/pixel skipping.

Well, yes and no, it really depends.

6K is obviously better than 4K, no argument here, if anything there is more room to adjust in post.

But practically speaking if the deliverable is 4:2:0 or even 4:2:2 then a 4K Bayer sensor pretty much produces the same visible resolution as a 6K Bayer sensor if the target resolution is 4K.
If the deliverable is 4:4:4 then there will only be visual differences for things like texts and VFX elements but not for natural imagery recorded through a lens.
Our eyes are simply far more interested in luminance resolution than color resolution.

 

[jcs]

Well, yes and no, it really depends.

6K is obviously better than 4K, no argument here, if anything there is more room to adjust in post.

But practically speaking if the deliverable is 4:2:0 or even 4:2:2 then a 4K Bayer sensor pretty much produces the same visible resolution as a 6K Bayer sensor if the target resolution is 4K.
If the deliverable is 4:4:4 then there will only be visual differences for things like texts and VFX elements but not for natural imagery recorded through a lens.
Our eyes are simply far more interested in luminance resolution than color resolution.

We discussed this at length, with math (Nyquist Sampling Theorem), computer image simulations, and example footage from real cameras: http://www.dvxuser.com/V6/showthread.php?356356-Actual-Camera-Resolution-Before-Aliasing

A sensor can provide ~2/3 the Bayer resolution before aliasing. Looking at it from the other direction, we need 3/2 or 1.5x oversampling:

2 pixels = at the Nyquist limit is aliasing since we need > 2x sampling:

Adding an extra pixel (3 total now) for > 2x = no aliasing:

Pixels are integers: we can only use 1 (vs. .5 etc.) extra pixel for sampling, thus we get 3/2 oversampling. As the image above shows, there's no way around this: Nyquist Sampling Theory- we've got to have > 2x (vs. = 2x) sampling.

This is for luminance, so 420, 422, 444 doesn't matter. With low(er) quality in-camera deBayer, there can be color artifacts which will be improved with post deBayer (before color subsampling).

 

[DLD]

https://www.instagram.com/p/BxzoIDhFIW9/

 

[Mitch Gross]

Right, to get real 4096 we need 4096/.66 = ~6206 Bayer pixels, and for 3840 we need 3840/.66 = ~5818 Bayer pixels with proper filtered downsampling vs. line/pixel skipping. This is to achieve real post-deBayer detail along with minimal aliasing (we also need a good optical low pass filter on the sensor to cut frequencies above the sensor resolution per Nyquist).

This is why the latest ~6K cameras producing 4K look so much more detailed, without aliasing, vs. say the C300 II and C200 (and any other camera with a 4K sensor producing 4K output), which look much less detailed/softer and also alias on fine fabric etc. With a properly tuned OLPF, the 4096 sensor cameras can output 2.7K without aliasing. Comparing e.g the 6K Sony A7 III and Fuji X-T3 to the 4K sensor cameras shows this quite well.

Your math is a bit out of date. To get proper DCI 4K requires a 5.7K sensor. That’s why we have one in the EVA1. Same he proof is in the pudding when you point the camera at a resolution chart. Most “4K” cameras can only resolve around 1650-1700 line pairs, while the EVA1 clearly resolves path 2000 lp.

The 2/3 rule for Bayer pattern is a good rule of thumb, but modern deBayer processing is up in the 70s percentage-wise.

 

[jcs]

Your math is a bit out of date. To get proper DCI 4K requires a 5.7K sensor. That’s why we have one in the EVA1. Same he proof is in the pudding when you point the camera at a resolution chart. Most “4K” cameras can only resolve around 1650-1700 line pairs, while the EVA1 clearly resolves path 2000 lp.

The 2/3 rule for Bayer pattern is a good rule of thumb, but modern deBayer processing is up in the 70s percentage-wise.

While there's certainly been improvements in deBayer methods, which improves color separation and reduces artifacts, etc., the 2/3 / 1.5x oversampling is specifically for representing detail without aliasing per Nyquist Sampling theory. See the rotated line grids above. If we use just 2 pixels, black and white (our 'sine wave'), and we rotate it we get a mess of aliasing. Basically because we don't have any 'anti-aliasing' pixels to help smooth things out (visual description). The Nyquist reason is it's because we are at exactly 2x sampling. To get > 2x, we've got to add another pixel, and with 1 black line and 2 white lines (now 3 pixels instead of 2 for each line), we can rotate it without aliasing: see the second image.

This really applies to any image / sampling system, not just Bayer (in this case we're really talking about luminance- the Bayer grid as a monochrome image, not taking into account the issues of deBayering and color separation/artifacts etc.). Again, the point being max detail without aliasing. For this to work, we also need the proper OLPF cutting off frequencies above our target resolution: this is what effectively blurs max resolvable detail to provide the antialiasing as shown in the second example (done digitally in Photoshop).

EDIT: In 2011 Alister Chapman said the factor was .5 (2x): http://www.xdcam-user.com/2011/03/measuring-resolution-nyquist-and-aliasing/
I used to think this as well, and after a few debates online, I did the line pair test in Photoshop where I realized 2/3 (1.5x) was (barely) sufficient.

I think the challenge is OLPFs aren't perfect, and thus the more sensor pixels we have before downsampling digitally (with a nice filter, e.g. Lanczos or similar), the less aliasing we'll get while preserving max detail (the ideal image). Thinking about this purely digitally, instead of 1/2 line pairs from the full resolution grid, which will alias when rotated, the max we can get is 1/3 lines, because we need those 'antialiasing pixels'. Instead of 4096 * 1/2 = 2048 lines, it's 4096 * 1/3 = 1365 lines. When shooting a test chart with vertical lines, this won't be as apparent as when shooting diagonal lines. The issue with test charts is each person's perception of the start of aliasing and the extinction of detail is different. IMO the best test is very high frequency/detailed moving images with little or no aliasing, as can be seen with recent 5.7+K 4K cameras.

Also note the X-T3 has no OPLF- instead sampling from ~6K to get ~4K. I've only examined footage online, but so far haven't noticed significant aliasing for video, so it appears they're doing a good job downsampling while preventing aliasing. I've downloaded X-T3 raw stills, and when zooming in (Photoshop, not Iridient), there are 'worms'. This appears to be due to a deconvolution filter expecting an OLPF (to deconvolve = recover real detail (vs sharpen). Space telescope images use this as well): https://www.dpreview.com/forums/thread/4364814 . In summary, it appears if we use a higher resolution sensor we can relax and even remove the OLPF and downsample to the target resolution for a high quality (mostly) alias-free result.

 

[DLD]

Depending on sampling - or oversampling, as it were - one can get "more from less". S1R has a 187 MPX res in stills with an 8-shot composite. ARRI claimed to deliver a 4K res out of a 3.2 MPX sensor with in-camera upscaling on Amira/Alexa.

 

[Run&Gun]

https://www.instagram.com/p/BxzoIDhFIW9/

Whatever it is, it doesn't appear to have a native EVF. That's a Graticle sticking out in the pic.

 

[jcs]

Depending on sampling - or oversampling, as it were - one can get "more from less". S1R has a 187 MPX res in stills with an 8-shot composite. ARRI claimed to deliver a 4K res out of a 3.2 MPX sensor with in-camera upscaling on Amira/Alexa.

Composite shots are something different than the concept of sampling and Nyquist/Shannon theories (and practice). Upscaling with sharpening can work well to increase perceived detail, AKA accutance, however it won't create real detail. Mathematically all that's happening is we're increasing the contrast/difference between pixels (high frequency) and groups of pixels (lower frequencies). There is a way to create 'real' detail with upscaling: you guessed it, AI upscaling which Samsung is doing for their 8K TVs (can be done in desktop software too). A better term for this kind of detail is synthesized from real data (original pixels). So not real, but realistically synthesized.

Most people who have done significant audio editing are familiar with the FFT (a very fast version of the DFT) where we can edit sound at arbitrary frequencies. Basically what happens is the sound is converted into the frequency domain, the frequencies are adjusted, then it's converted back to the time domain. High detail sound is sound with lots of frequencies (or energy), especially the higher frequencies.

For images, we use the DCT, which is effectively the DFT for 2D images. We can examine images, provided they are mostly noise free, captured from sensors to determine how much real information is present using the DCT and examine the various frequencies, especially the higher frequencies to determine how much real detail is being captured: https://en.wikipedia.org/wiki/Discrete_cosine_transform

The lower right corner represents the highest frequencies possible, and the higher the 'weight' of that frequency in an image, the more high detail information is present. For example, comparing the same image captured with a 3.2K sensor compared to a 6K sensor for 4K output, the DCT will show much more high frequency information. Upscaling won't ever compete with real detail, though AI upscaling can synthesize plausible detail.

What started this 'analysis journey' was the realization that HD and later 4K cameras were not providing anywhere near True HD / 4K and manufacturers weren't being honest about camera performance. It's interesting that right around the time we started 'waking up', ~6K cameras started coming out shortly afterwards. Our discussions where we settled on the .66 factor was 2 years ago, where only the F65 and Red cameras were producing real 4K (from much higher resolution sensors). Now all new 4K cameras have at least 5.7K sensors with downsampling (Canon's like a broken record, still skipping lol (except perhaps the C700 FF)). Canon's next 4K video camera will have to have downsampling to compete.

So many people are waiting for the A7S III- which will likely get 10-bit (because of HDR TVs and the X-T3 etc.) and hopefully less (or no) RS, and will have at least 6K and possibly 8K video (oversampled from 12K??). If the next S1 has phase detect support, that will be interesting too (else not very interesting unless they stick LIDAR (lol) or something else in there to give better AF vs. contrast / DFD (which doesn't work very well in practice (maybe with some AI it can be better)).

 

[Cary Knoop]

99.9999% of all videos are made for showing natural images recorded through a lens and often with motion blur.
So, yes, in theory, you have a point but practically speaking those charts do not matter for a thing.

 

[DLD]

Composite shots are something different than the concept of sampling and Nyquist/Shannon theories (and practice). Upscaling with sharpening can work well to increase perceived detail, AKA accutance, however it won't create real detail. Mathematically all that's happening is we're increasing the contrast/difference between pixels (high frequency) and groups of pixels (lower frequencies). There is a way to create 'real' detail with upscaling: you guessed it, AI upscaling which Samsung is doing for their 8K TVs (can be done in desktop software too). A better term for this kind of detail is synthesized from real data (original pixels). So not real, but realistically synthesized.
...

That's right. It's synthetic. I recall from an old, old reviews of a CD player that, when compared to vinyl, someone might complain that the oboe ending in some symphony was a second too long. And the remixes/remasters of the classic rock albums, such as Deep Purple's "Made in Japan", originally recorded on a couple of joined 4-tracks, don't sound like either the vinyl or reel-to-reel. And digital video doesn't look like either film or analog video/VCR. And, at the end, it all evolves into a synthetic form of what appears the most pleasant to an eye, a mix of the digital's resolution, low ISO and the film's color palette. Or something like that.

 

[jcs]

99.9999% of all videos are made for showing natural images recorded through a lens and often with motion blur.
So, yes, in theory, you have a point but practically speaking those charts do not matter for a thing.

Most people don't care about charts, graphs, math, or sampling theory- that is true, even me when shooting/creating: the less tech, greater the simplicity, the better!

When one sees aliasing on their $16K camera which supposedly shoots 4K, one might be interested in understanding why. A few online discussions and a bit of analysis got to the truth. I much prefer working on creative things, though doing a bit of tech/math/science is creative in its own way too. This information is also helpful for those looking to purchase new cameras: if one desires alias-free true detail, they'll need a camera that sufficiently oversamples, e.g. 6K to get True 4K.

 

[DLD]

Whatever it is, it doesn't appear to have a native EVF. That's a Graticle sticking out in the pic.

If it's a tweaked S1, it might have a few additional connections and mounting points, which makes it suitable for serving as a customizable "electronic box with a sensor" a la Alexa Mini, Z-Cam or every Red camera ever made. Which is more of a modular camera design, a trend of the moment anyway but more so at this end of the spectrum.

 

[jcs]

That's right. It's synthetic. I recall from an old, old reviews of a CD player that, when compared to vinyl, someone might complain that the oboe ending in some symphony was a second too long. And the remixes/remasters of the classic rock albums, such as Deep Purple's "Made in Japan", originally recorded on a couple of joined 4-tracks, don't sound like either the vinyl or reel-to-reel. And digital video doesn't look like either film or analog video/VCR. And, at the end, it all evolves into a synthetic form of what appears the most pleasant to an eye, a mix of the digital's resolution, low ISO and the film's color palette. Or something like that.

For audio, I used to debate with a good friend that digital would surpass analog in every way, including simulating amps for e.g. guitars. He was not convinced one iota. His day job (life really) is audio and music; he's also an excellent guitar player. Years later, he got one of these: https://www.fractalaudio.com/ (an earlier version a few years ago). He tested the synthesized models against his own amps and gear and couldn't believe it: not only did it sound as good, it was effectively better because he could tweak it in ways the real gear could not be. He was blown away. Now with very high sampling, better analog+digital filtering, especially when listening on electrostatic headphones, no one will be able to tell the difference between a record/tape vs. the exact same recording converted to digital.

For images/video, AI is just getting started, and for moving images (video), no one will be able to tell resolution/detail was synthesized vs. real. The bigger challenge for AI, are the so called Deep Fakes, which is a lot more complicated vs. image scaling or detail synthesis. I did some research into Deep Fakes for a project, and my take on the whole thing regarding impact on society is it will be effectively zero (other than fun for entertainment). The same exact effect fake Photoshop had: people call out the fakes right away. What's important is the source of the videos and witnesses etc. The other angle is real video being called Deep Fakes- again same pattern: source, chain of custody, witnesses, etc. People are now very skeptical, in a good way. Truth Tech integrated into cameras could be a thing soon (blockchain based, etc.).

 

[Cary Knoop]

This information is also helpful for those looking to purchase new cameras: if one desires alias-free true detail, they'll need a camera that sufficiently oversamples, e.g. 6K to get True 4K.

Even such a camera is no guarantee, aliasing can happen as well when you downsample from 8K or 6K to 4K.
Sure you can soften the image but that bites in the resolution tail and you are back where you started from.

One thing that avoids aliasing is to make sure your camera refrains from any kind of sharpening.

 

[Gary T]

Whatever it is, it doesn't appear to have a native EVF. That's a Graticle sticking out in the pic.

Alicia Robbins Instagram post tags #panasoniclumix and Panasonic Lumix also has the Paramount Theater booked at 12:30PM on Friday so I think she is talking about a new hybrid camera.

Mitch has said there will be relevant Pro annoucements and Panasonic's Pro group has the Paramount theater reserved at 3:15 on Saturday. It looks like Team Panasonic will be busy this weekend. My hope is they go big and releases updates across their various product lines including an new Varicam and possibly a EVA2. I'd love to see at least one update with internal ProRes RAW recording.

Looks like this show might be pretty exciting after all!

http://www.cinegearexpo.com/la-expo-seminar-schedule

 

[Bern Caughey]

https://www.instagram.com/p/BxzoIDhFIW9/

Whatever it is, it doesn't appear to have a native EVF. That's a Graticle sticking out in the pic.

Appears the Gractical is taking a HDMI feed so the camera could have the typical mirrorless body with eye-level EVF.

Hopefully one day a manufacturer will build a tilting EVF into a compact body. Bonus points if it also has SDI.

 

[Samuel H]

Hi. Just here to say thanks because I'm enjoying very much all these offtopic detours into oversampling, digital audio, AI...

 

[DLD]

So, S1H - 6K, 10-bit internal, ND ... priced at ~ $6,000. Supposedly, a slightly larger body than S1 (to handle the heat).

An interesting step. Maybe this is why there hasn't been a 4K C300 MKIII/C400. They're all going full frame with a 6K being an intermediate stop, with Sony likely jumping into 8K already.

 

[puredrifting]

So, S1H - 6K, 10-bit internal, ND ... priced at ~ $6,000. Supposedly, a slightly larger body than S1 (to handle the heat).

An interesting step. Maybe this is why there hasn't been a 4K C300 MKIII/C400. They're all going full frame with a 6K being an intermediate stop, with Sony likely jumping into 8K already.

If the trend in the next crop of cameras truly is to jump to 6K and 8K, I predict flat or declining sales of these new cameras.

We are cuttinga 5-6 minute marketing/fundraising trailer for our documentary, not even the film itself, for which we have already shot about 80TB of 4K footage and our editor is working with proxies and it's taking his older system days and days just to render out Prores Proxy from the 4K DCI. In a world where the majority of deliverables to paying clients are still 1080, for most projects and deliverables, shooting 6K and 8K is foolish, it basically is just an excuse to eat media, clog up drives, compromise archival workflows and generally even 4K is a PITA unless you are short form like commercials and music videos. But when you are shooting dozens of hours of interviews and b-roll for episodic, reality, documentary, event, it's truly just a nuisance. Take that inconvenience, esepcially when shooting 4K RAW as we mostly have and you soon find that shooting huge raster formats to be a hindrance more than an advantage. I think many users who are all excited about shooting 6K, 8K or beyond haven't actually shot and taken entire projects through workflow. As a producer, I hate 4K, 6K or 8K even for a 4K finish is just not worth the hassle, extra money and time, when you look at the on screen results. The extra money and effort is better spent on another crew position, better lighting, wardrobe, props, production value rather than pointless extra resolution. Maybe someday when computers that can handle this footage are common and affordable (ever notice that in the real world, most editors DON'T edit on a maxed out, top spec, latest hot rod, but typically pretty modest 2-4 year old computers?) and as soon as 30-60TB T-Bolt 3 FAST RAIDS are commonplace (I have a 16TB T-Bolt 2 RAID and when I benchmark it half full of 4K RAW, it's relatively sluggish, 500-600 Mbps) In the real world, 4K is a PITA. I shot and edited a corporate project in December, not even in RAW, just XF-AVC witjh two C200s and even with the relatively small amount of data, it still was a PITA because I had to make a lot of multi-cam clips that would barely play in RT to edit with, just two streams of UHD from a decently fast 12TB OWC T-Bolt 2 RAID. I made it through but it was painful having to start stop and try to make intelligent cuts using multicam clips. If we would have shot the same project in Cinema RAW Light, it would have been so much more painful.

IMHO, these formats cause Cognitive Dissonance with the real world of production, at least in my medium to low level of production. Clients (including most of the studios other than Netflix) don't care or want even 4K in most cases, shooting 6K or 8K is mostly delusion about "futureproofing, shooting at at least 5.7k to avoid aliasing in 4K, etc." If your clients don't want it or care about it, who are we buying these cameras for and shooting these formats for? To impress your fellow production people? I don't get it. I'll be at CineGear on Friday, shooting interviews with and having marketing discussions with companies, manufacturers and third party providers/accessory makers, but I never lose sight that the everyone I will be speaking with, their job depends on selling the illusion to all of us about what WE need and what OUR clients need. It's too bad more people in our business aren't circumspect and thoughtful about what value is and how to deliver it to their clients. In most cases, ultra high resolution content isn't a good value.