Masters of darkness - Panasonic GH5S and Sony A7R III in extreme low light

[horshack]

  On 2/2/2018 at 3:29 AM, Robert Collins said:

BTW I thought I would throw in the A7riii and A7rii comparison here because the difference is so striking (in FF video). I am not sure I see or understand where Sony found 3 extra stops of lowlight performance out of the same sensor,

 

Expand  

It's because Sony read the sensor out at 1/15 in FF mode on the A7rII, which isn't fast enough to support video frame rates, so it employed line skipping, which throws out both light-gathering ability and sharpness as well (soft video). On the A7rIII Sony reads out the sensor at 1/30, which is fast enough to support video frame rates without needing to toss out rows.

[Andrew Reid]

  On 2/2/2018 at 5:38 AM, horshack said:

It's because Sony read the sensor out at 1/15 in FF mode on the A7rII, which isn't fast enough to support video frame rates, so it employed line skipping, which throws out both light-gathering ability and sharpness as well (soft video). On the A7rIII Sony reads out the sensor at 1/30, which is fast enough to support video frame rates without needing to toss out rows.

Expand  

I don't think this is right.

I don't know precisely the difference in readout but the A7R 2 and 3 are pretty similarly detailed in full frame 4K, both with a similar amount of moire. The A7R II is not a lot softer.

The front-end LSI processing is certainly different though. Image processor is a lot better in the A7R3... Noise reduction on a different planet.

What source did you obtain the timing of the readout from? 1/15 seems too slow to me. The timing of the readout from top to bottom needs to be fast enough to support 30p at minimum shutter speed of 1/30, without frames doubling up.

[Robert Collins]

  On 2/2/2018 at 1:13 PM, Andrew Reid said:

What source did you obtain the timing of the readout from? 1/15 seems too slow to me. The timing of the readout from top to bottom needs to be fast enough to support 30p at minimum shutter speed of 1/30, without frames doubling up.

Expand  

http://blog.kasson.com/

This engineer has done extensive testing of the readout speed of Sony sensors.

A7r2 1/15 second

A7r3 1/30 second

A9 1/160 second but only in stills (not in video) (which could be very interesting in terms of the A7siii)

In terms of your above comment, he tests full sensor readout for stills. So the readout for a proportion of the sensor could be faster (I guess.)

[teddoman]

  On 2/2/2018 at 1:22 PM, Robert Collins said:

http://blog.kasson.com/

This engineer has done extensive testing of the readout speed of Sony sensors.

A7r2 1/15 second

A7r3 1/30 second

A9 1/160 second but only in stills (not in video) (which could be very interesting in terms of the A7siii)

In terms of your above comment, he tests full sensor readout for stills. So the readout for a proportion of the sensor could be faster (I guess.)

Expand  

Do you recall if he measured the readout speed for the A9 in video?

[Robert Collins]

  On 2/2/2018 at 2:40 PM, teddoman said:

Do you recall if he measured the readout speed for the A9 in video?

Expand  

No he didnt test the sensor readout for video.

But we know (for the A9) it is much lower than for stills - in fact by the rolling shutter tests I have seen, the readout is the same as the A7riii.

So did Sony intentionally cripple the A9 for video (quite probably) and was the reason they didnt include picture profiles because they deliberately didnt want people buying the A9 for video expecting that the fast electronic shutter would lead to less rolling shutter (again quite possible but just guessing.)

Still if a 24mp A7siii or a 12mp A7siii takes advantage of the incredibly fast readout from a chip like this, it would virtually eliminate rolling shutter on a FF sensor....??!!

[teddoman]

  On 2/2/2018 at 3:11 PM, Robert Collins said:

No he didnt test the sensor readout for video.

But we know (for the A9) it is much lower than for stills - in fact by the rolling shutter tests I have seen, the readout is the same as the A7riii.

So did Sony intentionally cripple the A9 for video (quite probably) and was the reason they didnt include picture profiles because they deliberately didnt want people buying the A9 for video expecting that the fast electronic shutter would lead to less rolling shutter (again quite possible but just guessing.)

Still if a 24mp A7siii or a 12mp A7siii takes advantage of the incredibly fast readout from a chip like this, it would virtually eliminate rolling shutter on a FF sensor....??!!

Expand  

The only A9 numbers (in video) I've seen are in a review by Imaging Resource which puts the readout speed at 1/40 - 1/75.

[horshack]

Regarding what actual low-light improvements have been made to the GH5s sensor (vs just video noise reduction), the best way to suss that out is to look at the raw stills performance compared to its predecessor. Any noise improvement in video which is not also demonstrated in stills performance can only be the result of noise reduction (presuming the predecessor's video processing didn't have its own faults/limitations, such as the need to sub-sample/skip lines due to limited readout performance, which the GH5 didn't have). With that said, here's a comparison from dpreview of the GH5s vs GH5 for ISO 1600 - they don't currently have higher ISOs depicted because they've only done ISO invariance testing so far.

Dpreview ISO Invariance Widget, ISO 1600, Normalized to common image resolution

And here's a comparison of the base ISO DR:

Dpreview ISO Invariance Widget, Base ISO, Pushed +6EV, Normalized to common image resolution

[anonim]

Quick noise reduction (Resolve 14) operation with GH5 ISO 6400 400% magnification fragment of Andrew Read's file.

 

[Stab]

I keep reading different crop factors for the GH5s.

Does anyone have the facts and the numbers? What would the crop factor be for the GH5s vs the GH5 for 16:9 and 17:9 modes?

[ntblowz]

  On 2/4/2018 at 5:12 PM, Stab said:

I keep reading different crop factors for the GH5s.

Does anyone have the facts and the numbers? What would the crop factor be for the GH5s vs the GH5 for 16:9 and 17:9 modes?

Expand  

from another site

Crop values

3680 pixel long in 4/3 = 2x crop factor in 4/3 still mode

3840 pixel long in 3/2 = 1,91x crop factor in 3/2 still mode

4016 pixel long in 16/9 = 1,832x crop factor in 16/9 still mode

4096 pixel long in 17/9 = 1.796x crop factor in 17/9 (DCI 4K) video mode

3840 pixel long in 3/2 = 1,91x crop factor in 16/9 4K UHD video mode.

Speed Booster Ultra + GH5s + DCI 4K = crop factor of 1,275x with 35mm lenses.

Speed Booster XL + GH5s + DCI 4K = crop factor of 1,149x with 35mm lenses.

[Stab]

Thanks ntblowz.

And what about 16:9 in video mode?

[Andrew Reid]

I still think something is wrong with the crop factor maths

If 4096 is 1.796 that would make my BMPCC 0.58x speed booster 1.04x crop, i.e. hardly any different to full frame, yet as I have already tried, much to my surprise this is not the case, looks more like 1.2x. And the XL (I have that too) is 0.64x, and looks like 1.3x crop not 1.149x.

Maybe the sensor is only 17mm wide after all?

[PannySVHS]

  On 2/5/2018 at 12:06 AM, Andrew Reid said:

Maybe the sensor is only 17mm wide after all?

Expand  

Or the regular mft sensor only provides us with 15.5mm. What a treachery!:)

[Robert Collins]

  On 2/5/2018 at 12:41 AM, PannySVHS said:

Or the regular mft sensor only provides us with 15.5mm. What a treachery!:)

Expand  

Most likely it is just the way that people are calculating the crop factor. The crop factor for M43 is 'considered' to be 2.0. But it is a 4:3 sensor while a 35mm sensor is 3:2. The crop factor is based on the diagonal but there is more of a crop on the width and less of a  crop on the height with a 4:3 sensor.

As in crop factor width = 36/17.3 = 2.08x

And crop factor height = 24/13 = 1.85x

[horshack]

Dpreview just posted their GH5s review, which includes a full set of High ISO raw stills for their IQ comparison widget. Here are few comps I generated:

GH5s vs GH5, ISO 6400, Low-Light, Normalized to Common Image Resolution

GH5s vs GH5, ISO 12800, Low-Light Normalized to Common Image Resolution

GH5s vs GH5, ISO 25600, Low-Light, Normalized to Common Image Resolution

[Andrew Reid]

  On 2/5/2018 at 1:42 AM, Robert Collins said:

Most likely it is just the way that people are calculating the crop factor. The crop factor for M43 is 'considered' to be 2.0. But it is a 4:3 sensor while a 35mm sensor is 3:2. The crop factor is based on the diagonal but there is more of a crop on the width and less of a  crop on the height with a 4:3 sensor.

As in crop factor width = 36/17.3 = 2.08x

And crop factor height = 24/13 = 1.85x

Expand  

Well this is what is so confusing. Image framing on a multi-aspect ratio sensor goes wider when you select a wider aspect ratio. To support this vs GH5, the GH5S sensor should be a bit wider than 17.3mm... more like 19mm wide. Then, if we do 36mm / 19mm we get 1.89x crop factor horizontally, which would make sense as you see more of the lens than on the 2x crop GH5, yet apply the 0.58x speed booster and it should be a 1.1x crop horizontally compared to full frame... and it just isn't... the field of view is significantly less wide horizontally.

So something is weird...

[PannySVHS]

If 3680 px width for 4 to 3 format photos goes along with sensor width of 17.3mm, then 4096px for C4K should have 4096px / 3860px x 17.3mm =19.25mm sensor width. Giving us a crop factor compared to FFrame as follows: 36mm/19.25mm = 1.87. Giving us theoretically a speedboosted factor of 1.87x0.58=1.08.

So math is right. But in practice it might be different, because the optics of the Bmpcc speedbooster work to its full extent with a factor 0.58 on the Bmpcc only. The larger the difference in sensor size goes, compared to the BMPCC s16 sensor, the less boost it´s giving us.

  On 2/5/2018 at 4:39 PM, horshack said:

Dpreview just posted their GH5s review, which includes a full set of High ISO raw stills for their IQ comparison widget. Here are few comps I generated:

GH5s vs GH5, ISO 6400, Low-Light, Normalized to Common Image Resolution

GH5s vs GH5, ISO 12800, Low-Light Normalized to Common Image Resolution

GH5s vs GH5, ISO 25600, Low-Light, Normalized to Common Image Resolution

Expand  

They show the 1.66 increase in sensitivity as advertised by Panasonic. In video it is rather a difference of 2-2.5 ISO steps in sensitivity, giving us better ISO 6400 than GH5 at 1600 ISO. They should come up with a video comparison as well.

[anonim]

  On 2/5/2018 at 8:40 PM, PannySVHS said:

... giving us better ISO 6400 than GH5 at 1600 ISO.

Expand  

I'd say it is far away from true. From all samples I saw - including that last one from Max Yurev - for me it seems that at iso3200 video frames from GH5 have obviously more details than GH5s. It looks like GH5s has some enormously agresive noise reduction algorithm deep inside of codec usage. Testing samples, I found that up to iso 3200, GH5 with simple temporal noise reduction values in Resolve 14 keeps more image resolution than GH5s. Yes, GH5 is completely useless at iso6400, but level of plasticity in GH5s image at the same iso 6400 value simple is not at all for professional use as advertised.

[PannySVHS]

From what I have seen GH5s shows much better color fidelity at low light and better overall image, an advantage of two iso steps, 6400 perfectly usuable. That is quiete the opposite to your findings Give extrashot on youtube a try.

 

  On 2/5/2018 at 9:15 PM, anonim said:

I'd say it is far away from true. From all samples I saw - including that last one from Max Yurev - for me it seems that at iso3200

Expand  

 

[anonim]

  On 2/5/2018 at 9:43 PM, PannySVHS said:

 Give extrashot on youtube a try.

 

Expand  

I gave a try... and I'm not at all convinced to Panasonic marketing states. Quite contrary, I saw lost of resolution in every comparative case up to iso 3200. What is called cleaner image is result of smoothing image... Out of discussion is that iso6400 really looks laptopic usable - but not at higher professional level or for central shots (simply, skintones are too smooth and unnatural, faces are cartoonish, without 3d natural look). But, to answer to your call: give a try, if you not did it already, to a very simple - and very fast - temporal noise reduction node in Resolve 14... and judge for yourself.

My conclusion is that Panasonic's technicians has long way ahead to really mastered new sensor. They hurried - and maybe that is the reason of first delay of GH5s. Next incarnation of new sensor might be very promising - but this one (with nonsense or pretentious explanation for cutting the ibis, with presumptuous price) is simply pilot-one for heavily GAS addicted persons, and as such, potential fiasco.