Re: Sony A7RII - FIRST video sample!!!

[paulcurtis]

OK Samuel, I read Paul's post to mean that he thought 2x2 direct read couldn't ever work properly because it couldn't access "the surrounding RAW pixels" - which is not true. There's enough information within each 2x2 block for it not to need to access any other photosites. My apologies if I misinterpreted.

No worries

Yes, i meant that some of the better debayering algorithms use the surrounding pixels to make a much better image, with edge detection and going beyond simple colour translations etc,. not that the 2x2 wouldn't work (as clearly it does!)

But doesn't the C300 suffer from some colour moire in some situations, perhaps as a result of this binning approach?

(of course most cameras suffer in some way because the debayering is optimised for speed, not quality, another plus for RAW workflows)

cheers
Paul

 

[combatentropy]

But doesn't the C300 suffer from some colour moire in some situations, perhaps as a result of this binning approach?

No, this binning approach eliminates moire.

Moire is when a line is focused onto just one photosite and not its neighbor. So that line gets rendered only in the color of the photosite it happens to fall upon. Hence rainbows.

Of course a line can fall onto just one photosite of the C300 just as well as it can on any other Bayer sensor. But the C300 merges a square of four photosites into one output pixel. That line of detail basically gets erased. It becomes a smudge, as it is merged with whatever detail fell on the other photosites.

(of course most cameras suffer in some way because the debayering is optimised for speed, not quality, another plus for RAW workflows)

There is something to this, but not as much as some people think. Most cameras use specialized chips whose sole purpose in life is to handle imagery. These ASICs are unlike the general-purpose processors found in your desktop machine, which not only must handle imagery but also open Word documents and surf the Internet. So, comparing the time that each spends debayering something is very misleading. An ASIC can do in an instant what might take a general-purpose processor minutes. That's not to say that a slow-baked debayer on your desktop computer might be better than the hard-coded formula in a camera. However, comparing just the render times is like comparing the bit rates of two different codecs. HEVC, for example, at 100 Mbps is probably better than ProRes at 200 Mbps.

 

[harddrive]

Yes, i meant that some of the better debayering algorithms use the surrounding pixels to make a much better image, with edge detection and going beyond simple colour translations etc,. not that the 2x2 wouldn't work (as clearly it does!)

The caveat to what you say above should be for the same number of photosites the better deBayering algorithms will indeed do a better job. For a grid of (say) 1000x1000 photosites, then if you used a simple 2x2 read system, you'd expect a maximum output resolution of 500x500, agreed? But note it would be effectively true "4:4:4" - 500x500 for red, green, blue (and luminance!).

A "proper" algorithm will indeed do better - squeeze out more of what the chip is intrinsically capable of. It's difficult to be precise, but the generally accepted figure is that you could expect something like 750x750 for luminance resolution, though the chrominance resolution will still only be around 500x500. The downside is more complex processing (especially if you need to rescale to a desired output format) which will likely mean more power and heat.

But what if we now decide to make our sensor not 1000x1000, but 1500x1500? It follows that using the "2x2" approach, it'll match the 750x750 luminance resolution of the previous example - but will also have full chrominance resolution to match. But it comes at the price of 2.25x as many photosites, though the benefit of simpler processing.

Looking at it another way, if our target output resolution is (say) 100x100, then we can get it via a Bayer sensor in two ways. Either start off with one of 200x200, and do the 2x2 approach, OR start off with approximately 140x140 and do a conventional deBayer. The results will be comparable in each case, it's just that the former approach will need 4x as many photosites as required output pixels, the latter approach will need something like 2x as many. That may seem like a disadvantage to the simpler (2x2) approach, but I come back to what I said earlier:

Having to have so many photosites may seem like a big disadvantage at first sight, but for what we're talking about here (hybrid stills/video cameras) a high photosite count is necessary anyway for still images.

What I believe may be different between the A7R ii and any other DSLR comes down to the other sentence from earlier: "The trick is to make it up to EXACTLY twice QFHD on each axis (7680x4320)." That's why about 33Mp becomes the magic number for 16:9, which translates to about 40Mp for 3:2.

 

[Samuel H]

Just for fun: if you're going to use 7680x4320 to get a 3840x2160 signal, you may as well use the fact that you're going to encode it as 4:2:2 and add some white and almost-black photosites to improve DR, losing color information that you're not going to store anyway.

I wrote this when the C300 came out.

Of course, this is hell to debayer, but you get 1/3 to 1/2 stop of extra low-light capability thanks to the W photosites, and thanks to the K ones you can improve DR by several stops!! (with almost-monochrome highlights if you push it too far, say with 4 or 5 stops of ND for the K pixel)

 

[harddrive]

I wrote this when the C300 came out.

It's an interesting concept! Off the top of my head, I'd suggest moving the lower two rows of photosites along by one block as a refinement - such that it stops red or blue blocks lining up in a column form?

Practically, a problem I can see with it is loss of versatility. When you say "the trend now is to have a sensor with 4 times as many photosites as pixels in the final image" I suspect there's a bit more to it than that. I strongly suspect that when Canon developed the sensor for the C300, they weren't SOLELY thinking of 2x2 read to HD, I suspect that it's use for 4K (via standard deBayering) was there from day 1. Indeed when the principles of the C300 first became known, several voices immediately said "expect a 4K camera soon.....!" And lo and behold it wasn't too long to wait until the C500.

And with what may be happening with the A7R ii, a similar thing holds true. We're not JUST thinking of 2x2 or 4x4 read to get 4K and HD from 7680x4320 - it can be normally deBayered to give max resolution stills off the whole sensor, and the count also means that a conventional deBayer off a centre crop can still give quality s35 4K. Going the way you describe wouldn't allow such versatility.

And with so many fewer red and blue photosites, it would substantially worsen coloured aliasing due to the much increased relative spacing between those coloured photosites. But an intriguing idea nonetheless!

 

[paulcurtis]

No, this binning approach eliminates moire.

Up to the nyquist limit of the density of photo sites you're using. The C300 did show some severe aliasing and i believe still does. I remember when it first arrived and pointing it at waves in the ocean looked unusable.

There is something to this, but not as much as some people think. Most cameras use specialized chips whose sole purpose in life is to handle imagery. These ASICs are unlike the general-purpose processors found in your desktop machine, which not only must handle imagery but also open Word documents and surf the Internet. So, comparing the time that each spends debayering something is very misleading. An ASIC can do in an instant what might take a general-purpose processor minutes. That's not to say that a slow-baked debayer on your desktop computer might be better than the hard-coded formula in a camera. However, comparing just the render times is like comparing the bit rates of two different codecs. HEVC, for example, at 100 Mbps is probably better than ProRes at 200 Mbps.

While i don't disagree in general it's important to show in real life that the hardware used in these cameras is rarely as good as dedicated non real time processing. A perfect example is in-camera scaling almost always results in negative lobes and black pixels and the look of sharpened edges because the requirements (real time) for camera hardware mean that choices are made to guarantee that - speed, heat and battery life. This is a very Sony look - black pixels around high contrast edges. You can mimic this in software easily but most software downsampling takes a more quality route.

There was a time when dedicated hardware was always faster but a modern CPU/GPU usually outpaces it. But look at the power requirements of both, it's not a fair competition. Plus you'd need to code close to the hardware in both cases. In Nuke i have a coding language Blink script, which compiles to GPU shader code directly. It's wicked fast.

It's the same for debayering, if you look at the options open to both Red and Alexa footage. Or even in software, comparing the debayer approaches can be a real eye-opener in terms of final result. I wrote something on this and compared side by side a while ago.

cheers
Paul

 

[paulcurtis]

The caveat to what you say above should be for the same number of photosites the better deBayering algorithms will indeed do a better job. For a grid of (say) 1000x1000 photosites, then if you used a simple 2x2 read system, you'd expect a maximum output resolution of 500x500, agreed? But note it would be effectively true "4:4:4" - 500x500 for red, green, blue (and luminance!).

Potentially daft question coming up

Within that 2x2 block are you just white balancing it to produce one pixel or is something else going on? (apologies if i missed this somewhere). The point about a debayer is that each RGB pixel's colour is based not just on local colour but the surrounding colour too (i.e. looking at what the corresponding pixels in other red, blue, green pixels are doing), extending beyond a single pixel group.

I think if you're just white balancing it to create a super pixel then nyquist generally suggest sampling at twice the highest frequency so wouldn't you need 4 times the number of photo sites not just double? Double for your super pixel then double again for sampling. Whereas in a normal debayer because you're dealing with single pixels then the oversample can be just double.

In real life situations i would always attempt to oversample again. I like shooting 4K for 2K delivery, or UHD for HD.


cheers
Paul

 

[harddrive]

Up to the nyquist limit of the density of photo sites you're using.

True. And that's an argument in favour of using more photosites than absolutely necessary to give the required resolution. This is oversampling - and the point is true regardless of which method is chosen.

Well, any method which fully reads and makes proper use of all available photosites! Once you get into the realms of line skipping and the poorer sorts of binning, then all ets are off!

The C300 did show some severe aliasing and i believe still does. I remember when it first arrived and pointing it at waves in the ocean looked unusable.

I don't know what you saw, but the C300 has never shown severe aliasing as a result of it's front end, and doesn't now. I don't expect you to take just what I say at face value, but look at https://tech.ebu.ch/docs/other/WHP034-add70_rev_1_canon_eos_c300.pdf - especially pages 11 & 12. Those are Alan Roberts tests on the C300 for the BBC, hosted now by the EBU, and whilst I wouldn't like to say there is NO aliasing, it's pretty small and unlikely to show up on normal pictures.

And it's worth comparing what you see in those pictures with the comparisons in https://tech.ebu.ch/docs/other/WHP034-add68_rev_1_sony_pmw-f3.pdf - Alan's equivalent report on the F3. They are both s35 sensors, but in the F3 it uses about 3Mp with a "conventional" deBayer, the C300 uses 8Mp with "2x2" processing. They both give highly acceptable results - that's the point - but you can see from the zone plate which goes up to 2160 resolutions (small one in between the red/green coloured plates) that for the C300 the out of band aliasing is further out from the centre. (Which is good! )

 

[harddrive]

Potentially daft question coming up

Within that 2x2 block are you just white balancing it to produce one pixel or is something else going on? (apologies if i missed this somewhere).

OK. Let's go back a bit. Bear with me if some of this sounds a bit basic, and we'll also assume it's before any subsampling of chroma prior to recording etc. We'll stick to R,G,B.

Most important is to distinguish between what I've seen called "input pixels" and "output pixels". The former are basically photosites on the chip, the latter are the pixels that go to form up a frame of the output. Taking the C300 as an example, we have 3860x2160 of the former, 1920x1080 of the latter. But the most crucial thing to appreciate is that input pixels only have a single value, output pixels have three - one each for red, green, blue. The value for an "input pixel" will be either for red OR green OR blue - depending what filter the photosite has in front of it.

Now, if we were deBayering conventionally, (to give 3840x2160 output pixels) the process has to calculate (ie guess!) what the red and blue values should be for all the green sites, and the green and blue values should be for all the red sites, etc. But the point is that even the best algorithm can't fully construct what isn't there. The best it can do is make a best guess. And because there as twice as many green photosites as the other colours, it also follows that it can make better guesses for green than red or blue! Hence why deBayering after this fashion inevitably means better luminance resolution than chrominance.

But for the C300 we don't want anything more than a 1920x1080 output. And this is where the 2x2 block idea comes in, as that sensor has 1920x1080 such blocks. Each block will directly give 4 values - one for each photosite - but two will correspond to green values, one to red, and one to a blue value. So what "2x2 direct read" does is simply take the red and blue values, and average the two green values to directly give red, green, blue values for the block. And with 1920x1080 blocks, that's directly given us 1920x1080 output pixels with three relevant values for each. No need to look at any other surrounding photosites.

I'm aware that a picture is worth a thousand words, so look again at Alan Roberts link - https://tech.ebu.ch/docs/other/WHP034-add70_rev_1_canon_eos_c300.pdf - page 11, where he shows part of the C300 Bayer pattern with a 1080 grid superimposed. Alan describes the process as:

Figure 3 shows part of the Bayer-pattern of photo-sites, with a grid structure for 1920x1080 superimposed. Clearly, within the area for each video pixel, there is one red photo-site, one blue, and two green. ....... Given this simple relationship between photo-sites and pixels, the camera must be simply reading out the red and blue sites as red and blue pixels, and summing or averaging the two greens. Thus, each of red, green and blue have the same spatial resolution, unlike the performance when the Bayer pattern is decoded in an attempt to produce higher resolution. I have not seen this done in any other camera.

It is very rare to see resolution so clearly as this, which is very encouraging.

Does that make it clear?

It's worth looking at p12 as well, (Figure 7) which shows the Bayer pattern with a 3x3 grid superimposed. If a similar process was done here (averaging corresponding R,G,B photosite values within each block) you'd correspondingly get a pattern of 1280x720 blocks - in other words a directly read 720p output. Alan describes the theory and potential issues in that link.

To this thread, the relevance is that it is exactly this which I think the A7R ii may be doing to derive 1080 from a s35 centre crop. Which means starting with 5760x3240 photosites in the centre of the sensor and binning in a 3x3 fashion to derive 1920x1080.

And that is why I am so interested to compare crop factors between the s35 HD and 4K modes. If my thinking is correct, the s35 1080 mode will be cropped less than the s35 4K by the amounts I previously said. That would explain Sonys statement that it bins to derive s35 1080, but does a conventional deBayer (off a center crop of the chip) to get s35 4K.

...... nyquist generally suggest sampling at twice the highest frequency so wouldn't you need 4 times the number of photo sites not just double? Double for your super pixel then double again for sampling. Whereas in a normal debayer because you're dealing with single pixels then the oversample can be just double.

Careful, because I think you're confusing "lines" of resolution with "line pairs". Nyquist says that you have to sample at twice the highest frequency (so 20,000 times a second to correctly sample up to 10kHz for audio), but for spatial sampling the equivalent to "frequency" is to talk about line pairs - not "lines".

So (ignoring colour and Bayer filters), a monochrome sensor of 1,000 photosites across (so 1,000 samples) will resolve up to 500 line pairs ...... or 1,000 lines.

 

[paulcurtis]

I don't know what you saw, but the C300 has never shown severe aliasing as a result of it's front end, and doesn't now. I don't expect you to take just what I say at face value, but look at https://tech.ebu.ch/docs/other/WHP034-add70_rev_1_canon_eos_c300.pdf - especially pages 11 & 12. Those are Alan Roberts tests on the C300 for the BBC, hosted now by the EBU, and whilst I wouldn't like to say there is NO aliasing, it's pretty small and unlikely to show up on normal pictures.

Well this for starters

http://www.nextwavedv.com/wp-content/uploads/2012/03/scarlet-vs-C300.png

A simple google will show some truly awful effects. I don't know whether the more recent firmware dealt with this, or rather mitigated it a bit but that's unusable IMHO. Worse than any artefacts on the Sony cams.

When i was pondering looking at the C300 this, and many other examples were a no-no at the time. I'm pretty sure this is a direct result of the approach canon took.

cheers
Paul

 

[paulcurtis]

Careful, because I think you're confusing "lines" of resolution with "line pairs". Nyquist says that you have to sample at twice the highest frequency (so 20,000 times a second to correctly sample up to 10kHz for audio), but for spatial sampling the equivalent to "frequency" is to talk about line pairs - not "lines".

So (ignoring colour and Bayer filters), a monochrome sensor of 1,000 photosites across (so 1,000 samples) will resolve up to 500 line pairs ...... or 1,000 lines.

I think where i'm am trouble is the balance between monochrome resolution and colour resolution. Theory vs Real world. I don't care about theoretical mono sensors.

In a bayer arrangement each RGB pixel is spatially in the correct place for the scene being imaged, yes there is reconstruction to determine the value of the missing two colours per pixel but those reconstruction techniques are pretty clever (i did at one point consider writing a debayer for a side project so spent some time looking into it).

In the 2x2 approach each full RGB comes from a spatially slightly different point of the image so when combined into a 'super pixel' then we loose that spatial arrangement of RGB, it become one slightly blurry RGB pixel. Or we get colour aliasing affecting the overall colour of that pixel. This is the crux of my comment that you would need 4 times as many sites because that first 2x2 average is not over sampling, it's doesn't count.

Does that make sense?

So yes this 2x2 approach can work, but you'd be looking at 4 times as many pixels. Doesn't the FS35 have something like this for it's HD output?

After all bayering exists for a reason.

cheers
Paul

 

[harddrive]

I think where i'm am trouble is the balance between monochrome resolution and colour resolution. Theory vs Real world. I don't care about theoretical mono sensors.

Theory is the same for Bayer and monochrome, and it does translate well into the real world, honest. But first envisaging what Nyquist means without the complication of Bayer filtration just makes it easier to grasp the basics - in this case that you only need one sample per resolved line, two per line pair. Not two per line as your previous post implied. Once you get that clear then you can extend it to a Bayer example. So in my 1,000 photosite example, if you put a Bayer pattern on, you'd only have 500 red sites (hence samples) on the line. So red (and blue) resolution can NEVER be better 500 lines in this case. (Or 250 line pairs if you prefer.)

In the 2x2 approach each full RGB comes from a spatially slightly different point of the image so when combined into a 'super pixel' then we loose that spatial arrangement of RGB, it become one slightly blurry RGB pixel. Or we get colour aliasing affecting the overall colour of that pixel. This is the crux of my comment that you would need 4 times as many sites because that first 2x2 average is not over sampling, it's doesn't count.

Does that make sense?

It makes sense in so far as I see where you're coming from, but there's a flaw in the argument. Take again the F3 and C300 as examples. Two s35 size sensors, both giving a 1920x1080 output, but the c300 with 3840x2160 photosites, the F3 with 2444x1374 (I believe). Now look again at Alan Roberts diagram of the C300 Bayer pattern, with the 1920x1080 grid superimposed, and imagine the F3 pattern superimposed over the top of that. Each of those photosites will be much bigger than the ones of the C300 (hardly surprising), so it follows that adjacent red sites (say) will be much further apart than in the C300 example as well. Which leads to an even higher degree of spatial uncertainty about where red will be.....

OK, that's before deBayering, but there's a limit to how much such can reconstruct. I'm not actually disagreeing with what you say in the paragraph above, but it must be remembered that in *output* terms the red and blue photosites are only 1/2 pixel (a 1080 pixel) apart. Practically, that's not going to be very visible on the vast majority of images. And the overall effect unlikely to be much different than for an F3 type sensor fully deBayered.

As far as coloured aliasing goes, then imagine a scenario of horizontal lines - alternately white-black - spaced exactly the same as the photosite dimensions. It will follow that alternate rows will get white light, the others none. Now since rows will be alternately G,R,G,R,G .... *OR* G,B,G,B,G .... it would follow that you will get a case of either only green and red photosites giving an output or only green and blue. No deBayering algorithm could possibly tell the difference between white-black lines with those characteristics, or the image being solely magenta or cyan light. Which sounds like a recipe for coloured aliasing......

And the theory can be seen very well in practice. Look at a F3 chart made by Adam Wilt - http://www.provideocoalition.com/ag-af100_and_pmw-f3_on_the_charts/page-2 - second chart down, with the yellow/cyan patterning on the horizontal and vertical blocks between about 1250 and 1500 lpph. Now what could be causing that....? Ah, the sensor has 1374 (approx.) rows.... funny that!?!

At this point look at equivalent tests done by Adam for the C300 - http://www.provideocoalition.com/quick_review_canon_c300_super35mm_lss_cine_camera/page-4 . You can see similar yellow/cyan lines (and for the same reasons as above) - but they are now only caused by much finer detail of about 2000 lpph and above. (And could be eliminated much more easily by an OLPF.) Flicking between the charts, then whilst neither is perfect, I'd say the C300 has the lesser of the imperfections. (And neither has anything like enough to give much in the way of real life problems.)

I don't know what has caused the problem in the image you linked to in the other post, but with both Alan Roberts and Adam Wilt coming to pretty much the same answer in their tests, it is not anything to do with the basic principles of the 2x2 read used by the C300. I did put it into Photoshop to examine, and as far as I can tell the blue dots are way, way too big to be caused by any mechanism such as you mentioned. They are many pixels across - what you describe would only give an effect of a single pixel at most.

So yes this 2x2 approach can work, but you'd be looking at 4 times as many pixels. Doesn't the FS35 have something like this for it's HD output?

If you are referring to 4x4 reading, then even greater oversampling will certainly reduce aliasing even more - but it's a question of where to draw the line. For 4K (or rather QFHD) it would mean a sensor of 15,360x8,640 very small photosites! Hmmm. And I wonder just how much better it would look in reality. But if my theory about the A7R ii is correct, 4x4 direct read sampling is exactly how it may derive a 1080 output from the full frame of the sensor.

As for the F35, then sorry, but it's actually worse than what the C300 does. Same principle of direct read of a "super-pixel", but in this case each consists of six photosites in three stripes, one each for R,G, and B in a corresponding "output pixel" block. See http://notesonvideo.blogspot.co.uk/2011/01/sony-f35-another-unusual-sensor.html . And yes, extreme pixel-peeping can see the effect of slight colour fringing if you REALLY go looking. Practically, using 2x2 on Bayer is actually a bit better in this respect.

 

[paulcurtis]

It makes sense in so far as I see where you're coming from, but there's a flaw in the argument.

I'm not disagreeing although i do 'feel' that for the 2x2 approach to work it needs more than double, but that's a IMHO. I understand the argument for pixel density of one large pixel vs 4 smaller ones for example. (although i believe a single big pixel will be more sensitive because it's not just surface area but all the electronics too and 4 pixels has more than 1. Hence the A7s being less noisy that the A7r for example).

Taking a step back from the trees to look at the forest though. Whilst at a pixel peeping perspective this is all good and well the F3 produces a nicer image than the C300 and there are many other places in the chain that affect this. The F35 is way better than the F3 and so on. These artefacts on the C300 prevented it of being any interest to me regardless of how sharp it's been tested as.

Back onto topic. The A7r may well be doing this, but the simple fact remains it still suffers some bad rolling shutter and therefore on a practical - could i take it out and shoot with it - level, it's not ideal. Also the output is still not good enough (8 bit slog).

My main camera is the FS700 at 4K RAW (which using a proper RAW pipeline i love the results), i have used the A7s at UHD as well which has some advantages in form factor but i'd only use when i have too.

However more recently i've been shooting vfx plates on the A7s with time-lapse and RAW stills. If *only* we had a way to get what this camera is capable off in motion. These stills are vastly superior in colour, range and quality. It's such a shame..

cheers
Paul

 

[ReelFilmGeek]

Finally looked at the video and skipped the thread (sorry but 24 pages is a lot to read and I just bought the D750 and haven't been as interested in cameras) but wow it looked good. I do a lot of studio photography first then video so this might actaully be something I can see picking up in a year or so (I rarely buy camera bodies new anymore due to how fast they get outdated). I was worried somewhat about low light but it seems useable at 3200 and that works for me. Would like to see a little bit more footage pushing the dynamic range but if its anything like the a7s we will be in good shape.

 

[harddrive]

Whilst at a pixel peeping perspective this is all good and well the F3 produces a nicer image than the C300 and there are many other places in the chain that affect this.......These artefacts on the C300 prevented it of being any interest to me regardless of how sharp it's been tested as.

I think I'd need to know exactly how those C300 samples that you refer to came about before drawing any such conclusions. From the nature of the image you linked to, then whilst I fully agree it has ...... errr, shall we say problems (!), I'd be very happy to put a very large bet they have nothing to do with what we've been talking about regarding 2x2 and deBayering. They are far too big - cover far too many pixels - and my suspicions would be far more down to something like a discrepancy between how the different colour channels are aligned. (Either due to a fault or misalignment.)

Frankly, I take "real world" examples posted on line with a huge pince of salt at the best of times, and that's especially true when the results are at total variance with tests such as those done by Adam Wilt and Alan Roberts.

Back onto topic. The A7r may well be doing this, but the simple fact remains it still suffers some bad rolling shutter and therefore on a practical - could i take it out and shoot with it - level, it's not ideal. Also the output is still not good enough (8 bit slog).

In terms of rolling shutter, then again, all that's been seen so far is far from scientific testing. I think it's worth waiting for more definitive tests. That said, what has been said so far has related to the s35 4K mode, and that's why all this geek talk about readout may have practical relevance. On it's basis I still strongly suspect that the other modes will have much *less* rolling shutter and in that respect at least may be fully acceptable to you. But let's wait for some accurate testing.....

As far as the more general point goes, then .... well, it's a hybrid stills/video camera, and I'm afraid that in operational etc terms I don't expect as much as from a true video camera. End of story. I have a RX10, which I basically use for still photography, and it's great. And looking at it's video mode I have to admit that in sheer quality terms it's on a par with what you'd have had to have paid several tens of thousands of £'s to get less than 10 years ago. But in terms of using it for filming I'd prefer not to - not for quality reasons, but because it's first and foremost a stills camera.

So whilst I'm intrigued by the new sensor in the A7R ii, and suspect it's (technically) a big leap forward, and (practically) means for the first time a single sensor can give quality 4K/1080, for both full frame and s35 sizes - then first and foremost the camera is primarily designed to be for still photography.

 

[DLD]

I have a RX10, which I basically use for still photography, and it's great. And looking at it's video mode I have to admit that in sheer quality terms it's on a par with what you'd have had to have paid several tens of thousands of £'s to get less than 10 years ago...

Technology marches on, you know.

 

[paulcurtis]

I think I'd need to know exactly how those C300 samples that you refer to came about before drawing any such conclusions. From the nature of the image you linked to, then whilst I fully agree it has ...... errr, shall we say problems (!), I'd be very happy to put a very large bet they have nothing to do with what we've been talking about regarding 2x2 and deBayering.

I don't own one so can't test, but they are single pixel artefacts, there are many examples either under aliasing or fringing. Just looked now. I do believe that a firmware update has lessened it now though. You may well be right, it may be a software interpretation but my guess is that the firmware update special-cased this circumstance and kludged it. I still suspect it has something to do with the sensor, but i have no way of knowing for sure. Just a hunch

I find some reviews a little overly gushing and don't always reflect real world.

This test http://www.alfonsoparra.com/images/articulos/investigacion/hd/pruebas_Alpha7sTest/A7Spruebas.pdf

is a good one, this guy knows his stuff and shoots for real.

I think it shows that there's a huge difference from theory to practise and many parts of the chain. So in theory 13 stops, but in real life 9. IMHO if you shoot UHD, denoise and downsample you can get 10/11 stops out of it, but nowhere near 13 unless you're shooting a test chart. I think one review claimed 14, which is actually impossible in slog2. The other take aways are very finicky colour response, these are somewhat more problematic than resolution. His comments on resolution again, are probably more to do with XAVC. Shoot UHD downsample and it is up there with other cameras, but still the FS700 is sharper in 4K.

I would urge Sony to sort the bl**dy colour out...

cheers
Paul

 

[LeoGrey]

Well this for starters

http://www.nextwavedv.com/wp-content/uploads/2012/03/scarlet-vs-C300.png

A simple google will show some truly awful effects. I don't know whether the more recent firmware dealt with this, or rather mitigated it a bit but that's unusable IMHO. Worse than any artefacts on the Sony cams.

When i was pondering looking at the C300 this, and many other examples were a no-no at the time. I'm pretty sure this is a direct result of the approach canon took.

cheers
Paul

Shot hundreds of projects with the c300 and what you speak of, I have never seen in any of my(or colleagues) footage. I understand this is not the Canon forum, but let's please be a little more realistic about one of the most dominant cameras ever created.

 

[paulcurtis]

Shot hundreds of projects with the c300 and what you speak of, I have never seen in any of my(or colleagues) footage. I understand this is not the Canon forum, but let's please be a little more realistic about one of the most dominant cameras ever created.

I did say that it seems that this was mitigated in firmware, but you can't deny there was a problem at the beginning. I raised it as a possible technical artefact from the 2x2 approach discussion - not bashing anything. just evidenced.

But this is a Sony forum, we've made decisions that led us here. The EF mount is enough to, personally, never go down that route as i loose access to myriads of adaptable lenses. That's just being practical.

cheers
Paul

 

[ReelWorksMedia]

Copied this from REDuser:


Great news if this is correct! I only heard the A7RII sends out an 8-bit 4:2:2 signal and the Shogun records in 10-bit but I think planet 5d should know better so Im hoping its true.

http://blog.planet5d.com/2015/06/ato...ew-sony-a7rii/
This means that straight out of the gate, the a7R II will be capable of bypassing its internal compression in favor of the much sturdier, 10bit ProRes and DNxHD codecs.


"While this is solid performance for a camera as small as the A7R II, Sony leaves room for users to go a little more professional by enabling uncompressed HDMI output of the sensor’s 4K signal.

This allows devices like the Atomos Shogun to capture 4K video to 10-bit ProRes or DNxHR codecs at an impressive 880Mbps, which provides a loads of benefits over the A7R II’s internal 50:1 compression."
http://www.photographybay.com/2015/0...video-capture/