New Panasonic 4k 2k CMOS DSLR Sensor with high frame rates

What I find interesting there are the references to "Movie Mode", and how it says "1080p/60 (4 pixel mixture), 720p/60 (9 pixel mixture)"

I interpret that as suggesting that in this mode it is starting with a basic 3840x2160 photosite matrix and deriving 1080 directly by taking 2x2 clusters of photosites, (exactly as we know the Canon C100/300 do) and 720 directly by taking 3x3 clusters. That should give good results in terms of definition and low aliasing etc.

Unfortunately, there may be a "but" and that is that it must lead to quite a high degree of cropping - to the extent that it will only use about half the total area of the chip in video mode. Not the end of the world - and probably a good price to pay for the high quality it should give - but expect it to narrow the angle of view of lenses, and lessen the shallow depth of field effect. Given the horizontal figure of 4644 they quote for total pixels, then if 3840 get used, it will be a crop of about 1.2x compared to full usage of the 4/3 sensor.

[EDIT - Saying "half the total area of the chip" above doesn't take into account aspect ratio, and the fact the chip is actually 4:3. Using a 3840x2160 crop therefore means usage of about 0.7x the area that would otherwise be used for 16:9 aspect ratio, if the full width was used. Still quite a big crop, but not as much as half, if looked at in that way.]
 
Samuel H said:
Summary: the spec sheet doesn't mean much, no way to know how the sensor works based on this info alone, we have to wait until we get further details and some sample videos.
(yes, I jumped to the same conclusion about video using a central crop of the sensor, and it seems it is not necessarily the case)
Click to expand...
Sorry - I'd missed all that!

Regarding the EosHD link - http://www.eoshd.com/content/11502/...nce-4k-video-capable-micro-four-thirds-sensor - then I 'm sceptical about several things there. Firstly, they talk about "We know also that it does a 4×4 pixel mix in 1080p mode (as per the PDF specs for the new 4K sensor) instead of .....". That doesn't make sense. I think they are interpreting "4 pixel mixture" as a 4x4 block (16 pixels) which seems unlikely. It makes far more sense to interpret it as 4 pixels total - so 2x2 blocks.

That interpretation seems far more likely when taken together with the 720p spec - "720p/60 (9 pixel mixture)" which by my interpretation would mean 3x3 blocks - not 9x9!! :)

WHY that makes sense is that 1280x720 sets of 3x3 blocks comes out at (1280x3)x(720x3) total - 3840x2160. And 1920x1080 sets of 2x2 blocks comes out at (1920x2)x(1080x2) total - 3840x2160! Funny coincidence that...... :)

And such should give very good results - albeit with the cropping caveat. Anything based on 4x4 blocks can only have a maximum horizontal resolution of about 4644/4 (1161) or about 650 lpph.

The EosHD link also says "In silent electronic shutter mode the GX7′s sensor uses a full pixel readout – essentially 4.5K raw at 16MP in 4:3. So in theory if you want to get a glimpse of the GH4′s 4K video image quality, have a look at the stills from the GX7 shot in silent shutter mode."

I'll believe that when I see it. There's nothing in the specs to indicate that full pixel readout happens in anything other than stills mode. It would almost certainly be power hungry to do it. But there IS something in the specs to indicate that in 1080 and 720 video modes at least, they are based on a 3840x2160 crop - at least if you accept that 4 and 9 pixel mixes relate to 2x2 and 3x3 blocks.

I've said many times before that joint video and still usage puts conflicting strains on sensor design - optimise for one and the other inevitably gets compromised. Here, we are talking of a sensor primarily for the stills market. Using a 3840x2160 crop is the best way I can think of to use it to get high quality video from it, and IMO high quality from a cropped sensor is better than compromising quality for the sake of full area utilisation. The only way you could have your cake and eat it is to design the whole area of the sensor to be 3840x2160 (as Sony and Canon have done) - but that would compromise the stills performance!
 
@harddrive yes, those are the calculations I made too, but then on the personalview forum they say the GX7 (supposedly this same sensor, but this one they could actually test) does line-skipping, then horizontal 4pixel mixing, which is completely different than what I concluded from the spec list
 
Samuel H said:
........ on the personalview forum they say the GX7 ...... does line-skipping, then horizontal 4pixel mixing, which is completely different than what I concluded from the spec list
Click to expand...
I'm struggling to think what could be meant by "line-skipping, then horizontal 4pixel mixing" and off the top of my head it doesn't sound as if it will very good if it was the case! :)

It really boils down to exactly what is meant by the phrase as used in the data sheet - "1080p/60 (4 pixel mixture)". I can't see why it should be taken to mean "horizontal mixing" (whatever that means), let alone associated with line skipping. It's far more logical to assume it means forming an R,G,B pixel directly from a 2x2 Bayer block. And don't overlook the 720p mode specifications and the use by Panasonic of "9 pixel mixing" there. That makes perfect sense if you accept the logic above, it's just using 3x3 blocks instead of 2x2. How do they explain the 720p specs on the forum you mention?

And it also makes sense because it should give about as good a result as it's possible to get from such a sensor. Albeit with the implication of extra cropping.
 
From what I read inthe personal view forum, it means they only read half the horizontal lines, and they then use a moving average of 4 pixels to get every final pixel (so, the first one would be an average of pixels 1,2,3,4, the second one would use 3,4,5,6, etc). Very weird if you ask me, but that's what they imply, if I'm reading it correctly.
For 720p, it would seem they read a third of the horizontal lines, then use a moving average of 9 pixels (first 1,2,3,4,5,6,7,8,9, then 4,5,6,7,8,9,10,11,12, etc). Not a bad idea if you can use a weighted avarage, and if all you can read at this speed is a third of the lines.

The obvious alternative of reading all lines and doing a square 2x2 mixing makes a lot more sense, but it does require much faster sensor readout.

Conflicting reports --> wait and see
 
Samuel H said:
From what I read in the personal view forum, it means they only read half the horizontal lines, and they then use a moving average of 4 pixels to get every final pixel (so, the first one would be an average of pixels 1,2,3,4, the second one would use 3,4,5,6, etc). Very weird if you ask me, but that's what they imply, if I'm reading it correctly.
Click to expand...
Skipping every other line wouldn't work. That's because for a Bayer sensor odd lines are RGRGRGRG, and even are BGBGBGBG - so if you skip every other line, you'll end up with either red and green OR blue and green pixels only. Likewise, you need to think of the Bayer layout when you talk about the horizontal averaging. An average of four pixels sounds a good idea - until you take into account that it can only be two green and two of either red or blue. So you're not averaging adjacent pixels, they're separated by ones of a different colour.
For 720p, it would seem they read a third of the horizontal lines, then use a moving average of 9 pixels (first 1,2,3,4,5,6,7,8,9, then 4,5,6,7,8,9,10,11,12, etc). Not a bad idea if you can use a weighted avarage, and if all you can read at this speed is a third of the lines.
Click to expand...
Every third line then certainly gets you a mixture of R,G and B pixels, but averaging over 9 horizontally brings with it the same problems as above - you have to group 1,3,5,7,9 together and 2,4,6,8 to be averaging ones of the same colour. I can't see that being very good for horizontal resolution, let alone aliasing....... And it also seems unnecessary.

It is very easy to tell whether any system is using such a system by looking at a chart. Such as they describe will give a very asymmetric result, and likely pretty bad coloured aliasing. (It's believed something like it was used in the earliest DSLR video - see http://downloads.bbc.co.uk/rd/pubs/whp/whp-pdf-files/WHP034-ADD39_Canon_5D_DSLR.pdf More recently, a better approach has been to read one out of four 2x2 blocks from the whole sensor and upscale to 1080. That will give resolution of a quarter the sensor dimensions (so about 650 lpph), but (more importantly) no coloured aliasing.)
The obvious alternative of reading all lines and doing a square 2x2 mixing makes a lot more sense, but it does require much faster sensor readout.
Click to expand...
Not that much necessarily. The 16:9 crop of the whole sensor will be about 4644x2612 - 12,130,128. Skip half the lines and that brings it to about 6,065,064. Use a 3840x2160 crop and the figure is 8,371,200 - just over a third more. But reading that area gives a direct read out of 1080 and 720 video with no need for any further debayering or rescaling - that's the real beauty of the 2x2/3x3 readout when coupled to a 3840x2160 sensor. A block directly forms a pixel. Best if those sites fill the full sensor area, but if it can be done with some cropping - because you need more photosites in total than for still photography - it's still worth it IMO.

If you accept the idea of reading one 2x2 block in four as the previous best way of doing it, that means then reading about 3,032,532 photosites at frame rate. But that's the technology of a few years ago, so I don't find the approx 8 Megapixel figure impossible to believe now. That opinion is made more likely by the talk of high frame rate modes - 240/481fps - though they are obviously likely to be at reduced quality compared to normal HD modes. (As with the FS700.)
 
when I said "read half the horizontal lines", I meant it reads two lines then skips two lines, etc; same with columns, 1,2,3,4, etc mean first, second, third, fourth RED photosite, etc
for 720p I would think they read two lines then skip four lines, etc
It's all speculation based on what the GX7 sensor is supposed to do

the cropped square-2x2 mode would make a much crispier image with few (if any) artifacts, but then the sensor wold become 2.5x crop for video, so, halfway between the sensor size of the BMCC and the BMPCC

(also: when a sensor reads one pixel from one row, it usually means it is going to read all the pixels in that row; I've just read a couple of sensor manuals, but they all do "windowing" in that way: you choose which rows you want, but all columns are always read; further processing can ditch columns, but that won't affect sensor readout time) (other designs are possible, I'm just saying what I have found in my very limited low-level experience)
 
Samuel H said:
when I said "read half the horizontal lines", I meant it reads two lines then skips two lines, etc;
Click to expand...
OK - I'll go along with that.

In cameras where I'm pretty sure that's what is being done, it usually seems to go hand in hand with also skipping columns - read two columns, then skip two. (Though yes, I'll also go along with what you say about "all columns are always read; further processing can ditch columns, but that won't affect sensor readout time.") The evidence for that is charts show identical performance on the horizontal/vertical axes. Consequently, it's logical to assume it's doing the same horizontally as vertically. Hence there is a single 2x2 block read out of each 4x4 block. That also predicts resolution of one quarter the total photosite count on each axis. So if 4644, that means 1161, or 653 lpph - which is indeed typically what we get.

But it's difficult to then see why you'd need a different readout method for 720. The above gives an intermediate raster of 1161x653 - which needs scaling up to either 720 or 1080 frame sizes - and upscaling is much more easily done than downscaling. So why not read the sensor the same for both 720 and 1080 modes, then use different amounts of upscaling?

Excuse me if I've missed something, but is it known for sure that the talk about the new sensor and the GX7 are the same thing? If maybe the GX7 was doing it the "old" way, and the new sensor is going to do something else that may explain a lot. It would also mean that a future new product using such a sensor would have a very significant quality improvement relative to the previous generation - though with higher levels of cropping.
 
That makes a lot of sense, and it was my first thought when I read the spec, but for some reason the guys at the personalview who tested the GX7 found a different behavior. That's all I can say.
 
You must log in or register to reply here.
Back
Top