What camera has the biggest sensor today?

[David Heath]

Originally Posted by Andrew Bower

Steve,
I am intrigued with your comment. Would this still be true if one would capture the FS100's output via the HDMI output?

Yes, the EX is likely to be sharper. It's difficult (for a 1080 output) for a single chip design to equal, let alone better, a design of 3 chips, each of 1920x1080. For every output pixel, unique red, green, blue values are available. Use a single chip (as is necessary once you start getting much larger than 2/3") and you're in to the whole world of deBayering, and different colours having differing resolutions. It (theoretically) could be overcome by chips with large numbers of photosites, but then it becomes increasingly difficult to read out the whole chip at up to 60 times a second. Hence why still cameras have to resort to only reading out a lower percentage of the total available every frame.

I find that the EX compression algorithm gets slightly soft when dealing with lateral movement, is this evident on the FS100 as well?

Is it really the compresion algorithm, or just the natural effect of motion blur?

As far as lenses, big sensors, and sensitivity go, then the ONLY fact that ultimately determines how a camera works in low light is the physical size of the front element of the lens - which may not sound logical until you think about it. (Assuming a given basic technology and resolution.)

The reason is that if you quadruple the area of a sensor, then it may well lead to an increase of 2 stops in the ISO rating. (Keeping all else equal.) But think of the other implications. For the same angle of view, it will mean the focal length of the lens must be doubled. If the same diameter of lens is kept, that means when expressed in f stops, the max aperture of the lens will be two stops down on the first (smaller sensor) case. Keep the same diameter of lens and you're back to square one - two stops better camera sensitivity, but two stops less lens. Same overall performance.

It's usual to think in terms of a doubling of ISO rating meaning much better in low light. It's true - but ONLY if the same f stop in each case. In the above example, that would mean quadrupling the area of the lens as well as the sensor. That means a lot of extra glass, and I'll leave you to guess how much extra weight and cost........

[Steve Game]

Andrew,

The sensor in the FS100 (the same as the F3) is stated to be 2440x1373 (effective). (The Alexa is similar at 2880x1620, so its in the right ball park.)
So the FS100 luminance resolution will be largely determined by the green content, which is half the total, i.e. 1220x690, augmented by the red and blue content. The first stage of processing is to 'demosaic' the sensor data to produce the raw video which can be made available at the HDMI port. This is compressed to 4:2:2, (4:4:4 on the F3 but as the chroma source data is lower resolution than the luminance, it cannot provide a genuine full resolution).
The EX1R will provide 4:2:2 video at its HDMI port based on chroma source from three full 1920x1080 sensors.

So in simple terms, the luminance on the FS100 can represent 610x345 cycles and the chroma can be somewhere between 610x345 and 305x173 cycles depending on the proportions of primary colours.
The EX1R represent 960x540 cycles equally for all colours.

Now practice is not that simple as there are other issues and of course both cameras are part of a commercial products, each designed to address, (and protect) a particular market.

All digital cameras should have an optical low pass filter to limit the detail arriving at the sensor to below the nyquist limit of the sensor. That's fine for a three chipper like the the EX1R, but the FS100 has a chroma resolution ranging between 1/2 and 1/4 of the luminance, so a single filter will be a compromise between maintaining the luminance detail whilst preventing excessive chroma sampling artifacts, (the so-called Bayer pattern signature of single chip cameras).

The other issue is one of noise. In more recent times, some of the shine has rubbed off the EX camera's reputation as other newer products have better lower light performance. That's where large sensors really score, so DCT and interframe compression does not sacrifice detail at the expense of bandwidth.

So for low-light clips, the large sensor cameras may well have better details recorded despite their theoretical lower resolution front ends.

These are just my thoughts so they could be wrong, - stands back for flaming from the experts here. :¬}

Oops, This post was started before David commented so it may repeat some of his comments - or contradict them.

[Steve Game]

Originally Posted by David Heath

As far as lenses, big sensors, and sensitivity go, then the ONLY fact that ultimately determines how a camera works in low light is the physical size of the front element of the lens - which may not sound logical until you think about it. (Assuming a given basic technology and resolution.)........

David,

I've now had time to read your post. One question:
If I read you correctly, the same 'basic technology and resolution' means that each photosite is about 3 microns square on a 1/2 inch sensor and about 10 microns square on a S35 sensor (of 2440x1373 resolution used to get a 'full HD' output). That's about three times the surface area so the SNR would be that much better. If the same 'basic technology' meant 6 micron sites, then the basic resolution would be about 7963*4480 which is over 35M pixels.The SNR would be no better, but there would be enormous processing overheads to get it down to 1920x1080, unless the inadequate pixel skipping techniques of SLRs was used.

Have I understood your comment correctly?

[David Heath]

Originally Posted by Steve Game

The sensor in the FS100 (the same as the F3) is stated to be 2440x1373 (effective). (The Alexa is similar at 2880x1620, so its in the right ball park.)
So the FS100 luminance resolution will be largely determined by the green content, which is half the total, i.e. 1220x690, augmented by the red and blue content................

So in simple terms, the luminance on the FS100 can represent and the chroma can be somewhere between 610x345 and 305x173 cycles depending on the proportions of primary colours.
The EX1R represent 960x540 cycles equally for all colours.

In practice, the F3 and FS100 manage a lot better than the numbers you quote (and 610x345 cycles is another way of saying 690lpph), the reason is that whilst indeed green is the largest contributor to luminance, red and blue do both contribute. Without going into too much technicality, it means that a good deBayering algorithm should give luminance resolution corresponding to about 80% of the sensor linear dimensions. (Practically, it tails off, so it's not possible to give an exact number.)

So if you want 1920x1080, working backwards means that you'd expect a sensor of about (1.25x1920)x(1.25x1080) - or about 2400x1350 to give that sort of resolution

Use less, and the resolution will be less than optimum, use more and there are improvements - but at the expense of reading extra data off the chip, with power, heat etc penalties. It's no accident that Sony and Arri chose the figures they did!

As you say, this is true for LUMINANCE - chrominance is another matter, and it's inherent in such as a Bayer sensor that chrominance resolution will be less than luminance.

But resolution is one thing, aliasing is another. It follows from lum resolution being higher than chrom that you may get chrominance aliasing at resolutions not much higher than the wanted luminance band - and as such it will be impossible to design an OLPF good enough to get rid of them.

Assuming a 2400x1350 sensor, think of a set of white/black horizontal lines, 1350 in total - so 675 white, 675 black - effectively one line per sensor row. In an extreme case you will get alternate rows of the sensor white, alternate black. But alternate rows are R,G,R,G,R and G,B,G,B,G. Hence, only red and green photosites will give a response - OR only blue and green if the pattern is shifted by a row. It's pretty easy to see why you'll get significant coloured aliasing around this spatial frequency - and it will be irrespective of the deBayering method used.

And look at a chart of the F3 and it's exactly what you do get. Look at a chart from FS100, and you don't. You see aliasing at the corresponding point - but it's not coloured.

I've said in another thread that (contrary to popular belief) the only explanation that makes sense to me is that the F3 and FS100 don't share the same sensor, that the FS100 must have one with a much higher photosite count than the F3. If anyone can explain to me a mechanism whereby the FS100 can avoid coloured aliasing around 1350lpph if it does have the same sensor as the F3, I'd be very interested to hear it. I don't believe it would be physically possible without severely reducing the overall resolution. (Which measurements don't show.)

[David Heath]

Originally Posted by Steve Game

David,

I've now had time to read your post. One question:
If I read you correctly, the same 'basic technology and resolution' means that each photosite is about 3 microns square ........

Have I understood your comment correctly?

By "same basic resolution" I meant (say) 1920x1080 (or whatever) in each case, by "same basic technology" I meant not comparing (say) CMOS with CCD etc.

Hence, yes, for the larger sensor, the photosites will be bigger, yes, you'd therefore expect better SNR, so yes, you'd therefore expect a higher ISO rating. The important part is that that translates to correspondingly better lower light performance if - and only if - the comparison is done at THE SAME F STOP.

Now, because the sensor is bigger, to maintain angle of view the focal length of the lens must be longer, so by the laws of optics TO KEEP THE SAME F STOP the diameter of the lens must be correspondingly bigger as well. (If l=focal length, and d=diameter of the lens, f no = l/d by definition - F-number - Wikipedia, the free encyclopedia.) That's why to get better low light performance you need a bigger diameter lens, period. Keep it the same, and optically, as you vary sensor size, ISO rating will go up, f stop go down in proportion - and performance stay the same!

Incidentally, in your example (3 micron v 10) they represent linear distances. For area, the corresponding differences are the squares - 9 versus 100, so the differences are well over 10 - not 3.

[Steve Game]

Interesting info. Thanks.

So the case of pixel skipping SLRs is different as although the total sensor size is larger than a conventional camera, only a portion of the total light falling on it is used for video. I have a Canon 550D which has an effective resolution of 5184x3456. If Canon skip every other pixel then the sensor looks like 2592x1458 before debayering. That means that less than half the total light is used and the light sites are around 1/4 the area of those on the FS100, thus increasing the noise.
Incidentally, that would also explain why the setup struggles to reach 600lines pph and still creates a mess of artifacts from any detail in the scene.

And yes, I did forget the area comparison rather than the linear sizes. It took so long to find out the resolution of the FS100 sensor. Despite Sony's hints about the sensor being the same on the FS100 and F3, they don't seem to have confirmed it in any publically released specifications, and some commentators seem to suspect differences (yourself included).

[David Heath]

Originally Posted by Steve Game

If Canon skip every other pixel then the sensor looks like 2592x1458 before debayering. That means that less than half the total light is used and the light sites are around 1/4 the area of those on the FS100, thus increasing the noise.
Incidentally, that would also explain why the setup struggles to reach 600lines pph and still creates a mess of artifacts from any detail in the scene.

Yes to the point about sensitivity. It's also the case that there is an unused "blank" portion around each photosite. Increase the number of photosites for a given sensor area and the percentage "wasted" by such boundary areas will also increase.

Sensitivity aside, IF your Canon was still able to look like 2592x1458 Bayer after skipping every other pixel, then it could still give comparable results to the F3 and Alexa - theoretically! That's the same number as they have in the first place.

Trouble is, that simply skipping every other pixel on a Bayer sensor will give only a single colour! (If used sites are capital letters, because then you'd get something like g,R,g,R,g,R,g etc!) Hence it needs to be a bit more complicated. Even allowing for that, full deBayering and high quality downconversion (from the 2592x1458 matrix to 1920x1080) at full frame rates is processor intensive and power hungry - witness the power consumption of such as the F3 and Alexa.

Early DSLR systems used to skip several whole lines, then do something complicated within each line - it gave severe colour aliasing and asymmetrical results on charts. It was really intended to give live viewfinding initially - the idea of video from a still camera followed on.

I believe more recent systems use a form similar to what Canon do with the C300, basically reading out blocks of 2x2 and getting direct R,G,B values - no deBayering, and as the C300 chip has 1920x1080 blocks, no downconversion either. Hence the low power consumption.

For "designed for still" chips, it seems they skip BLOCKS rather than photosites, and that's why the fundamental resolution then becomes a quarter that of the sensor - so in your case would be 1458/2, or 729 lpph as a theoretical maximum. Apart from reducing the amount that needs to be read every frame, this also means the matrix that results will be 1296x729 (using your figures) which will need UPCONVERSION, not downconversion. Which is easier to do. All this helps explain why the power consumption of DSLRs etc is less than such as the F3.

Despite Sony's hints about the sensor being the same on the FS100 and F3, they don't seem to have confirmed it in any publically released specifications, and some commentators seem to suspect differences (yourself included).

At the end of the day it's the results that are important, and the FS100 is pretty good. How it gets it is really of academic interest. My own suspicion is that the FS100 uses the same sensor as the FS700, and if that is true it's then pretty obvious why Sony would want to be coy over not letting the details out originally - it would have been evident that something like the FS700 was bound to follow......... (As the C500 was an inevitability given the C300 sensor details.)