4K JVC Consumer Camera with 3D at CES 2011- Red Scarlet blushes...

[Steve Castle]

My understanding of the AF100 is that it doesn't line skip, it pixel bins. Which allows you to use that lost area.

Although, I hope you noted that I said "Size isn't all that matters" which you go on to explain the other possibilities for extended DR using back illumination. I absolutely agree that it is a factor, but just one factor.

I also don't see the photosites as massive in the Arri Alexa, they are around 8.25um maximum, whereas the 5Dm2 is 6.4um square. But, look at the Red MX sensor, it has 5.4um square, smaller than the 5Dm2 (and I am talking about using the 5D in photo mode, not movie mode) and it has around 13 stops (not including the HDRx stuff). That's pretty amazing. Granted, DR does not have an industry standard as you say.

There is A LOT of magic that goes into making a sensor. Basing the sensitivity just on the sensor size and resolution is NOT an indication of what you are going to get. So much depends on other factors of the design, including the AD conversion, noise in CMOS circuits, special methodologies in the CMOS technology.

I mean, let's even take it a step further, the M sensor has less dynamic range than the MX sensor, and they are the same exact size. Point being, trying to judge how this camera from JVC is going to behave at 4K just isn't possible till you know everything they are doing. They may be employing methods, just as red has proven, that can allow higher resolutions and higher DR by following different methodologies, which may in fact, not work as well at a larger size. Don't forget, the larger the active area, the longer it is going to take to drain out the current as well. This might be why Red's read/reset time on the Red is so fast.

The AF100 doesn't pixel-bin, in fact, pixel-binning is very difficult for a rolling shutter CMOS sensor since it needs to capture multiple photosites simultaneously. In the past, pixel-binning has been reserved for CCD sensors with global-shutters that capture the entire frame and are easier to 'bin'. For CMOS sensors to do pixel-binning you need to use summation techniques that require additional logic and are significantly more complex. However, the AF100 does use video-centric OLPF that does minimizes the aliasing issues associates with dSLR sensors.

As far as reading into claims of how many 'stops' a sensor has in DR a lot of that can be deceptive, if not down-right dishonest. You'll get different results based on what test you do; Stouffer, DSC, Sekonic, DRTC. Worse, you can further manipulate the results based on how a test is conducted and what you want the results to be. We have companies fabricating their own in-house methodologies on how to measure DR and using it in their advertising material.

Also, as a CMOS sensor we know exactly what the sensor is doing and what this FalconBrid is doing. ADC is done on sensor, and modern CMOS sensors distributes current over multiple ADCs and isn't an issue for large photosites. Interpolation is done off-sensor on the FalconBrid, hence for this camera's DR is also a function of the gamma curve that is being applied downstream from the sensor.

 

[Gabrobot]

The AF100 doesn't pixel-bin, in fact, pixel-binning is very difficult for a rolling shutter CMOS sensor since it needs to capture multiple photosites simultaneously. In the past, pixel-binning has been reserved for CCD sensors with global-shutters that capture the entire frame and are easier to 'bin'. For CMOS sensors to do pixel-binning you need to use summation techniques that require additional logic and are significantly more complex. However, the AF100 does use video-centric OLPF that does minimizes the aliasing issues associates with dSLR sensors.

Problem with that theory is the GH2...which also appears to lack moire and aliasing. Something about Panasonic's new sensor allows them to get a good 1080p image, and it can't just be a different OLPF. Whether that's pixel binning or not is another question, but it does appear they have a way to do some type of downscale as opposed to simply line skipping. But this shouldn't necessarily be a surprise, Panasonic's other CMOS sensor cameras can do something similar when doing 720p on a 1080p sensor. This appears to be unique to their CMOS sensors since even their high-end CCD full raster 1080p cams can't do 720p.

As far as reading into claims of how many 'stops' a sensor has in DR a lot of that can be deceptive, if not down-right dishonest. You'll get different results based on what test you do; Stouffer, DSC, Sekonic, DRTC. Worse, you can further manipulate the results based on how a test is conducted and what you want the results to be. We have companies fabricating their own in-house methodologies on how to measure DR and using it in their advertising material.

It's been documented by quite a few people that the MX sensor has significantly better DR than the M sensor. Specific numbers vary as with any cam, but relative comparisons seem consistent.

 

[Rick Burnett]

The AF100 doesn't pixel-bin, in fact, pixel-binning is very difficult for a rolling shutter CMOS sensor since it needs to capture multiple photosites simultaneously. In the past, pixel-binning has been reserved for CCD sensors with global-shutters that capture the entire frame and are easier to 'bin'. For CMOS sensors to do pixel-binning you need to use summation techniques that require additional logic and are significantly more complex. However, the AF100 does use video-centric OLPF that does minimizes the aliasing issues associates with dSLR sensors.

As far as reading into claims of how many 'stops' a sensor has in DR a lot of that can be deceptive, if not down-right dishonest. You'll get different results based on what test you do; Stouffer, DSC, Sekonic, DRTC. Worse, you can further manipulate the results based on how a test is conducted and what you want the results to be. We have companies fabricating their own in-house methodologies on how to measure DR and using it in their advertising material.

Also, as a CMOS sensor we know exactly what the sensor is doing and what this FalconBrid is doing. ADC is done on sensor, and modern CMOS sensors distributes current over multiple ADCs and isn't an issue for large photosites. Interpolation is done off-sensor on the FalconBrid, hence for this camera's DR is also a function of the gamma curve that is being applied downstream from the sensor.

But you are assuming a row dependent pixel bin, nothing is saying they can't use a different filter layer and pixel bin pixels within a row, which may be entirely possible and not that prohibitive. Further, Jan the product manager has stated that the AF100 DOES NOT use line skipping. Clearly both Canon (which also by the way claim they don't actually skip lines) have two different methods as their moire is very different, and is not just a factor of the OLPF. The coloring of the Canon moire is something else in their downsample algorithm.

Sure, it is one thing for a company to boast about the DR of their product, but it's another to listen to people talk about the actual usable dynamic range they get OR shoot the same cameras on the same test charts. The dynamic range of the M and MX sensor are different, and the MX sensor is better, regardless, and they are THE SAME pixel size. In fact, they even upped the default sensitivity of the sensor because it's so much more sensitive. Again, same sensor size, same pixel size, different results. This is why comparing a RED to an ARRI to a CANON to a SONY to a PANASONIC isn't just pixel size or sensor size. There is so much more to go on.

Pertaining to modern CMOS sensors, the larger your sensor, the higher the resistance in the lines and the further they have to travel to the ADC, it's that simple. You don't want to make larger wires because you don't want to block off active area. When I refer to the current being read out, I mean in terms of a single site. If you read out a row in parallel of a larger site, it's going to have a wider range of values for what it can equal. The time that you'll need to read the value and more importantly reset the photosite, which is draining all the current out of it, is going to take longer. This is going to impact your read/reset times for that line. Now, given you have less rows, maybe it'll be the same. But comparing the rolling shutter on the 5D and the 7D, the 7D is quite a bit faster in tests. (25ms compared to 21ms) Which makes sense to me.

Having worked with footage from the Red One, and how amazing it was, Red really changed my mind about resolution. But, the caveat is, I also thing resolution isn't everything because these sensors are designed by different teams and their operations can be totally different. I like the freedom 4K gives. Even when working at 1080p as my final resolution, with 4K, if I want to crop in on a shot, I wont lose resolution in the final. Whereas, if I am starting with a camera that resolves 800 lines, and I try and crop in on shots, I start to seriously degrade my footage immediately.

 

[Steve Castle]

But you are assuming a row dependent pixel bin, nothing is saying they can't use a different filter layer and pixel bin pixels within a row, which may be entirely possible and not that prohibitive. Further, Jan the product manager has stated that the AF100 DOES NOT use line skipping. Clearly both Canon (which also by the way claim they don't actually skip lines) have two different methods as their moire is very different, and is not just a factor of the OLPF. The coloring of the Canon moire is something else in their downsample algorithm.

Sure, it is one thing for a company to boast about the DR of their product, but it's another to listen to people talk about the actual usable dynamic range they get OR shoot the same cameras on the same test charts. The dynamic range of the M and MX sensor are different, and the MX sensor is better, regardless, and they are THE SAME pixel size. In fact, they even upped the default sensitivity of the sensor because it's so much more sensitive. Again, same sensor size, same pixel size, different results. This is why comparing a RED to an ARRI to a CANON to a SONY to a PANASONIC isn't just pixel size or sensor size. There is so much more to go on.

Pertaining to modern CMOS sensors, the larger your sensor, the higher the resistance in the lines and the further they have to travel to the ADC, it's that simple. You don't want to make larger wires because you don't want to block off active area. When I refer to the current being read out, I mean in terms of a single site. If you read out a row in parallel of a larger site, it's going to have a wider range of values for what it can equal. The time that you'll need to read the value and more importantly reset the photosite, which is draining all the current out of it, is going to take longer. This is going to impact your read/reset times for that line. Now, given you have less rows, maybe it'll be the same. But comparing the rolling shutter on the 5D and the 7D, the 7D is quite a bit faster in tests. (25ms compared to 21ms) Which makes sense to me.

Having worked with footage from the Red One, and how amazing it was, Red really changed my mind about resolution. But, the caveat is, I also thing resolution isn't everything because these sensors are designed by different teams and their operations can be totally different. I like the freedom 4K gives. Even when working at 1080p as my final resolution, with 4K, if I want to crop in on a shot, I wont lose resolution in the final. Whereas, if I am starting with a camera that resolves 800 lines, and I try and crop in on shots, I start to seriously degrade my footage immediately.

The fact is that the AF100 doesn't have the capability to process the entire sensor resolution at 60 frames a second, if it could this GH sensor based camera wouldn't be limited to a mere 1080p. The aliasing issues on dSLRs sensor based cameras is clear evidence that every single line of the sensor is not being read. The proof is in the pudding.

Also, modern CMOS sensors have ADCs that are usually column or area dependent. This means that each column/area has its own ADC and the current doesn't need to travel very far for each photosite and the time for each readout is shortened. There are literally thousands of ADCs on each sensor. As far as the MX, its probably using Correlated Double Sampling (CDS) to increase appearance of dynamic range. But again, its still a function of the actual readout of the sensor:

Dynamic Range = 20 × Log[Full Well Capacity (electrons)/ Read Value (electrons)]

The larger the photosite, the more electrons you get per well, the better the dynamic range. It improves everything downstream of that. Obviously the ADC and signal processing still play role, its not a mononuclear environment, but its still dependent on how well the camera captures light. This is the inescapable physics of it, its not the only thing that determines DR, but it is still the main factor.

As far as Red marketing talk, it never ends very civil so I'll avoid it, but If resolution is what you're after than this JVC also fits the bill. Personally, I think that it's a good thing. Different horses for courses, but it still comes at a sacrifice. Overall sensor size and the size of the photosite is still the fundamental characteristic that the camera is designed around.

 

[Gabrobot]

The fact is that the AF100 doesn't have the capability to process the entire sensor resolution at 60 frames a second, if it could this GH sensor based camera wouldn't be limited to a mere 1080p. The aliasing issues on dSLRs sensor based cameras is clear evidence that every single line of the sensor is not being read. The proof is in the pudding.

Which is why I brought up the GH2 which doesn't seem to suffer from moire and aliasing, and has an OLPF designed for stills. Proof is in the pudding. I don't know what exactly it's doing, but you can see the results.

Also, modern CMOS sensors have ADCs that are usually column or area dependent. This means that each column/area has its own ADC and the current doesn't need to travel very far for each photosite and the time for each readout is shortened. There are literally thousands of ADCs on each sensor. As far as the MX, its probably using Correlated Double Sampling (CDS) to increase appearance of dynamic range. But again, its still a function of the actual readout of the sensor:

Dynamic Range = 20 × Log[Full Well Capacity (electrons)/ Read Value (electrons)]

The larger the photosite, the more electrons you get per well, the better the dynamic range. It improves everything downstream of that. Obviously the ADC and signal processing still play role, its not a mononuclear environment, but its still dependent on how well the camera captures light. This is the inescapable physics of it, its not the only thing that determines DR, but it is still the main factor.

The "inescapable physics" are only cut and dried when comparing the exact same sensor technology. Otherwise, you can't say much without knowing how close the sensor is getting to theoretical efficiency. That is, a big leaky bucket's going to collect less water than a small one without leaks. But equally leaky (i.e. the same sensor technology from the same company) buckets will indeed collect water relative to their size.

Here's a comparison of M vs MX:
http://www.lightirondigital.com/LightIRON_Digital/LiDigital__MX_4K_TEST_CHARTS.html

The big difference between the sensors and the reason the MX has better DR is that it has vastly reduced noise. That is, it's a much less leaky bucket so it collects more water even though it's the same size.

 

[Steve Castle]

Which is why I brought up the GH2 which doesn't seem to suffer from moire and aliasing, and has an OLPF designed for stills. Proof is in the pudding. I don't know what exactly it's doing, but you can see the results.



The "inescapable physics" are only cut and dried when comparing the exact same sensor technology. Otherwise, you can't say much without knowing how close the sensor is getting to theoretical efficiency. That is, a big leaky bucket's going to collect less water than a small one without leaks. But equally leaky (i.e. the same sensor technology from the same company) buckets will indeed collect water relative to their size.

Here's a comparison of M vs MX:
http://www.lightirondigital.com/LightIRON_Digital/LiDigital__MX_4K_TEST_CHARTS.html

The big difference between the sensors and the reason the MX has better DR is that it has vastly reduced noise. That is, it's a much less leaky bucket so it collects more water even though it's the same size.

Its not 'leaky buckets', sensor technology does improve over time. This is a given. A 2.7mp Nikon D1 from ten years ago performs poorer than a 12MP Nikon D3s from today. But all things remaining equal, larger photosites do have better light sensitivity. I'm not sure what's so hard to understand about this. A 3.5K version of the MX sensor would have better DR and light sensitivity than 4.5K version of the sensor of similar size. The larger the photosite the more light you're capturing, there is more signal to work with, everything downstream of that benefits.

As Canon themselves advertises, "More megapixels. great for Photos, not for full HD video"...

Also in terms of making points of comparisons between manufacturer's DR claims is pretty useless as they are measured using different tests and different settings and conditions. Also, for a camera like this JVC, or any dSLR, your DR is limited by your gamma curve.

And BTW I'm not the one making DR comparisons between cameras in this thread...

In fact, I was pointing out the folly of comparing dSLR sensors, which aren't even using all their photosites while taking video, to a camera such as the Alexa, and that size does matter when it comes to photosites and dynamic range.

 

[Rick Burnett]

The fact is that the AF100 doesn't have the capability to process the entire sensor resolution at 60 frames a second, if it could this GH sensor based camera wouldn't be limited to a mere 1080p. The aliasing issues on dSLRs sensor based cameras is clear evidence that every single line of the sensor is not being read. The proof is in the pudding.

Also, modern CMOS sensors have ADCs that are usually column or area dependent. This means that each column/area has its own ADC and the current doesn't need to travel very far for each photosite and the time for each readout is shortened. There are literally thousands of ADCs on each sensor. As far as the MX, its probably using Correlated Double Sampling (CDS) to increase appearance of dynamic range. But again, its still a function of the actual readout of the sensor:

Dynamic Range = 20 × Log[Full Well Capacity (electrons)/ Read Value (electrons)]

The larger the photosite, the more electrons you get per well, the better the dynamic range. It improves everything downstream of that. Obviously the ADC and signal processing still play role, its not a mononuclear environment, but its still dependent on how well the camera captures light. This is the inescapable physics of it, its not the only thing that determines DR, but it is still the main factor.

As far as Red marketing talk, it never ends very civil so I'll avoid it, but If resolution is what you're after than this JVC also fits the bill. Personally, I think that it's a good thing. Different horses for courses, but it still comes at a sacrifice. Overall sensor size and the size of the photosite is still the fundamental characteristic that the camera is designed around.

You do realize that even the Red has aliasing and moire right? And it reads out every row, every column. The point is, most bayer sensor OLPF choices pick a value somewhere between the green density and the red/blue density so as not to totally make the image too soft. If the project manager from Panasonic tells me they are not line skipping, then they are doing something else. I agree that they may not be actually reading the values fully out of it, but somehow they claim to be using the data from the rows.

Pertaining to the column, this is exactly what I am talking about with the ADC travel distance. Even if you had an ADC at the top and bottom, you'd still have to travel half the height of the ADC at minimum at the worst case, which, would dictate your clocking speed for all the rows. Skew is constant, at least in all measured tests I have seen, so the larger the sensor, the more time a read/reset takes. Like I said, the 5D read/reset is longer than the 7D read/reset.

But regardless of all this, all I am saying is you can't look at a sensor design and immediately know that it will be more sensitive because the pixels are larger. There are other factors, including the age of the design, and new methodologies, coming out constantly, that are part of the process. EVERY person who I've read their posts that has upgraded their M sensor to MX who have done tests have shown how much better the results are. I don't see how you can argue with all these end users who are absolutely thrilled with how much better the sensor is. Trust me, I don't listen to the Red marketing talk. I'm an electrical engineer who does CMOS back end design for a living. I laugh when they say they've made the most complicated ASICs ever, because I know better, maybe the most complicated they've ever done.

I think it is a good thing as well, and I absolutely understand the trade-offs that you point out. It's getting that balance between resolution AND DR and sensitivity. The Alexa and the F3 look like they do a pretty good job, but, from the samples of the Epic (footage) I have worked with, I really like that extra resolution. Would I want more resolution than what the Epic has? No. Past what it has is overkill to me, but that is a personal preference. At the end of the day, it's what is important to each of us.

One thing we don't know is how the skew is on this camera (JVC). That would be my main concern. They really need to get the skew down to sub 10us on these cameras.

 

[Barry_Green]

The fact is that the AF100 doesn't have the capability to process the entire sensor resolution at 60 frames a second, if it could this GH sensor based camera wouldn't be limited to a mere 1080p.

Steve, let's be very clear on one thing -- you don't know what Panasonic is doing. You're making these proclamations of what they are or aren't doing, but you aren't a Panasonic engineer, you don't work in the company, you don't have access to their tech documents, so you don't know. You're guessing. You don't know that the sensor is "DSLR based" or "GH based". You don't know that it skips rows. You don't know whether or not it bins pixels. You don't know, really, anything about it. Nobody outside of a few engineers in a factory in Osaka know what it does.

The only things we know about the AF100 sensor are that it has 12.4m pixels, that it has a primary color filter, and that it is 17.8 x 10mm in size, and that it's MOS technology. That's it. We don't know that it's Bayer, or that it skips rows, or that it bins. I would think it's entirely possible that they could have designed a primary color filter specifically for this camera that would be maybe 2x2 for each pixel, allowing them to bin pixels in a 2x2 grid and giving them effectively a 3-megapixel Bayer chip with comparatively quite large pixels (about 5.7 microns).

Why would the AF100 be limited to a "mere 1080p"? Perhaps for the same reasons that the F3 and F35 and the Alexa and the HPX3700 are "limited" to a "mere 1080p." That's the video standard, and this is a video camera. There is no video standard beyond 1080p, there's no video transmission standard beyond 1080p, there's no recording format they have that goes beyond 1080p, there's no broadcast standard that goes beyond 1080p, and there's no delivery system beyond 1080p. Seems like 1080p is a perfectly reasonable choice.

 

[Barry_Green]

Which is why I brought up the GH2 which doesn't seem to suffer from moire and aliasing, and has an OLPF designed for stills. Proof is in the pudding. I don't know what exactly it's doing, but you can see the results.

The GH2 has just as much aliasing as the GH1 has. The difference is that it doesn't do the horrible purple/orange color aliasing like the Canons and Sony NEX cameras do. But it has the same degree of luma aliasing that the GH1 has, and can exhibit the same moire patterns.

 

[Steve Castle]

You do realize that even the Red has aliasing and moire right? And it reads out every row, every column. The point is, most bayer sensor OLPF choices pick a value somewhere between the green density and the red/blue density so as not to totally make the image too soft. If the project manager from Panasonic tells me they are not line skipping, then they are doing something else. I agree that they may not be actually reading the values fully out of it, but somehow they claim to be using the data from the rows.

Pertaining to the column, this is exactly what I am talking about with the ADC travel distance. Even if you had an ADC at the top and bottom, you'd still have to travel half the height of the ADC at minimum at the worst case, which, would dictate your clocking speed for all the rows. Skew is constant, at least in all measured tests I have seen, so the larger the sensor, the more time a read/reset takes. Like I said, the 5D read/reset is longer than the 7D read/reset.

But regardless of all this, all I am saying is you can't look at a sensor design and immediately know that it will be more sensitive because the pixels are larger. There are other factors, including the age of the design, and new methodologies, coming out constantly, that are part of the process. EVERY person who I've read their posts that has upgraded their M sensor to MX who have done tests have shown how much better the results are. I don't see how you can argue with all these end users who are absolutely thrilled with how much better the sensor is. Trust me, I don't listen to the Red marketing talk. I'm an electrical engineer who does CMOS back end design for a living. I laugh when they say they've made the most complicated ASICs ever, because I know better, maybe the most complicated they've ever done.

I think it is a good thing as well, and I absolutely understand the trade-offs that you point out. It's getting that balance between resolution AND DR and sensitivity. The Alexa and the F3 look like they do a pretty good job, but, from the samples of the Epic (footage) I have worked with, I really like that extra resolution. Would I want more resolution than what the Epic has? No. Past what it has is overkill to me, but that is a personal preference. At the end of the day, it's what is important to each of us.

One thing we don't know is how the skew is on this camera (JVC). That would be my main concern. They really need to get the skew down to sub 10us on these cameras.

I don't think that JVC is claiming "most complicated ASICs ever". Also, the main purpose of this FalconBrid that JVC is claiming to do is debayering, signal processing, encoding into MP4 or Mjpeg and output. For Red they aren't doing debayering and the signal processing stuff, which is the point of raw. So there is really no point of comparison with Red with the JVC is trying to with the FalconBrid.

As far the sensor itself. We can very clear assumptions about the size of the sensor relative to the what results we should be expecting out of this camera in terms of DR, low-light sensitivity, dof, in terms of noise, etc etc etc. A small sensor (~1/3") with a 4k resolution is not going to perform the same as a larger sensor (~S35). Everything downstream of that needs to work with what's provided by the sensor.

From a personal level, I really like the concept of this camera. An affordable 4K camera is interesting, and is certainly a more convincing move by JVC instead of making another 1080p camera that would have been forgotten in the flood of new cameras coming out. FalconBrid is a way for JVC to deploy 3D, 4k, and ultra high-speed 2k cameras using a common components. Its their competitive advantage, and its a really interesting one.

 

[Rick Burnett]

My comment about most complicated ASIC ever was about RED, not JVC, they made the claim. Given the results I have seen with back illuminated sensors, I guess time will tell. I was pretty impressed with the results of the sensor on my iPhone4 which takes better pictures or video than ANY other phone I have tested, by quite a large order of magnitude. With a smaller sensor you can use other technologies you cannot on a large sensor, so I still don't believe that size is everything, because all things are not equal. You can build a part that is more consistent and doesn't suffer from as extreme on chip variation. And, you can make it faster, so your rolling shutter isn't as big a problem. Which is neither here nor there because it's not a part of the sensitivity. Though I will say, for large sensor, deep dof requires quite a lot of stopping the lens down. So how low light sensitive they are in terms of usability does depend on your needs. But, that's really not about the sensitivity persay.

I have no doubt that Red and JVC on this aren't worthy of comparing to each other on these cameras, that was never my intention. What was my point was the incremental advances in the same pixel size between the M and MX sensor, and the fact the resolution is what 4K on the Red One. From the footage I have worked with, and especially some Epic footage, it doesn't look like Red is having a problem getting great sensor results at such a small pixel size that compare in the same range as larger pixels. 5.4um versus 8.25um. The area of the Alexa is MORE than twice the area of the Epic sensor, yet from the tests I have seen, and looked at, they don't look that far apart from each other. RED tested the Alexa and the Epic, and I think the Alexa had 1/2 stop more sensitivity on the same charts. That's not a huge difference, 3.7% in terms of stops.

 

[Steve Castle]

Steve, let's be very clear on one thing -- you don't know what Panasonic is doing. You're making these proclamations of what they are or aren't doing, but you aren't a Panasonic engineer, you don't work in the company, you don't have access to their tech documents, so you don't know. You're guessing. You don't know that the sensor is "DSLR based". You don't know that it skips rows. You don't know whether or not it bins pixels. You don't know, really, anything about it. Nobody outside of a few engineers in a factory in Osaka know what it does.

The only things we know about the AF100 sensor are that it has 12.4m pixels, that it has a primary color filter, and that it is 17.8 x 10mm in size, and that it's MOS technology. That's it. We don't know that it's Bayer, or that it skips rows, or that it bins. I would think it's entirely possible that they could have designed a primary color filter specifically for this camera that would be maybe 2x2 for each pixel, allowing them to bin pixels in a 2x2 grid and giving them effectively a 3-megapixel Bayer chip with comparatively quite large pixels (about 5.7 microns).

Why would the AF100 be limited to a "mere 1080p"? Perhaps for the same reasons that the F3 and F35 and the Alexa and the HPX3700 are "limited" to a "mere 1080p." That's the video standard, and this is a video camera. There is no video standard beyond 1080p, there's no video transmission standard beyond 1080p, there's no recording format they have that goes beyond 1080p, there's no broadcast standard that goes beyond 1080p, and there's no delivery system beyond 1080p. Seems like 1080p is a perfectly reasonable choice.

Fair enough, we can only speculate on information that hasn't been confirmed by Panasonic. Though at the same time I don't know why Panasonic would confirm or release information that can be interpreted negatively by some. Though as an forum discussing the technical merits of different cameras we need to form reasonable conclusions based on available information.

While I understand why you are defensive, but let's be logical here. We know that its a 12MP FourThirds sensor for a camera that is designed to output 1080p. If as you're suggesting this camera is using a specially designed sensor that is doing 2x2 pixel binning, why not just make a sensor that is closer to the target resolution instead of 12MP?

Now again we're speculating on things that isn't part of the official marketing material. Technically speaking, to bin a 2x2 pixel would require additional logic on a active pixel sensor such as the MOS (this we do know). Unlike passive sensors such as CCDs, charged-domain binning/summation requires the readout of each photosite of the 2x2 array to be delay integrated until all photosites in that array are read. Now, when you doing summation on a 2x2 array you are dumping the charge on the read-out bus, this means you'll need to preserve color information. Companies like Cmosis that work on pixel-binning for active pixel sensors actually have a modified CFA where the shared pixel are staggered. Obviously it would be a lot easier to just make a 3MP FourThirds sensor.

If its the case that they are using a 2x2 binned pixel as you're suggesting, than it should be something Panasonic put in their marketing material. Because that would be very compelling.

As far as 'line skipping', you're right in that its an over oversimplification of many techniques of sparse sampling the array to get down to a lower resolution. Technically they may not be using 'line skipping' but they may be using similar techniques. Another technique is to use on-chip column circuitry to perform block averaging using programmable 2x2 kernel sizes. However, this technique will not act as a ~5.7 micron photosite as you say, or gain the advantage of pixel-binning, it'll still be comparatively small photosite of a 12MP sensor.

EDIT:

Actually, looking back at Panasonic's marketing material they are pretty clear the sensor is derived from still-photography technology:
http://mirror.dpreview.com.s3.amazonaws.com/AG-AF100.pdf [PDF]
Also, Japanese news site AV Watch Impress (which is really a PR mouth-piece for companies and has never written a critical article ever) also reports that its derived from a GH based sensor even though there are some specification differences. I think its fairly clear that this sensor is derived from a dSLR sensor (or in this case MicroFourThirds still camera) and there is no evidence of summation techniques like pixel-binning. And again, its not slight against the AF100, it is a sub-$5k camera that offers a lot, but its a technical reality.

 

[Steve Castle]

My comment about most complicated ASIC ever was about RED, not JVC, they made the claim. Given the results I have seen with back illuminated sensors, I guess time will tell. I was pretty impressed with the results of the sensor on my iPhone4 which takes better pictures or video than ANY other phone I have tested, by quite a large order of magnitude. With a smaller sensor you can use other technologies you cannot on a large sensor, so I still don't believe that size is everything, because all things are not equal. You can build a part that is more consistent and doesn't suffer from as extreme on chip variation. And, you can make it faster, so your rolling shutter isn't as big a problem. Which is neither here nor there because it's not a part of the sensitivity. Though I will say, for large sensor, deep dof requires quite a lot of stopping the lens down. So how low light sensitive they are in terms of usability does depend on your needs. But, that's really not about the sensitivity persay.

I have no doubt that Red and JVC on this aren't worthy of comparing to each other on these cameras, that was never my intention. What was my point was the incremental advances in the same pixel size between the M and MX sensor, and the fact the resolution is what 4K on the Red One. From the footage I have worked with, and especially some Epic footage, it doesn't look like Red is having a problem getting great sensor results at such a small pixel size that compare in the same range as larger pixels. 5.4um versus 8.25um. The area of the Alexa is MORE than twice the area of the Epic sensor, yet from the tests I have seen, and looked at, they don't look that far apart from each other. RED tested the Alexa and the Epic, and I think the Alexa had 1/2 stop more sensitivity on the same charts. That's not a huge difference, 3.7% in terms of stops.

My bad, I misread that claim to be JVCs not Red's. It sounds like Red more than JVC now that you mention it.

As far as back-illuminated sensors they aren't very unique anymore, and primarily benefit smaller sensors. Which is why they are fairly common on cellphone sensors and point & shoots these days; I think there were a dozen mentioned at CES recently using back-illuminated sensors. As far as photosite size, we aren't talking about comparing a 5 micron camera to an 8 mircon, this JVC sensor is likely going to around 1-2 microns. Back illumination is suppose to double its sensitivity, but its still a good order of magnitude different in size compared to S35 equivalent sensor.

As far as Red's comparison between the Epic and Alexa, a half-stop is actually significant. But I will further question the DR methodologies, I also find it odd that they conducted the comparison on both DSC and DRTC tests but didn't release the more accurate DRTC certificate. Whatever, its not really that important. There is just too much grey area in measuring DR to really compare claims.

 

[Rick Burnett]

Heh, well I won't argue about the grey area of DR testing. One thing for sure is acceptable noise is absolutely not quantifiable, and makes taking someone's opinion on a camera difficult to judge. The only real way is to get the camera and shoot what you know. You'll know in no time (once you get the camera settings right of course).

 

[robertrogoz]

Having worked with footage from the Red One, and how amazing it was, Red really changed my mind about resolution. But, the caveat is, I also thing resolution isn't everything because these sensors are designed by different teams and their operations can be totally different. I like the freedom 4K gives. Even when working at 1080p as my final resolution, with 4K, if I want to crop in on a shot, I wont lose resolution in the final. Whereas, if I am starting with a camera that resolves 800 lines, and I try and crop in on shots, I start to seriously degrade my footage immediately.

But you are comparing apples and oranges. The reason you get to the difference with RED ONE (or similar cameras) is the size of the sensor and amount of detail lens can transfer onto the sensor. With a small sensor (like I predict on JVC case will be between 1/4 to 1/3 max) there will be no visual difference between 1920x1080 2K or 4K.

 

[Rick Burnett]

But you are comparing apples and oranges. The reason you get to the difference with RED ONE (or similar cameras) is the size of the sensor and amount of detail lens can transfer onto the sensor. With a small sensor (like I predict on JVC case will be between 1/4 to 1/3 max) there will be no visual difference between 1920x1080 2K or 4K.

Sorry, but that's not true. First, a smaller lens is MUCH cheaper and easier to manufacturer than a large one that has to cover a wider sensor that can resolve more detail. Using a lens designed for the sensor is not going to be the issue, it will not be the limiting factor. Second, that a 1080p bayer DOES NOT resolve to 1080p. it's mathematically impossible. So, a 4K bayer dropped down to 1080p is going to have more detail than shooting a native 1080p. It will also have more color resolution as well.

Look at the size of the HV30 sensor. The lens in it is not the limiting factor, the sensor is. How can I tell? Easy, I can get the HV30 to show moire/aliasing. This means the lens is showing detail past the limit of what the sensor can handle.

 

[Camera Expert]

Was a chip size even mentioned? If not then why complain about the size if we don't know what the size is?
The chip size in the top JVC consumer camcorders are 1/2.3" and the chip for this camcorder could be even bigger.