How does BMPCC video/film compare to Canon DLSR with Magic Lantern raw video?

[l_d_allan]

from:
http://www.luminous-landscape.com/reviews/camcorders/black_magic_pocket_camera.shtml
> this camera is capable of producing stunning image quality. Quality that far exceeds that of any video DSLR,

I'm curious how much of an "equalizer" the Magic Lantern raw video capability (and other advanced video + focus + metering capabilities) provides for Canon DSLR's, especially the newer 5d3.

Or is this "apples and oranges"?

[jjj]

I saw some footage of Canon raw DSLR footage when the hack first appeared and it was markedly better than the BMCC in some comparison shots. The link to it would be in the Magic Lanterns forums somewhere.

[paulmcmurrick]

hard to know what markedly better means. Depends on grading, capture codec etc

I have both a 5d and BMPCC and markedly prefer the output from the latter

[John.Murray]

raw aside, color quality from ML is 4:2:0, the bmpcc is 4:2:2

[jjj]

hard to know what markedly better means. Depends on grading, capture codec etc

I have both a 5d and BMPCC and markedly prefer the output from the latter

Is that raw from the Canon too?

The shot of the path up to the trees is in this comparison is where the BMCC wasn't as good as the Canon with the raw hack - which has been much improved since this test. And the article also mentions poorer high ISO performance compared to the Canon.

[bcooter]

Is that raw from the Canon too?

The shot of the path up to the trees is in this comparison is where the BMCC wasn't as good as the Canon with the raw hack - which has been much improved since this test. And the article also mentions poorer high ISO performance compared to the Canon.

I know when you test these motion cameras, it's much more than just color information, or bit depth, even at times mbps/kbps.  

When you look at 444 which is 100% color info, 422 50%, 420 25% you'd think at 420 you'd have a file that would be lifeless, but there is so much more to digital video than just the numbers.

What most of these tests don't show is cameras at their true iso.

I don't know about the black magic (other than the people I work with that have used them on set), or for that matter a hacked 5d2/3, but for the RED 1's, the base iso is 800.  Actually the iso is always 800 and all your doing when dropping or raising the iso is pushing or pulling a curve.

Actually with the R1's, pulling it down to 100 is worse than pushing in regards to highlight clipping, (at least in my experience) and now we just set the camera at 800 and leave it, using filters, or light to get to the desired exposure.

It's funny, the Reds shoot 444, the gh3's we use shoot 4:2:0 but under the right conditions you can't tell them apart, even when moving the files in post.  Then again, under other conditions the difference is huge.

I'm not an engineer, but I know there is a lot going on under the hood of these things that makes a huge difference, especially combined with different lighting, subject, ambient color, ambient bounce and the big killer, heat.

With the R1's I would leave them on, when shooting stills, or just to avoid the long boot up time and if you then go into a long take, you start building up noise.

The Gh3's will do the same, though not as pronounced and I've been told the original bm 2.5k cameras are also affected by heat.

So comparing these cameras is a difficult task, especially if you just go from the specs, because all of these cameras have a sweet spot.  When your on, they are all marvelous, when your off, they can be a mess.

I think Digital Video is where digital still capture was about 10 years ago, with the quality and use all over the place and I assume in a few years with improved and more standard software, graphics cards and better camera design, they will all level off like still capture has.

Not having any inside information, I think the next step is a true combination camera, something like a Canon 1dc 4k and a Panasonic gh3, but both with better sound inputs, maybe even some dedicated coloring/processing suites.



IMO

BC

[bcooter]

Alexa's native iso is 800 too, like the R1.

IMO, a part from Red (Arriraw is too expensive), raw video has still a long path to recover until it becomes
as matured and widely supported as still imagery, hopping they will avoid the proprietary mess, wich I think they obviously won't.

But guys, I don't get a point there: how can Canon hack claims raw capture if deliver 4:2:0 and BMPC 4:2:2 ?
No possibility of 4:4:4 ? Absurd. Better get a good Prores or DNx 444.

I've graded a lot of imagery, from a lot of cameras and all I can say and be definite, is test, test, test.

Every camera looks a little different, but lighting and subject, including ambient bounce have as much to do with the final look (and resolution) as anything the camera does.

Obviously a 1/3" chip, avchd file won't hold up under high iso and hard 13 stop lighting, but then again hard 13 stop lighting probably shouldn't be used anyway.

I just finished an edit with RED and GH3 footage mixed.   I rarely compare imagery direct, but had one scene shot A cam with the R1 and  we shot b camera with the gh3's and it's a fairly challenged look at the end of the day.

Bottom line is I can't see any difference, I mean any difference and the RED is 444 the GH3 420 and if there is a difference it's so minor it will never be noticed.

This doesn't hold true for all imagery, but then again, a 420 file is 1/4 the color information (on paper) of a 444 file and we should see a difference . . . we don't.

I know other people will disagree, but I have nothing against a baked in gh3 file vs. a 444 raw.    The secret is to shoot the gh3 slightly flat, but not c-log, flat and colorless, but just slightly smoothed out, then grade to your hearts extent.

If anything I would like to see is the next round of cameras shoot a prorezz file, rather than an h264, even if it's a larger size and not for quality, because a high bitrate h264 file looks virtually identical to a prorezz file, but just do prorezz so we don't have to transcode.

IMO

BC

[John Brawley]

Bottom line is I can't see any difference, I mean any difference and the RED is 444 the GH3 420 and if there is a difference it's so minor it will never be noticed.

This doesn't hold true for all imagery, but then again, a 420 file is 1/4 the color information (on paper) of a 444 file and we should see a difference . . . we don't.

I don't know about the black magic (other than the people I work with that have used them on set), or for that matter a hacked 5d2/3, but for the RED 1's, the base iso is 800.

It's funny, the Reds shoot 444, the gh3's we use shoot 4:2:0 but under the right conditions you can't tell them apart, even when moving the files in post.  Then again, under other conditions the difference is huge.

Just want to mention a couple of things.

It's not really right to describe a RAW camera as having colour encoded space like 4:4:4 or even 4:2:2.  This is the wrong terminology.  This is video encoding terminology.  

RAW cameras like the RED and the BMCC and even the ML hacked 5D are just that RAW cameras. They can be Log or Linear encoded, and you can TRANSCODE 4:4:4 or 4:2:2 or even 4:2:0 files from their RAW files, but you don't refer to them by these descriptions.

That's because there's a lot of other things that affect the colour subsampling ratio.

It also gets confusing because Bayer sensors, which these RAW cameras generally have, have differing ratios of GREEN to BLUE to RED pixels so some would even argue that you can never have a 4:4:4 image without massive oversampling. (an 8K sensor like the SONy F65 for true 4K 4:4:4 encoding).

A lot of people ALSO make the MISTAKE of assuming that the RATIO in bayer sensor must have something to do with the RATIO's used to describe video encoding.  For example, that because bayer sensors generally have 2 green photo sites for every red and blue photo site, this lines up nicely with 4:2:2.  It's a coincidence.  You need to look way beyond this gross simplification and understand it's a lot more complex than this.

Someone will likely jump on this and elaborate, but it's all kind of futile.  RAW sensor data can't be described as video encoded space like 4:4:4 until it's transcoded.  And even then you'll have endless arguments about what the original sensor size needs to be to derive that number.

I should also add, that even though can't see the difference, you will certainly see it once you start to grade some files.  And here its' just like comparing a JPEG to a RAW.  So if you can get a JPEG right out of camera then you're golden, but as soon as you want to push it around, you are quickly in trouble....

You also describe that the R1 has a native ISO of 800.  It's actually 320, as is the R1 MX and the EPIC (which uses the same MX sensor)

RED have long advocated exposing at 800 ISO to protect the highlights, and that's what they call a "recommended" exposure.  I personally find the RED has very little HEADROOM and rating at 800 is their way of trading some shadows for highlights.

jb

[Chris L]

Good to see you here JB. I have a BMCC and am loving it.

Does anybody here have the BMCC and the Canon 5d3 with Raw? I heard the Canon raw has 14 stops of DR, and the BMCC has 13, I am curious if this is true.

[bcooter]

John,

Ok your right, it is linear until encoded to 444 or whatever, but regardless, a 444 out of cinex doesn't look much different than a 420 out of a h264 gh3 of the same scene.

I've run em side by side enough to know that part.

Now the color response and look can be different, but in the end there is not a heck of a lot of difference in color visual depth or dr.  Maybe the REDs have a stop more.

In regards to iso, yea, I've heard 320, I've heard 640 but every dp I know says lock it on 800 and don't pull it down to 200 or  it clips.  In fact with all digital video I find I have a lot better chance to recover shadows than pull in detail from highlights.  I know from experience don't go past 1000 with the mx sensor, ever.

Still it depends on what, how you shoot.   I do know that with any motion image how you set up in the front end can save you a week in the back.   These things, raw or cooked are sensitive, have their own properties and getting a base white balance then filtering for color I find preferable to moving the knobs trying to get a look from the camera.  

I now treat all dv, either raw or cooked like I'm shooting transparency film.  I get my base, filter for the rest and don't try to produce a 100% different look in post.  That's a fight and meant for people with a lot larger budgets than mine.

People always like to say, _______________was shot with a _________ s it must be good, but they kind of forget that when they are looking at big budget theatrical they are looking a huge money and time.

Gravity took 4 years to do and they invented technology as they went along.  Not that it wasn't worth it because it's a good movie, but it's also something you can't do with any camera, working with two lights in the basement.

Today, someone brought over some underexposed 5d3 footage from a recording session shot at a billion iso, with a wb of daylight, shot tungsten. I dropped it into color and opened it up, moved some of the red out and it was pretty, though very noisy, but still pretty and that's from a cooked h264 file shot at the wrong wb and about 2 stops under.

Then again, everyone works different though nothing and I mean nothing improves the look of a digital motion file like time and money.  With time and money in the pre, shoot and post production you can make anything look good if you work professionally.

Given that I don't think raw is a magic godsend any more than I think one camera is 100% better than the other.

IMO

BC

BTW:  I love the shot on your site titled Reef Ireland.  Very pretty.

Just want to mention a couple of things.

It's not really right to describe a RAW camera as having colour encoded space like 4:4:4 or even 4:2:2.  This is the wrong terminology.  This is video encoding terminology.  

RAW cameras like the RED and the BMCC and even the ML hacked 5D are just that RAW cameras. They can be Log or Linear encoded, and you can TRANSCODE 4:4:4 or 4:2:2 or even 4:2:0 files from their RAW files, but you don't refer to them by these descriptions.

That's because there's a lot of other things that affect the colour subsampling ratio.

It also gets confusing because Bayer sensors, which these RAW cameras generally have, have differing ratios of GREEN to BLUE to RED pixels so some would even argue that you can never have a 4:4:4 image without massive oversampling. (an 8K sensor like the SONy F65 for true 4K 4:4:4 encoding).

A lot of people ALSO make the MISTAKE of assuming that the RATIO in bayer sensor must have something to do with the RATIO's used to describe video encoding.  For example, that because bayer sensors generally have 2 green photo sites for every red and blue photo site, this lines up nicely with 4:2:2.  It's a coincidence.  You need to look way beyond this gross simplification and understand it's a lot more complex than this.

Someone will likely jump on this and elaborate, but it's all kind of futile.  RAW sensor data can't be described as video encoded space like 4:4:4 until it's transcoded.  And even then you'll have endless arguments about what the original sensor size needs to be to derive that number.

I should also add, that even though can't see the difference, you will certainly see it once you start to grade some files.  And here its' just like comparing a JPEG to a RAW.  So if you can get a JPEG right out of camera then you're golden, but as soon as you want to push it around, you are quickly in trouble....

You also describe that the R1 has a native ISO of 800.  It's actually 320, as is the R1 MX and the EPIC (which uses the same MX sensor)

RED have long advocated exposing at 800 ISO to protect the highlights, and that's what they call a "recommended" exposure.  I personally find the RED has very little HEADROOM and rating at 800 is their way of trading some shadows for highlights.

jb

[jjj]

I've graded a lot of imagery, from a lot of cameras and all I can say and be definite, is test, test, test.

Yup.

Every camera looks a little different, but lighting and subject, including ambient bounce have as much to do with the final look (and resolution) as anything the camera does.

Even the same camera can look very different at times. Sometimes matching two cameras in same lighting setup is probably easier than matching same camera with very different lighting conditions. Particularly if ISOs are very different. Though this is where a raw file is particularly useful as at least you can colour balance after the fact which can to help match shots.

I also think video is like shooting transparency, so is like stepping back in time as a result.

[bcooter]

Yup.
Even the same camera can look very different at times. Sometimes matching two cameras in same lighting setup is probably easier than matching same camera with very different lighting conditions. Particularly if ISOs are very different. Though this is where a raw file is particularly useful as at least you can colour balance after the fact which can to help match shots.

I also think video is like shooting transparency, so is like stepping back in time as a result.

In an interview type scene we have set up two R1's with every setting identical and one Scarlet.    Just working three angles when you go to cinex out of the can looks, there will be a difference, mostly just from the camera angles and color ambient bounce, maybe due to a different exposure setting.  The R1's to the Scarlet never match exactly, regardless of RED gamma and RED color settings.

I just believe that dv is very, very receptive to ambient color and bounce or there is something else going on under the hood that I don't understand, but the looks are different.

In fact watch a news roundtable discussion where they run multiple cameras and you will also see a slightly different skin tone and color look from each angle and this is from studio situations where nothing moves.

IMO

BC

[jjj]

In an interview type scene we have set up two R1's with every setting identical and one Scarlet.    Just working three angles when you go to cinex out of the can looks, there will be a difference, mostly just from the camera angles and color ambient bounce, maybe due to a different exposure setting.  The R1's to the Scarlet never match exactly, regardless of RED gamma and RED color settings.

I just believe that dv is very, very receptive to ambient color and bounce or there is something else going on under the hood that I don't understand, but the looks are different.

Lenses can make a big difference too. F stops can be somewhat shall we say optimistic at times and is why t-stops are used on cine lenses. My f2.8 primes are much faster than my f2.8 zooms. Not to mention the colour variation between different lenses. Had issues matching shots from 2 different Canon bodies because of this once when doing a two camera setup.

[John Brawley]

John,

Ok your right, it is linear until encoded to 444 or whatever, but regardless, a 444 out of cinex doesn't look much different than a 420 out of a h264 gh3 of the same scene.

I was making a few points, but I do think it's worth making distinction here about what "looks" to the eye fantastic, though its easily fooled and then what happens when you try to grade it later.

This is when H.264 will fall flat on it's arse.  It can look great OOC but once you push the grade around you will soon "SEE" a difference.  And that's where RAW cameras do have an advantage....once you need to move into any kind of PP work, higher bit depth and less compression means better results. Just like RAW Vs JPEG....

In an interview type scene we have set up two R1's with every setting identical and one Scarlet.    Just working three angles when you go to cinex out of the can looks, there will be a difference, mostly just from the camera angles and color ambient bounce, maybe due to a different exposure setting.

I shoot narrative TV drama and we regularly use more than one camera, and often 3 and as many as 6 on a single scene.

I'm in pre now for a drama series which in it's 5th season.  We shot the first RED one, then RED ONE MX and then EPIC and now Alexa.

None of the cameras ever match with all the settings being the same.  Every camera is different even in RAW and it just gets compounded by slight differences in lenses (even the same versions), ND filters and the biggest of all, different shooting angles.

Even at the same exposure, shooting from one angle light gets reflected off of a persons face totally differently to shooting at 90 degrees offset, so they will never ever match at "default".  That's exactly why you have a colourist and the footage needs to be graded.  To me, that's an even bigger leap than the simple small differences between individual cameras.  And that's also why higher bit depth and less compression counts for a lot to me....

I have four Alexa bodies on the main unit of the show I'm now and two on the second unit.  Out of 6 bodies, I'm lucky if two of them line up the same.  There's even bigger differences when shooting EPIC.  
 
jb

[hjulenissen]

When you look at 444 which is 100% color info, 422 50%, 420 25% you'd think at 420 you'd have a file that would be lifeless, but there is so much more to digital video than just the numbers.

For most natural scenes, 4:2:2 and 4:2:0 throws away information that was hardly there in the first place.

For "reasonable" editing, 4:2:2 and 4:2:0 seems to matter little to most human viewers.

For imagery that will be lossy encoded for distribution, 4:2:0 is the norm.

For camera sensors that have a Bayer CFA at the same pixel count as that of the luma channel, "full"/true 4:4:4 information is not available.

...Which is not to say that 4:4:4 _never_ have merits. I understand that green-screening have benefits from 4:4:4. In that case, one is "encoding" specific and critical encoding into the color difference channels, and using nonlinear editing that serves to make chroma subsampling more visible. Colored, sharp text (end credits) on low-resolution DVD can have annoying 4:2:0 artifacts.

If you look at the e.g. 100:1 "throwing away of information" that lossy compression, 8bit/gamma, 4:2:0 etc does to a raw signal, one might be surprised that our digital images resemble the scene at all. Then compare the massive loss of information that any sensor/lens does to a scene. I guess the take-away point is that bits & bytes are poor indicators of perceived quality.

-h

[John Brawley]

For camera sensors that have a Bayer CFA at the same pixel count as that of the luma channel, "full"/true 4:4:4 information is not available.

.

As I mentioned earlier in this thread that however tempting it is to do, it's really incorrect to link video encoding terminology like "444" to the ratio of pixels in a bayer sensor and to talk of it never having "444" worth of information.

While the ratio in both examples seem to correlate and seems to indicate otherwise the truth is that it's much more complex than simply looking at the ratio of pixels.  

Yes. Bayer sensors have a differing ratio of Colour pixel ratios but you never describe them using video encoding terminology.  You're only perpetuating an incorrect understanding of how these numbers affect the end result.

JB

[hjulenissen]

As I mentioned earlier in this thread that however tempting it is to do, it's really incorrect to link video encoding terminology like "444" to the ratio of pixels in a bayer sensor and to talk of it never having "444" worth of information.

I believe that you are wrong. Of course a Bayer CFA is not a "4:2:0" format, but there is some value to comparing the two.

While the ratio in both examples seem to correlate and seems to indicate otherwise the truth is that it's much more complex than simply looking at the ratio of pixels.  

Yes. Bayer sensors have a differing ratio of Colour pixel ratios but you never describe them using video encoding terminology.  

I describe sensors using sensor terminology and video standards using video standard terminology. Whenever I think that it makes sense to compare those two, I attempt to use appropriate terminology.

A Bayer sensor of 1920x1080 sensels will have approximately 2 million samples.

A 1920x1080 4:4:4 image frame will have approximately 6 million samples.

Intuition and information theory tells us that a 2 million sample sensor cannot generally feed a 6 million sample file with independent information; there will be some redundancy.

If you have a scene where all spatial information happens in the "blue" channel, the effective resolution of the Bayer CFA is 960x540. Clearly, this does unjustice to a 1920x1080 4:4:4 format where the Cb channel is sampled at 1920x1080.

Scenes tends to behave in certain ways that are beneficial to the Bayer CFA. Humans tends to perceive visuals in certain ways. That is where things get complex. The maths of standardised color 3x3 matrixing and spatial up/downsampling is not that complex, and there is no reason to make it into something complex.

You're only perpetuating an incorrect understanding of how these numbers affect the end result.

JB

I am sorry that you see things that way.

-h

[Morgan_Moore]

None of the cameras ever match with all the settings being the same.

What I don't get is why, especially with raw cameras, one cannot shoot a gregtag, or other brand card and have the software build a profile for each camera - colour by numbers totally exact.

One click, perfect match*.

Sinar had this in stills in 2005.

S


*obviously angle change and the like may make the shots not match but that is a different story.

[jjj]

What I don't get is why, especially with raw cameras, one cannot shoot a gregtag, or other brand card and have the software build a profile for each camera - colour by numbers totally exact.

One click, perfect match*.

Sinar had this in stills in 2005.

You calibrate a monitor to match outputs, it seems daft to not match inputs particularly when you have to match shots from [expensive] multiple cameras.

[John Brawley]

I believe that you are wrong. Of course a Bayer CFA is not a "4:2:0" format, but there is some value to comparing the two.

So we're in agreement that using "chroma sub-sampling" terminology is not the correct language to use when describing a bayer sensor performance. You're wanting to "compare them" using the wrong terminology.

http://en.wikipedia.org/wiki/Chroma_subsampling
(above link has no mention at all of Bayer sensors)

http://en.wikipedia.org/wiki/Bayer_filter
(Above link has no mention of 4:4:4 or 4:2:2)

I'm not really disagreeing with the point you're actually trying to make about differences in the way chroma is "captured" but using chroma sub-sampling terminology is the wrong way to make your point, which is I believe, that there are less blue and red pixels relative to green in a given bayer sensor.

I presume that you know what those actual numbers mean in that ratio. 4:4:4 ? Or 4:2:2 ?

They refer to the way the video is "encoded". The first number is the brightness information and is also the "green" channel. The second two numbers are the "colour difference" signal. So they contain no brightness information, only colour information. You "re-create" the colour gamut by summing and subtracting these two signals from the Lumininace / green channel.

This video encoding is also known as YUV or YCbCr or component.  

http://en.wikipedia.org/wiki/YUV

This is why it's described as "encoded" video. It's NOT the same as RGB. Apple only recently created an actual ProRes RBG codec option, called ProRes 4444 which offers the user the choice of RBG or traditional YUV encoding.

A bayer sensor doesn't work that way. At all. There is no "encoding" into this colourspace until the sensor data is transcoded to a "video" format.  There is not separate brightness channel and colour difference channels.  A sophisticated algorithm "creates" The colour information and although there are differences in the sampling by virtue of the fact that there are generally twice as many green pixels as any other, it's only co-incidence that it falls into the same ratio as 4:2:2 chroma sub-sampling.

I'm being pedantic about this because your next leap of logic is really a gross simplification and in my opinion gives a misleading idea about Bayer sensors having "half" the chroma resolution.

A Bayer sensor of 1920x1080 sensels will have approximately 2 million samples.

A 1920x1080 4:4:4 image frame will have approximately 6 million samples.

I see it differently.

At 1920 x 1080 there are a fixed number of pixels.  And this is the point. You're throwing pixel resolution into a discussion about colour fidelity and making a direct leap to the ratio of RGB pixels and equating that to "colour" resolution.

The inference of the sensor only having "960x540" worth of blue pixels doesn't make sense because you never ever only have the blue channel, because nothing we shoot is ever so highly monochromatic.  Even BLUE LED's would have a range of "blue"  And for that matter, even the "blue" pixels have a huge overlap of sensitivity to the other colours.  It's not so highly monochromatic. If it was, then we'd have a very odd looking image indeed.

It also discounts the "mathematics" of debayering.  You only have to look at the visual differences between ACR, C1 and even Resolve to see that those algorithms can make a large difference to the way the image presents in post.

So while I would absolutely accept there *IS* a difference in the chroma resolution, you can't use video encoding terminology as it over simplifies what's actually going on and leads to these kinds of numbers being kicked around and it leads to silly conclusions about cameras not being good enough for chroma keying for example.

Nobody ever talks about bit depth and I think that has a bigger vector on captured colour gamut than the sample size of the bayer sensor.  In this case the BMPCC is greater than 16 bit at the sensor, 16 bit linear internally before going to 12bit LOG when recording to DNG or 10 Bit ProRes (422 HQ).  In fact they unpack as 16 bit files again in Resolve or ACR once you start to work from the DNG's.

Yes there is a case for oversampling with Bayer sensors, and that's exactly the thinking behind a camera like the Sony F65 having an 8K sensor for 4K raw files.  Can you name for me another RAW camera that oversamples in this way to address this "problem" of reduced chroma resolution ?  The Panavision Genesis (Sony) is the only other one I can think of, though it's not really a RAW camera.  

RED oversample and Arri oversample somewhat but not enough to deliver 4K cinema files that are flawed by the logic you're presenting.

jb