AF-100 Wishlist/Suggestion Discussion Thread

[Bruce Watson]

The 3700 does have variable frame rates so all it misses is slow motion. But how important is slow motion? When it's 10x it allows you to shoot special effects in scale or capture an action when the emphasis is on the physics (sports, etc). When it's only 2x slow motion, most hollywood films use frame duplication instead of real slow motion, it works better because it lets you see each frame for more time, and, as I would put it, it puts less emphasis on the motion and more on the emotion. True slow motion is like watching physics in action, it works well for a car ad or anything that emphasizes technology, but frame duplication works better for drama. Frame duplication actually stops time for 2x instead of showing time with 2x detail. A single photograph or even a duplication of a frame for 5 seconds can work dramatically very well, while a 1000fps shot will easily become a caricature of the subject. When the brain is given enough time it focuses on a face or an object/form as a whole. When slow motion is present it focuses on every single insignificant detail, dust, particles, wardrobe in motion, anything that moves fast from one frame to the next. Impressive but usually inappropriate.

+1. Excellent point.

 

[Graeme_Nattress]

Chroma sampling notation is not useful to describe images beyond what it was designed for, which is to show the relative subsampling of CbCr chroma compared to Y in a YCbCr system. Back when the notation was first used, it did tend to mean something for comparison purposes just because most camera systems were very similar - three chips, prism, and YCbCr video recording with chroma sub-sampling.

Today we have a vast variety of ways of making a high quality image, and not all those methods convert to YCbCr or chroma subsample. And even if you have a "4:4:4" camera, it's no guarantee of any quality because chroma sub-sampling notation only refers to relative resolutions, and does not tell us how well those channels align with each other. Some 3 chip and prism systems use a "precision offset" to improve the quality at 4:2:2, but actually reduce it when they record 4:4:4, and other systems that use a single chip and RGB stripe array do do 4:4:4, but the colour channels don't align at high details thus leading to a rainbow moire effect.

To say a Bayer pattern sensor cannot deliver it's nominal rated pixel resolution is true, but such a caveat also applies to all sensor types due to a simple application of sampling theory and a desire to avoid aliasing artifacts in images. You can make a sensor see it's rated resolution, but you will also see artifacts. It's not a free lunch.

A bayer pattern sensor camera is indeed hard to quantify, but only because it's tricky to ascribe a single number to an image quality factor that is multi-dimensional. When all professional cameras had the same basic technology, a single number could describe the difference between them quite well when looking at a parameter like "resolution", but today, that just doesn't work.

To go back to the point about chroma sub-sampling referring to YCbCr format data, the sensor on a bayer pattern camera is RGB - it doesn't have luma and chroma pixels. The green is generally close to luma (it's still green though), but the red and blue are indeed red and blue. It's not the same thing as YCbCr. In YCbCr you start with full resolution Y and Cb and Cr and then "throw away" some of the Cb and Cr resolution.

Now, with any camera system you can measure the resolution of the resulting image. In YCbCr 4:2:0 you know the resolution of the chroma will be always at most half that of the luma horizontally and vertically. It could be worse, but it will never be more. But 4:2:0 only describes the recording, not the camera system as a whole, so you could be in a situation where the resolution of the luma going into the compression is already lower than it should be. It's still "4", but if you measured it, it could be a 3, or 2.5. The chroma may be 2:0, but that says vertical and horizontal resolution are the same, but when you measure it, they may be very different horizontally than vertically. You recording may be 4:4:4, but it may be 4:4:4 where the RGB don't line up right. 4:4:4 should measure the same horizontally and vertically for luma and chroma (or R,G & B ) but depending on how the camera works, it could actually measure very differently horizontally than vertically.

In the end, 4:2:0 or 4:4:4 tells you if chroma sub-sampling has occurred or not. That's it. It doesn't tell you anything about the resulting image quality of the final image.

So, what can you expect from a Scarlet 2/3" at 3k. Luma resolution will be around 2.5k, with strong MTF at 2k resulting in a very detailed image with minimal aliasing or moire. The exact resolution you get for any particular colour will vary. It will, however, not drop below 1.5k and depending on how much luma detail there is in that colour, it will between 1.5k and 2.5k, again, with minimal moire. But it would be wrong to use chroma sub-sampling notation to describe such an image as no chroma sub-sampling has occurred. After demosaic, the RGB colour channels will however line up. And the image will be gorgeous.

The way to think of chroma sub-sampling is that it's a particular (and simple) form of image compression. If a camera system does not use that particular system, how can you say it has a chroma sub-sampling notion x:y:z ?

Graeme

 

[Camera Expert]

Real slow motion may not be important to some people but it is to others. I mean not everybody will shoot the same way and when talking specifically about the AF100, I'll be bought for many different reasons.

 

[Gabrobot]

The 3700 does have variable frame rates so all it misses is slow motion. But how important is slow motion? When it's 10x it allows you to shoot special effects in scale or capture an action when the emphasis is on the physics (sports, etc). When it's only 2x slow motion, most hollywood films use frame duplication instead of real slow motion, it works better because it lets you see each frame for more time, and, as I would put it, it puts less emphasis on the motion and more on the emotion. True slow motion is like watching physics in action, it works well for a car ad or anything that emphasizes technology, but frame duplication works better for drama. Frame duplication actually stops time for 2x instead of showing time with 2x detail. A single photograph or even a duplication of a frame for 5 seconds can work dramatically very well, while a 1000fps shot will easily become a caricature of the subject. When the brain is given enough time it focuses on a face or an object/form as a whole. When slow motion is present it focuses on every single insignificant detail, dust, particles, wardrobe in motion, anything that moves fast from one frame to the next. Impressive but usually inappropriate.

I disagree, the brain always focuses heavily on faces, and slow motion doesn't change that fact. The difference between frame duplication and true slow motion is that frame duplication emphasizes a picture style analysis, and slow motion emphasizes the motion. Neither is innately more emotional than the other...emotions are not static, they are in motion. I really enjoy shooting people's faces in slow motion, especially when they're full on doing something and focused on it too much to be self-conscious. I shot part of one music video in 50% slow motion while the band played twice as fast, and the intensity of the emotion (especially on the lead singer) was really amazing.

 

[Duke M.]

So, what can you expect from a Scarlet 2/3" at 3k. Luma resolution will be around 2.5k, with strong MTF at 2k resulting in a very detailed image with minimal aliasing or moire.

@Graeme

I've seen a couple of people recently quoting the 3k (3072) Scarlet as having true resolution of chroma at 2.5k. (Of course we don't know yet.)

Originally on Scarletuser people were speculating at a 20-30% loss in resolution due to the debayer process on a single chip camera, because that's where it is on most cameras of that type. (And generally closer to 30% loss.) At the time they were saying 2.2-2.3k resolution.

This is just for curiosity, because the expectation will be a 2k output, but at a minimal 20% loss we are at 2457 in true resolution. Any over sampling is usually good, but... How much over sampling?

Has something changed to make people now believe 18.5% resolution loss is to be expected? Or are people just playing what if again?

 

[Otis Grapsas]

I disagree, the brain always focuses heavily on faces, and slow motion doesn't change that fact. The difference between frame duplication and true slow motion is that frame duplication emphasizes a picture style analysis, and slow motion emphasizes the motion. Neither is innately more emotional than the other...emotions are not static, they are in motion. I really enjoy shooting people's faces in slow motion, especially when they're full on doing something and focused on it too much to be self-conscious. I shot part of one music video in 50% slow motion while the band played twice as fast, and the intensity of the emotion (especially on the lead singer) was really amazing.

Hm we are out of topic again Skip this post guys!

I'm only expressing a school of though that is less popular on the forums. I know most people believe slow motion is the best thing since sliced bread, it has sold tons of cameras for sure. But some cinematographers with very high budget and a solid grasp of the art consider slow motion to be a cheap looking solution, and when it involves violence it can even lead to an exploitation look. In any case, there are far more interesting ways to treat time in film.

With slow motion you can exploit a scene to the maximum. Think of the Matrix, in spaghetti. You can get multiple angles of the expensive explosion, or the cowboy going through the sugar glass window, you can have them in slow motion and you can also put all shots in the final cut, one after the other, total coverage.

But the exact same thing works with excellent results in Zabriskie Point. Explosions from multiple angles for minutes. It has a meaning and a message. It's not for show, it's not cheap.

Have you seen Henry, the Portrait of a Serial Killer? It's been more than 10 years, But I remember some shots with little or no motion. Crime scenes, you just see the victim and the surrounding and hear what happened in audio. Frozen time and a powerful memory.

Taxi Driver has an entire sequence that shows a single moment in time while the camera exits the crime scene. That's what makes it powerful, the lack of subject motion, the only thing that moves is the camera and that's as slow as possible. Everything is frozen, death is in the air, the music is a slow beat that brings a funeral in mind.

Any photographer can tell you that a single photo can have great impact. A single photo of a soldier can be more powerful than an entire documentary as a vehicle for expressive one's feeling against the war. It can touch a viewer more than a 500fps slow motion death of a soldier, because it shows something, it does not exploit it. There is no motion to follow, we have time to see his eyes and his body effected by the forces of politics. Even the milliseconds of extra time given by frame doubling is a lot. It allows the user to examine the subject to greater extend. That's why some people use it.

Regarding the video clip, I'm surprised it turned out good, going faster usually leads to a rushed performance, with less control and authority. You could try recording a band at half tempo. It might look great and it could even improve their real time performance

 

[Gabrobot]

Hm we are out of topic again Skip this post guys!

I'm only expressing a school of though that is less popular on the forums. I know most people believe slow motion is the best thing since sliced bread, it has sold tons of cameras for sure. But some cinematographers with very high budget and a solid grasp of the art consider slow motion to be a cheap looking solution, and when it involves violence it can even lead to an exploitation look. In any case, there are far more interesting ways to treat time in film.

With slow motion you can exploit a scene to the maximum. Think of the Matrix, in spaghetti. You can get multiple angles of the expensive explosion, or the cowboy going through the sugar glass window, you can have them in slow motion and you can also put all shots in the final cut, one after the other, total coverage.

But the exact same thing works with excellent results in Zabriskie Point. Explosions from multiple angles for minutes. It has a meaning and a message. It's not for show, it's not cheap.

Have you seen Henry, the Portrait of a Serial Killer? It's been more than 10 years, But I remember some shots with little or no motion. Crime scenes, you just see the victim and the surrounding and hear what happened in audio. Frozen time and a powerful memory.

Taxi Driver has an entire sequence that shows a single moment in time while the camera exits the crime scene. That's what makes it powerful, the lack of subject motion, the only thing that moves is the camera and that's as slow as possible. Everything is frozen, death is in the air, the music is a slow beat that brings a funeral in mind.

Any photographer can tell you that a single photo can have great impact. A single photo of a soldier can be more powerful than an entire documentary as a vehicle for expressive one's feeling against the war. It can touch a viewer more than a 500fps slow motion death of a soldier, because it shows something, it does not exploit it. There is no motion to follow, we have time to see his eyes and his body effected by the forces of politics. Even the milliseconds of extra time given by frame doubling is a lot. It allows the user to examine the subject to greater extend. That's why some people use it.

Regarding the video clip, I'm surprised it turned out good, going faster usually leads to a rushed performance, with less control and authority. You could try recording a band at half tempo. It might look great and it could even improve their real time performance

Well, I'm not arguing that any technique is innately better than another...there are many ways of telling a story. Some people would argue that capturing anything visually is exploitative...it's all pretty subjective.

EDIT: I remember reading an essay by Maya Deren about slow motion that explained it really well...I definitely recommend reading what she wrote about it.

As far as the band speed, it's actually harder to play slowly (of course this is relative and isn't always true depending on the base tempo), playing fast forces you to have faith in your body and leaves no room for second guessing. As a drummer I can vouch for that...a slow tempo means you have to be mind numbingly precise, whereas a faster tempo lets muscle memory do all the work. It's that removal of inhibitions that I look for when having people play faster for slow motion...same reason I like having models do some kind of major movement and use a high shutter speed to capture them when doing stills. It peels away the layers we build up to make us look like everyone else, and reveals the quirks that make us unique.

 

[Jay Birch]

Not sure if this has been suggesting for the AF-100...

HDMI 1.4

Now i'm just speculating (i'm not technical at all).... but as HDMI 1.4 can port 4K video.... theoretically, depending on where the HDMI is located in the chain.... could we not find a way to extract that 4K stream before it gets downsized to 1080p?

Tech guys, don't fire me down too much.... i'm just guessing (or hoping would be a better word!)

 

[Barry_Green]

The AF100 has a 4K chip, but it doesn't process it internally as 4K. All the internal processing is almost certainly going to be done at 1080p, just like all their other video cameras do. I doubt the chip will even be read at 4K, it will probably be pixel-binned to take maximum advantage of the sensitivity gains possible, and keep the total amount of individual data pixels read to a manageable level.

 

[AGMedia]

I assume this is in the "wish list" somewhere in these 99 pages, and I'm sorry if I'm beating a dead horse -- but this is important.

DEAR PANASONIC: As an almost certain purchaser of this AF-100 camera, I request the fastest possible refresh rate on the CMOS sensor. The faster the refresh on the CMOS sensor, the less the skew/jello, and the more viable this camera becomes for use in the next generation of digital cinema.

Please -- push that refresh rate to the max.

 

[JMG100]

Refresh Rate

Refresh Rate

I assume this is in the "wish list" somewhere in these 99 pages, and I'm sorry if I'm beating a dead horse -- but this is important.

DEAR PANASONIC: As an almost certain purchaser of this AF-100 camera, I request the fastest possible refresh rate on the CMOS sensor. The faster the refresh on the CMOS sensor, the less the skew/jello, and the more viable this camera becomes for use in the next generation of digital cinema.

Please -- push that refresh rate to the max.

I second that. The refresh rate needs to at least be comparable to an EX1. I know that may be a tall order due to the huge sensor, but it will definitely have an impact on whether I buy the camera or not.

 

[Rick Burnett]

The AF100 has a 4K chip, but it doesn't process it internally as 4K. All the internal processing is almost certainly going to be done at 1080p, just like all their other video cameras do. I doubt the chip will even be read at 4K, it will probably be pixel-binned to take maximum advantage of the sensitivity gains possible, and keep the total amount of individual data pixels read to a manageable level.

With regards to pixel binning, if you are using a Bayer pattern, how exactly does this work when any two adjacent pixels are different colors? Do they bin on the diagonal?

I've wondered this for awhile. My research of pixel binning shows there are certain non-bayer filter patterns that some companies use to make binning easier, but I cannot find a lot of info on how in hardware pixel binning is done on a bayer pattern, and what specific gains there are and what disadvantages there are as well.

I imagine there are a host of different ways to do it. Thoughts?

 

[Graeme_Nattress]

With regards to pixel binning, if you are using a Bayer pattern, how exactly does this work when any two adjacent pixels are different colors? Do they bin on the diagonal?

I've wondered this for awhile. My research of pixel binning shows there are certain non-bayer filter patterns that some companies use to make binning easier, but I cannot find a lot of info on how in hardware pixel binning is done on a bayer pattern, and what specific gains there are and what disadvantages there are as well.

I imagine there are a host of different ways to do it. Thoughts?

In binning, you'd only bin pixels of the same colour - essentially think of making bigger pixels by joining together smaller ones. Now that you've effectively got bigger pixels, what was the right amount of optical low pass filtering for the smaller native pixels is no longer the right amount for the larger binned pixels. Binning itself is therefore a very poor downsampling filter.

Usually the advantage of binning is that it's done on chip, so there's a lot less data to read out, and therefore a bottle-neck is reduced. If you're reading out the whole sensor and post binning you're doing it because it's the 2nd cheapest downsample filter there is after "nearest neighbour".

Graeme

 

[Graeme_Nattress]

@Graeme

I've seen a couple of people recently quoting the 3k (3072) Scarlet as having true resolution of chroma at 2.5k. (Of course we don't know yet.)

Originally on Scarletuser people were speculating at a 20-30% loss in resolution due to the debayer process on a single chip camera, because that's where it is on most cameras of that type. (And generally closer to 30% loss.) At the time they were saying 2.2-2.3k resolution.

This is just for curiosity, because the expectation will be a 2k output, but at a minimal 20% loss we are at 2457 in true resolution. Any over sampling is usually good, but... How much over sampling?

Has something changed to make people now believe 18.5% resolution loss is to be expected? Or are people just playing what if again?

It's not so much resolution loss though - it's resolution you shouldn't have anyway because to acquire such resolution leads to excessive aliasing which is ugly. All cameras are in very much the same position. Look at the F35 which is soft horizontally, and over-sharp vertically. It measures a full 1080 lines vertically, but at the expense of aliasing that can show up on broadcast when you shoot people who have fine detail on their clothing.

Empirically, based on our RED One experience, you can expect ~80% of the rated pixel resolution, with negligible aliasing, which will make for a very nice image. Even on a 3 chip system, if you push your resolution beyond that 80% figure, you'll get into the land of objectionable aliasing due to optical filters not being "fast" - something to do with the lack of negative photons (darkons?).

Graeme

 

[Barry_Green]

In binning, you'd only bin pixels of the same colour - essentially think of making bigger pixels by joining together smaller ones. Now that you've effectively got bigger pixels, what was the right amount of optical low pass filtering for the smaller native pixels is no longer the right amount for the larger binned pixels. Binning itself is therefore a very poor downsampling filter.

Yes, BUT -- the difference here is that the camera isn't going to be used at its native resolution for stills. So the OLPF being designed for a binned sensor, could result in very lovely footage after all. You won't get the 4K res, but you're not going to get that anyway because it's a 1080p system.

Even if they only binned adjacent green pixels, they'd still cut the amount of sensor data down dramatically, while keeping it very high res.

 

[Barry_Green]

With regards to pixel binning, if you are using a Bayer pattern, how exactly does this work when any two adjacent pixels are different colors? Do they bin on the diagonal?

As Graeme said, you only bin those of same color, so that would mean skipping and binning across the gaps. Which, in general, would be a pretty awful thing to do. However, in this particular case, I think it could work, specifically because the sensor has 6x as many pixels as the destination frame size. Those pixels will go to waste unless they can find a way to use them, and binning might be that way.

The gain for binning is to increase the sensitivity and lower the noise by creating effectively larger "super-pixels". There's been some discussion of Panasonic implementing something they called a 4K2K sensor... that might be exactly what this is, a 4K sensor that delivers 2K resolution images. If they binned two green pixels together, and then used an appropriate-strength OLPF to blur the detail just enough, they could end up with an effectively 4:4:4 2K+ sensor, because that chip would have 2176 horizontal red and 2176 horizontal blue pixels on each line, and 2176 green super-pixels, created out of the 4352 pixels on each line. With similar performance in the vertical.

 

[Graeme_Nattress]

The complexity comes in that now that as we see, bigger pixels leads to a new OLPF. In a design that is only for that lower resolution, the OLPF is tuned for that resolution, but you'll still have the very same issue that you can't get your rated pixel resolution unless you accept aliasing, and how much you accept is a complex thing to get right. If you're going to use an over-sampled sensor only for an over-sampled image, the very best thing to do is a full adaptive demosaic followed by a high quality downsampling filter. That can lead to superb results.

Just binning both greens in the bayer macro pixel gives you a 1/4 bandwidth saving, which is nice, but not the significant saving of 2/3 that basic line skipping gives you. Binning probably doesn't help skew, whereas line skipping helps it a lot (by 2/3 again).

But when it comes down to it - almost everything you do is a compromise among image quality, size, weight, power consumption, heat and especially price.

One thing I'm getting increasingly unhappy with in images is when image quality is different horizontally and vertically, which I'm noticing in images and it bugs me I really like to keep everything symmetrical!

Graeme

 

[Rick Burnett]

I see, I guess it also depends on the sensor and how much binning you are doing. For instance, taking the 7D sensor you'd have to skip every 2 of 3 lines, do they bin multiple pixels? Like:

GRG
BGB
GRG

In that pattern, if line skipping picked just the center G, a binning approach could use all 5 greens, or if you were looking at R and B, two of each.

One question I would have is, does the binning process take as long as reading 3 lines? I know line skipping can give you 1/3 the time in this approach, but depending on how the binning occurs, I could see it being 2/3 the time of a full approach (giving at least one reads worth of time to combine the results).

I'd really like to see some circuitry that does this! My EE side is frothing for knowledge.

Given that the AF100 will have a stronger OLPF, it will be interesting to see how much softer the image looks. I say this in regards to the fact that anything you see in the pictures you are taking now that is aliased is going to have to get softer. People have commented that if a strong enough OLPF were used on say the 7D, the image would be so soft, people wouldn't like it.

Given the density of R and B pixels in a given 9 pixel area, I think you'd have to still have a pretty strong OLPF which would really soften the image (on the 7D).

On the GH1, how many lines are pixel binned into 1?

 

[Barry_Green]

Well, again, in this specific application, it's an unusual scenario. Red has a different design priority, in that you want stills and video at the same resolution. DSLRs have a different design priority, which is a $799 pricetag. But the AF100 is looking to get the best 1080p image it can, out of a 4.3k sensor.

Skipping rows is a way to go, which is what we all presume the DSLRs are doing. But they *have* to do that, because they have to preserve the full resolution for the stills mode. The AF100 doesn't have to do that. All it needs is 1080p. So if it starts with an oversampled 2160 and downsamples to 1080p, they might end up with a fantastic 4:4:4 1080p image for internal processing.

Furthermore, as far as savings go, there's not really all that much reason for them to even have to go with the 4:4:4 aspect, because the recording format is 4:2:0. So they could bin a couple of reds together, and a couple of blues together, and end up with 1080 chroma samples per line, and 2160 luma samples, and they've slashed the data rate in half, while giving up *nothing* in terms of ultimate image quality as per its recorded format.

If you then slice off a third of the chip for the 16:9 shape, instead of its native 4:3 shape, you chop off the necessity to read a whole bunch of lines that aren't doing anything for you anyway. I think it's possible, even practical, to say that the combination of properly-engineered binning and ignoring unused areas of the chip could very well result in a fast-enough read off the chip that they won't have to resort to pixel-skipping.

Granted they won't get 120fps or 4K out of it, but then again they weren't going to get that anyway. As you said, it's all about the compromise of heat vs. price vs. imager size vs. resolution vs. aliasing vs. everything, all together. You could get a Red One with an MX and get 4.5k and all that, but at $23,000. Or you could get a Scarlet and get 3K and all that, but with 2/3" chips and less-than-16mm DOF. Or you could get an AF100 and get 35mm-style DOF and a hopefully pristine alias-free 1080p image that should rival the EX1/HPX370 for sharpness, but you don't get 120fps or 4K res.

For me, I'm plenty happy with that. EX1/HPX370 sharpness, 35mm cine DOF, interchangeable lenses, PL-mount options, pro audio, variable frame rates, less than $5995. I'm not arguing! :thumbsup:

 

[Barry_Green]

Given that the AF100 will have a stronger OLPF, it will be interesting to see how much softer the image looks. I say this in regards to the fact that anything you see in the pictures you are taking now that is aliased is going to have to get softer. People have commented that if a strong enough OLPF were used on say the 7D, the image would be so soft, people wouldn't like it.

That's because the 7D is designed for stills, and has a video mode grafted on. Its stills resolution is the priority. The video mode delivers barely 720p of resolution. If you used an appropriate OLPF for the 7D, you would end up with a 720p camera, that has to record in 1080p mode to get that 720p.

With the AF100 we're talking about something that's taken into the design shed by the video engineers, starting from the perspective that stills are irrelevant. This unit will be tuned to deliver the best video image they can make. I fully expect HPX370/EX1 sharpness out of it.

On the GH1, how many lines are pixel binned into 1?

Nobody knows, outside of the engineers inside Panasonic Japan. You'll get people making guesses, even educated guesses, but that's all. Nobody actually knows.