Re: Sony A7RII - FIRST video sample!!!

[paulcurtis]

FWIW as i mentioned in another thread there's some stabilisation software from prodad that will tackle rolling shutter artefacts, including jello and skew. It does a magical job at times and i would run A7s footage at 4K through it then down-sample. You can grab a demo version.

I have nothing to do with the company except being a happy customer, i also have access to a lot of tools but even so this still pulls through the best.

cheers
Paul

 

[Samuel H]

...

Agreed, for almost everything you said in your post.

* 5D3: I believe it skips rows, then averages same-color photosites within each row. I concluded this because the video RAW files have more noise than a full-res RAW still downsampled to 1080p using bilineal (which skips nothing and is not particularly smart), but more than one downsampled using nearest neighbor (which skips both lines and columns). It is in fact very similar to the amount of noise you get when you first do a vertical downsample with nearest neighbor, then a horizontal-only downsample with bilineal. But it's just an observation, maybe they're doing a full-readout with the ADCs working faster with a higher read-out noise or less resolution, who knows.

* Analogue binning: I know some photosite designs that allow for this, but from my understanding, it's far from common to implement them. It would be great if it is there. 7680x4320 to 3840x2160 through analogue binning would look absolutely stunning. But I don't think it is there, because there wouldn't be any aliasing/moire issues (well, hardly any; you'd probably need really sharp glass and a chart to make it appear), and because the high-ISO performance would be better than in the APS-C mode (I'm not sure but from the newsshooter bit it seems -the wording is not clear- that FF has more noise than APS-C, which would be in line with APS-C downsampling from 15 Mpix but FF doind lineskipping, then throwing away columns, and debayering with just 8 Mpix).

 

[photochris]

I'm really surprised there aren't more samples popping up since the Camera ships in about 3 weeks. It would be nice if some of the questions posed could be answered before the camera is released.

 

[photochris]

It lacks fast zooms and fast teles, as for instance. It beats on the lightweight department, though.

There's always the A-mount glass on the LAEA3, they have the "big 3" f/2.8 zooms (16-35, 24-70, 70-200) and some good long lenses. They won't be any smaller when they're eventually remade into E-mount lenses.

 

[DLD]

I'm really surprised there aren't more samples popping up since the Camera ships in about 3 weeks. It would be nice if some of the questions posed could be answered before the camera is released.

I thik RX100II reviews got out over the last couple of days. RX10II is next and then it'll be A7RII turn.

There's always the A-mount glass on the LAEA3, they have the "big 3" f/2.8 zooms (16-35, 24-70, 70-200) and some good long lenses. They won't be any smaller when they're eventually remade into E-mount lenses.

Those zooms also just got updated.

 

[harddrive]

7680x4320 to 3840x2160 through analogue binning would look absolutely stunning. But I don't think it is there, because there wouldn't be any aliasing/moire issues........

The best account of the camera I've so far read is the newsshooter account, and that actually says: "The second big surprise is that the full-frame 4K mode also looks remarkably good. .......... The 4K full-frame output of the a7R II is really excellent. There are some traces of aliasing visible on lines of high contrast, but you really have to look for them." I think unfortunately there has been a lot of misinformation about regarding this mode. People have heard "it bins in FF mode" and just assumed "the FF mode must therefore be inferior". Well, let's wait for more definitive REAL tests.

But if it's doing as I imagine, then each output pixel gets simply formed from a 2x2 block of one red, one blue, and two green photosites. And 7680x4320 photosites translates directly to 3840x2160 blocks - so for each block the only binning that needs to happen is two green photosites at a time. Does that sound feasible? Heck, it could even ignore one green value per block, and simply translate the values of the three remaining directly into the R,G,B values of the 4K pixel! The processing could hardly be simpler, which may help to explain a fast read out time, and much less rolling shutter in this (FF) mode than s35 (when it has to do full deBayering and scaling.)

But let's think even more about the geometry. Let's assume it is using 7680x4320 photosites, and think what if it did do the "read two, skip two" approach for lines and columns? It would mean you'd have a total of 1920x1080 blocks of photosites, each 2x2, and three times as many not read. Out of each block of 16, only 4 would be read.

But how to then process that? The simplest way would be to simply take the values for red, green, and blue for each block as above but that would only give 1080 output, not 4K. What about the processing doing a standard deBayer, "pretending" the skipped lines and columns aren't there, pretending the 8Mp read photosites are as a normal 8Mp sensor? I don't see it somehow. I suspect the aliasing would be awful, and it wouldn't really work well anyway.

So the idea of reading all the photosites and doing a simple 2x2 direct read might actually be the EASIEST thing to do! And if a simple method would give the highest quality output, then why on earth do anything else!?!

.......and because the high-ISO performance would be better than in the APS-C mode (I'm not sure but from the newsshooter bit it seems -the wording is not clear- that FF has more noise than APS-C, which would be in line with APS-C downsampling from 15 Mpix but FF doind lineskipping, then throwing away columns, and debayering with just 8 Mpix).[/COLOR]

I may be missing it, but I don't see that result in the newsshooter article? Closest I've seen is "What surprised me is that the a7R II 4K S35 image was less noisy than the a7S APS-C 4K image ......" - but that's comparing the A7R II to the a7S.

 

[cpc]

But if it's doing as I imagine, then each output pixel gets simply formed from a 2x2 block of one red, one blue, and two green photosites. And 7680x4320 photosites translates directly to 3840x2160 blocks - so for each block the only binning that needs to happen is two green photosites at a time. Does that sound feasible? Heck, it could even ignore one green value per block, and simply translate the values of the three remaining directly into the R,G,B values of the 4K pixel! The processing could hardly be simpler, which may help to explain a fast read out time, and much less rolling shutter in this (FF) mode than s35 (when it has to do full deBayering and scaling.)

But let's think even more about the geometry. Let's assume it is using 7680x4320 photosites, and think what if it did do the "read two, skip two" approach for lines and columns? It would mean you'd have a total of 1920x1080 blocks of photosites, each 2x2, and three times as many not read. Out of each block of 16, only 4 would be read.

But how to then process that? The simplest way would be to simply take the values for red, green, and blue for each block as above but that would only give 1080 output, not 4K. What about the processing doing a standard deBayer, "pretending" the skipped lines and columns aren't there, pretending the 8Mp read photosites are as a normal 8Mp sensor? I don't see it somehow. I suspect the aliasing would be awful, and it wouldn't really work well anyway.

So the idea of reading all the photosites and doing a simple 2x2 direct read might actually be the EASIEST thing to do! And if a simple method would give the highest quality output, then why on earth do anything else!?!

What you describe is a full sensor readout sans proper debayering. You still need to read all the photosites, and there is no binning at all. It would have rolling shutter slower than the s35 mode, not faster.

 

[Policar]

Agreed, for almost everything you said in your post.

* 5D3: I believe it skips rows, then averages same-color photosites within each row. I concluded this because the video RAW files have more noise than a full-res RAW still downsampled to 1080p using bilineal (which skips nothing and is not particularly smart), but more than one downsampled using nearest neighbor (which skips both lines and columns). It is in fact very similar to the amount of noise you get when you first do a vertical downsample with nearest neighbor, then a horizontal-only downsample with bilineal. But it's just an observation, maybe they're doing a full-readout with the ADCs working faster with a higher read-out noise or less resolution, who knows.

* Analogue binning: I know some photosite designs that allow for this, but from my understanding, it's far from common to implement them. It would be great if it is there. 7680x4320 to 3840x2160 through analogue binning would look absolutely stunning. But I don't think it is there, because there wouldn't be any aliasing/moire issues (well, hardly any; you'd probably need really sharp glass and a chart to make it appear), and because the high-ISO performance would be better than in the APS-C mode (I'm not sure but from the newsshooter bit it seems -the wording is not clear- that FF has more noise than APS-C, which would be in line with APS-C downsampling from 15 Mpix but FF doind lineskipping, then throwing away columns, and debayering with just 8 Mpix).

You're really overthinking things. The 5D Mark III's noise is largely from the off-chip ADC introducing a lot of read noise. It does on-chip binning to 1920X1080 then read noise is introduced at the same amount per-pixel as you'd get at full read out. So more noise, but not because there's line skipping.

 

[harddrive]

What you describe is a full sensor readout sans proper debayering.

I'd use the term "conventional debayering" instead of "proper". That may sound simply pedantic but both forms will give fully acceptable outputs - it's just that what I describe will need a higher pixel count than a conventional deBayer process. (And needs to be an exact multiple of output resolution.) But for a camera whose stills ability is of as much importance as video ability, that is no hardship if it enables a vastly simpler processing system - you need the high pixel count for the stills resolution.

You still need to read all the photosites, and there is no binning at all. It would have rolling shutter slower than the s35 mode, not faster.

Yes, you need to read all the photosites (and only bin the two green per block) but the question is what is the limiting factor for speed? Is it the actual act of reading off the sensor, or processing downstream of that? If the latter, you could have a situation where the sensor has to be read slower than it's capable of, because the following processing is creating a bottleneck. A bit like a traffic jam as three lanes are forced to merge into one. *IF* that was the case, then having to read all the photosites on the chip - but being able to process them very quickly - could indeed be faster than reading far fewer (the s35 crop) but complex deBayering etc slowing the process down.

To follow the analogy through, it's like the average speed of a lot of traffic moving along a 3 lane motorway could be a lot faster than a smaller number of vehicles merging from three lanes into one.

 

[cpc]

You're really overthinking things. The 5D Mark III's noise is largely from the off-chip ADC introducing a lot of read noise. It does on-chip binning to 1920X1080 then read noise is introduced at the same amount per-pixel as you'd get at full read out. So more noise, but not because there's line skipping.

The readout noise is the same, but the binned signal has a level times the number of pixels in the bin, so the SNR of the binned signal is greater. This is exactly the original motivation of the binning concept: to reduce the apparent effect of readout noise by decreasing the number of a/d operations per frame.


@harddrive:
What you describe is exactly how the Canon C100/300/500 cameras work when deriving a 1080p image from a 4K sensor, btw. (If you look at some aliasing test samples from these cameras you can see how the red channel manifests more aliasing than the green for this very reason. Same for the blue channel, but it is perceptionally darker, so less noticeable.)
Your analogy is a fancy way of saying that a pipeline works at the speed of its slowest component. Unfortunately in the case of the a7rII, even with additional ADC units (were Sony touting additional ADCs for the a7rII or the new rx cams? can't recall) increasing the total number to more than one per column, I think the readout is the most likely bottleneck.

 

[harddrive]

@harddrive:
What you describe is exactly how the Canon C100/300/500 cameras work when deriving a 1080p image from a 4K sensor, btw.

Indeed, and I said as much when I first suggested what may be happening - http://www.dvxuser.com/V6/showthrea...ple!!!&p=1986555224&viewfull=1#post1986555224

It was after the C300 news that it first occurred to me that THEORETICALLY the ideal hybrid stills/video camera would be gained by quadrupling the photosite count of the C300 sensor - to enable stills by normal deBayering, 4K by 2x2 direct read, HD by 4x4 binning and direct read. So it seemed only a matter of time........ And with Sony here talking about "just the right number of pixels", and 42Mp being "better than 45Mp", and the numbers being so close to what was predicted (allowing for a 3:2 sensor, and cropping to get 16:9) everything seems to fit into place?

As for readout speeds, well, I can't be sure. In the press release though Sony do say:

Additionally, the sensor’s back-illuminated structure, with an expanded circuit scale and copper wiring design, enables faster transmission speed and ensures content can be captured in high resolution without sacrificing sensitivity. Data can also be output from the sensor at an approximately 3.5x faster rate compared to the original α7R.

So maybe they've been able to increase the sensor output speed by 3.5x, but deBayer processing hasn't been speeded up by as much? I can only theorise......?

I don't know if you saw it, but earlier in this thread a Sony chart was shown which said it was binning in both FF modes (4K & HD), and also binning for s35 HD, but not for s35 4K - and it's why both the s35 modes aren't the same that puzzled me. The best explanation I can come up with is that for s35 HD, it's binning in 3x3 blocks. As said before, that can be easily proved or disproved (by anyone with access to the camera! ) by comparing horizontal crop factors for the different modes. If I'm right, I'm expecting:

But we need somebody to test this in practice. See what the crop figures are for the different modes. My prediction would be:

mode ------------------------ crop factor --- hor res

FF stills ------------------------1.0 -----------7952
FF video (HD & 4K) -----------1.035 --------7680
s35 HD ------------------------1.38 ----------5760 (NB crop factor 1.33 relative to FF video)
s35 4K -------------------------1.55 ----------5120 (approx. - assuming 14.7Mp, and crop factor then 1.5 relative to FF video)

(See post 157 in this thread if you want the reason.)

 

[DLD]

https://www.youtube.com/watch?v=h9fY3yer6Ww

An AF test with the A-mount lenses.

https://www.youtube.com/watch?v=N1OnaCPVpuU

and with EF 70-200 and Nikon F 70-200

Anyone speaks Chinese?

 

[paulcurtis]

A decent debayer needs the surrounding RAW pixels to work properly, the binning and 'mini debayer' of 4 pixels into one final colour would prevent this. But it does seem like a likely scenario.

WTF can't we just get the raw data off this thing (rhetorical question, i know the answer)...

cheers
Paul

 

[Samuel H]

RAW data out of one of these Sony sensors with tons of DR... that would be sweet, but yes, we all know the answer: they want so sell the more expensive cameras too.

Regarding sensor readout, debayering, crop modes, etc, we should know soon.
My bet is that the bottleneck for 4K FF is the sensor readout and not the processor (so I think we'll have line-skipping and not binning in FF mode: aliasing/moire on charts and with very sharp lenses but not likely to happen in real-world shooting, noise about a stop worse than 4k crop mode) and all crop modes at 1.54x (with line-skipping in 1080p because of lack of registers in the processor for a one-step debayer+big.downsample).

 

[Emanuel]

There's always the A-mount glass on the LAEA3, they have the "big 3" f/2.8 zooms (16-35, 24-70, 70-200) and some good long lenses. They won't be any smaller when they're eventually remade into E-mount lenses.

Yes, the point is I was talking about the E-mount. A-mount is another one, to be Minolta/Konica (Sony today) affiliated doesn't count for the matter : ) And as we can see from the early reports, distinct AF response.

 

[Emanuel]

FWIW as i mentioned in another thread there's some stabilisation software from prodad that will tackle rolling shutter artefacts, including jello and skew. It does a magical job at times and i would run A7s footage at 4K through it then down-sample. You can grab a demo version.

I have nothing to do with the company except being a happy customer, i also have access to a lot of tools but even so this still pulls through the best.

cheers
Paul

Paul, thanks for bringing it to here.

E

 

[harddrive]

A decent debayer needs the surrounding RAW pixels to work properly, the binning and 'mini debayer' of 4 pixels into one final colour would prevent this. But it does seem like a likely scenario.

No, that's not true - what I've been calling "2x2 direct read" will work "properly" - the C300 is testament to that - but the caveats are that compared to conventional deBayering it needs roughly twice as many photosites to give comparable results, and the sensor dimensions must be an exact multiple of the required output resolution. But on the plus side it needs far less processing - which translates into simpler electronics, and likely lower power consumption and heat generation.

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. The trick is to make it up to EXACTLY twice QFHD on each axis (7680x4320).

 

[Samuel H]

I think that's what he said: that many debayer algorithms need extra surrounding RAW pixels but the 4-to-1 thing the C300 does doesn't need that.

 

[harddrive]

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.

 

[Samuel H]