Canon 5D Mark III - 3.5K and 4K raw video with Magic Lantern

[cantsin]

31 minutes ago, maxotics said:

Many of us have hashed this to death.  The core problem is you can't save 12 stops of full-color dynamic range in an 8-bit space.  It doesn't matter how those numbers are represented, LOG or otherwise.  The 8-bit space was designed for our viewing, not for our capturing.  LOG can't magically "compress" the sensor data into an 8-bit space.  It can help in telling you where you can trade off some data in one area for another, as Don mentions above, but the rub is the area you're getting a lot of noise is also the area you can least afford to lose if you're extending beyond the 6-8 stops you can capture in 8-bit!\

I am fully aware of this. It's just that (digital) cinema projection isn't 8bit (and much more than 8 stops), and HDR video isn't 8bit either. Even with run-of-the-mill displays, we're already beyond 8 stops. 8bit color is becoming a thing of the past for this very reason. The cheapest TFT displays on the market now have contrast ratios of about 1:900, for which you need 10bit (9.8 to be really precise....) + 10 stops dynamic range if you want to display an image without color banding.

Plus, as your photographic print example suggested, we need to differentiate motif dynamic range and display dynamic range. Of course you can take a picture of a landscape at noon with 15 or more stops dynamic range in the motif, capture it on negative film or a good digital camera which can hold those 15 stops, and print the recorded image on a postcard (which, btw., if it's printed on a high-quality, 2.3 density paper, will have about 7 stops dynamic [display] range) without the sky being white or the shadows drowning in black - but of course you'll lose differentiation in the midtones. In this respect, photography has routinely compressed dynamic range because our mental image of the landscape at noon always has blue, never white skies, and detail in the shadow, not a pitch-black mass - never mind the fact that, as you correctly wrote, the eye actually only captures 6.5 stops and achieves this 'HDR' picture through quick brightness adaptation and mental compositing. 

(Our ear can also hear sound frequencies that a speaker cannot reproduce, for example a low note in a symphony playing on a cheap radio, because our brain reconstructs the missing bass frequency from the overtone spectrum. In that sense, the mere physical parameters of our seeing and hearing capabilities literally do not 'paint the whole picture'.)

[Andrew Reid]

I think you guys should debate this in a new topic, because it's killing this one for the rest of us.... title of the topic is "What Magic Lantern reveals - 5D Mark III capable of 4K all along"

I really mean it will all the good will in the world, please for the love of all that is holy open a new thread instead of using this one... I've nothing against you guys arguing about the science of dynamic range over 20 pages....

Just put it in the proper place so this one can stay on the topic of Magic Lantern and the 5D Mark III raw. PLEEEASE!

[tugela]

On ‎9‎/‎7‎/‎2017 at 4:31 AM, Andrew Reid said:

The hardware is 100% capable of 4K video and likely with the same MJPEG codec as the 5D Mk IV and 1D C. The reason it is not in there is due to marketing. They wanted to reserve the feature for the 1D C and target the 5D3 predominantly at a stills audience and consumer crowd for whom 4K would be somewhat overkill or wasted and result in more support costs, distracted focus in the marketing messages, confusion on the shop floor, etc. The 5D3 also wasn't allowed to compete against any of the pro Cinema EOS cameras which were fresh on the market at the time it was released, because the margins would be much greater on these cameras and video shooters should not have been tempted by something cheaper offering such a recording high spec... That's another reason the 1D C cost $15,000 on release.

Plenty of compact flash cards in 2012 were fast enough for MJPEG 4K, otherwise they wouldn't have released the 1D C the same year doing MJPEG 4K to compact flash cards. You talk like the cards didn't even exist or would drop frames.

 

Every camera with enough resolution on the sensor is capable of 4K video. What limits it is the processing requirements and interface data transfer rates.

MJPEG may not be an efficient compression but it is still compression and requires processor overhead (which raw does not). So putting mjpeg into a 5D3 is not as simple as you seem to think, even if it can be hacked to record raw. Sustained operation at very high bit rates on equipment not designed for that may also reduce the service life of components within the camera, which is a huge issue if you have to cover the product with a warranty. Users may not worry about stuff like that, but manufacturers sure as hell do.

One other thing that people tend to forget is that manufacturers have to sell products that deliver their specs. Dropping frames, even occasional frames, means that mode will likely not make it into the product. Whatever is included has to be reliable, and reliable in all cameras produced in that model, not just the copy you have. Otherwise you will get returns, warranty services requests and all sorts of crap that eats away any profit you make.

For example, you might be able to make your car go really fast by removing all sorts of safety features and adding nitro, but no manufacturer would sell the car like that because of the liability and reliability issues that go along with doing that.

[Andrew Reid]

Well in my opinion, the hardware is just as capable as the 1D C as doing 4K MJPEG. None of us really knows for sure though, as we didn't design the camera.

Seen as the 1D C is basically a 1D X, that's probably capable of 4K as well.

The lossless raw from Magic Lantern uses the 5D3's existing MJPEG compression by the way, in the 5D Mark III, applying it to raw 3.5K

It's in there.

I'm not denying that had business reasons not to switch it on.

[cantsin]

34 minutes ago, tugela said:

For example, you might be able to make your car go really fast by removing all sorts of safety features and adding nitro, but no manufacturer would sell the car like that because of the liability and reliability issues that go along with doing that.

I like your analogy. We're car tuning nutcases with our cameras, taking them to unofficial, off-industry races like in "The Fast and the Furious". ;-)

[Andrew Reid]

Except car tuning is mainly hardware based, and you are not changing the hardware in the 5D3 or even overclocking it with Magic Lantern. It's purely software, running as normal, like an app on your phone.

[cantsin]

7 minutes ago, Andrew Reid said:

Except car tuning is mainly hardware based, and you are not changing or overclocking / turbo-charing the hardware in the 5D3 at all with Magic Lantern. It's purely software.

With a few exceptions: Maxing out memory card bus bandwidth way beyond what Canon would allow in its end-user product; manipulating the hardware registers in order to reduce global drawing, display color and display FPS in order to achieve higher video resolution and/or fps; allowing resolution/fps combinations that make the camera drop frames after a few seconds, and that are therefore outside the legal specs...

Yes, it's all software-based, but the controller chip hardware is designed in such a way that it doesn't have hard limits for features that overtax the rest of hardware; instead these limits are implemented in software and can therefore be tweaked/maxed out by MagicLantern. In that sense, it's not that far from tuning a car motor. (Which, as we know from Volkswagen, is nowadays done in software, too...)

[Andrew Reid]

You're still wrong.

A car is a mechanical system, with moving parts in the engine, and the engine mapping is under the influence of software.

When you force a mechanical system to work faster it breaks.

So that would be more analogous to overclocking a CPU without adequate cooling.

Magic Lantern aren't doing that.

The hardware is still running under the speed limit set by Canon's spec.

Is the data throughput higher in raw recording? Yes. But still within what it is designed to do.

Does it consume more power and therefore get a bit hotter? Yes. But still within the spec.

And that is why the comparisons to removing safety features from cars and adding nitro is complete nonsense.

This is sadly what Canon reps seem to believe as well.

An image sensor, processor and card are not mechanical parts. It's not a fucking car.

The 'manipulating hardware in order to avoid overtaxing the system' you observe, is simply about allocation of existing system resources to different tasks. Reverse engineering and hacking is simply not as efficient with resources, versus rewriting the source code with a full understanding of it. The camera is a black box of ARM processors and everything Magic Lantern knows is a limited subset of what code and routines exist.

The only choice Magic Lantern have to maximise the efficiency of their features like 3.5K raw is by memory management and tweaking the running of existing tasks like global draw, so that the ARM processors spend more time on something else. They are not 'working harder'.

Oh and by the way, the raw codec is less computationally intensive than H.264.

[cantsin]

34 minutes ago, Andrew Reid said:

Is the data throughput higher in raw recording? Yes. But still within what it is designed to do.

Does it consume more power and therefore get a bit hotter? Yes. But still within the spec.

How do you know that? And data throughput can be set to values that are higher than what the hardware is designed to do, otherwise there would be no frame dropping or aborting. This is fully analogous to overclocking a CPU - only that it is done with the memory bus. 

EDIT: And MagicLantern is fully open and transparent about that, see the official FAQ:

Quote

 

Is it safe?

No. Magic Lantern was created by reverse engineering an undocumented system that controls hardware. Therefore, we can't be certain that it's 100% safe.

 

http://wiki.magiclantern.fm/faq

[Andrew Reid]

The memory clock speeds and bus speeds are fixed.

You are increasing resolution, bit-depth and so on... that's all.

Then when it gets close to the maximum of the card controller's transfer speed or the sustained write speed of the card, yes it stops or skips frames. I have it set to stop recording when a frame is skipped. Nobody wants to be fooled into thinking they have a flawless recording when it is actually missing frames. With it stopping instead, I know to reduce the resolution for a longer recording time.

The sensor modes already exist. Canon meant this sensor to do 4K. It does so every time you hit the 5x magnification key, to supply that 1:1 image to the live-view display, which you can then scroll around across the entire sensor. The window on the sensor moves around the full 5.5K width, for example.

I have been using Magic Lantern raw for 4 years now and my 5D Mark III hardware did not have any issues. There's no firm data on failure rates or whether the lifespan of the hardware reduces depending on how it's used by Magic Lantern. The "not certain that it's 100% safe" is just a disclaimer, to stop them getting into any legal trouble with users.

I think it takes away from their achievements a bit by saying that it stresses the hardware to breaking point. That discourages use. Are you trying to discourage use of Magic Lantern? Poor form if you are.

Shoot and be happy!

[cantsin]

51 minutes ago, Andrew Reid said:

I think it takes away from their achievements a bit by saying that it stresses the hardware to breaking point. That discourages use. Are you trying to discourage use of Magic Lantern? Poor form if you are.

Nope, not discouraging it at all being an ML shooter myself. But one also needs to be frank/brutally honest to others (who lack the experience) that it is and will remain an experimental system. Maybe a better analogy than tuned cars is how in the past, indie filmmakers modded their 16mm cameras (including budget Bolex and Krasnogorsk cameras) to Super 16. The s16 format was, in fact, created as a DIY hack by a filmmaker in 1969:

http://www.fdtimes.com/2009/08/20/the-early-years-of-super-16-and-how-it-all-started/

[Andrew Reid]

Yes it's experimental. We all realise that.

No amount of analogies with mechanical cars and mechanical cameras (Super 16 film) will convince me they are relevant to software and processors though

[Drew Allegre]

Yeah, this is pretty impressive stuff.  I'm used to Canon h264 footage, so this 3.5k raw stuff is a little difficult to comprehend.  Here's a dng still run through ACR.  

And thank you to Don Kotlos and maxotics for your explanations on the raw/log.  Makes total sense to me now.

[Andrew Reid]

The h.264 is pixel binned as well, so it isn't just compressed to hell, 8bit, 4:2:0 and probably with some weird Canon intentional crippling in terms of resolution

The 3.5K is a full pixel readout of a Super 35mm width window of the sensor with no binning or line-skipping.

Crop the centre out of a full frame 22MP RAW still and that's now your video mode.

Enjoy!

[tugela]

On ‎9‎/‎8‎/‎2017 at 5:08 PM, Andrew Reid said:

Well in my opinion, the hardware is just as capable as the 1D C as doing 4K MJPEG. None of us really knows for sure though, as we didn't design the camera.

Seen as the 1D C is basically a 1D X, that's probably capable of 4K as well.

The lossless raw from Magic Lantern uses the 5D3's existing MJPEG compression by the way, in the 5D Mark III, applying it to raw 3.5K

It's in there.

I'm not denying that had business reasons not to switch it on.

The 1D C has dual processors, the 5D3 has a single processor, so quite a bit more processing power. Maybe more, IIRC the 1D cameras have a third processor as well which is dedicated to helping with focusing.

The 1D C almost certainly has additional hardware logic included specifically to help with video as well, which is what provides it with capabilities that the 1D X does not have.

[Andrew Reid]

1 hour ago, tugela said:

The 1D C almost certainly has additional hardware logic included specifically to help with video as well, which is what provides it with capabilities that the 1D X does not have.

Prove it

[tugela]

18 minutes ago, Andrew Reid said:

Prove it

I don't need to. The fact that that they had to make a separate camera to do video should be proof enough. If you think that is otherwise, the onus of proof is on you, not me.

It would have been more cost effective to simply have added the feature to their existing professional camera and expanded the market for that.

[Don Kotlos]

The only info comes from an unknown source at Canon but states:

"reworked circuitry and design to dissipate heat for the 4K recording”

http://www.canonrumors.com/the-canon-eos-1d-c-is-different-than-the-eos-1d-x-on-the-inside/

If that is the only thing that came from Canon, then Andrews' point about 1Dx being 4K capable is correct even if that meant overheating or more noise. Now that source might have been wrong, but nowhere did Canon claim otherwise. 

Maybe at that time point they could charge more by introducing a new camera than a paid firmware upgrade. Market is all about perceived value, and a replacing X with C did it. 

[Andrew Reid]

2 hours ago, tugela said:

I don't need to. The fact that that they had to make a separate camera to do video should be proof enough. If you think that is otherwise, the onus of proof is on you, not me.

It would have been more cost effective to simply have added the feature to their existing professional camera and expanded the market for that.

Yes but two things...

1. My logic is superior to yours

2. I actually own the 1D C

So try again fat boi.

Listen... They did not HAVE to make two cameras, they chose to do so, due to the different target markets and pricing.... between Hollywood and pro stills.... And smaller quantities of sales for the 4K filmmaking tool vs the all-round number 1. pro stills camera in the world.

The sensor, and all the features are identical, they look the same, handle the same and weigh the same.

It makes no sense to design two cameras so similar and then flippin switch all the insides around, new sensor, new motherboard, etc.

It's nonsense.

Also numerous Canon reps are on record as saying it was just a 1D X with firmware update... Until they were told to shut up about it and lie instead. Heatsink!! Erm!! Hot!

[zerocool22]

How is the 3,5K compared to the 1080 RAW? Is it more detailed. 1080p is enough resolution for me, but It always lacked detail in my experience(besides DR). Not sure if the the 3,5K is a detail improvement over the 1080p. If not, then I would just use the 1080p.