Finally how many AF pixels on the Z9 /Z8 sensors ?

[Restoc]

Does somebody found a proven source from Nikon or expert studies letting us know the real number of AF pixels implemented on the Z9 /Z8 sensors.

I speak and care about real AF pixels on the sensor, not about the 493 Focus Point selectable and displayed in the viewfinders.

The only source that i could find out is the study of Jimm Kasson as a side result of his dark level spectra study here

Jim Kasson has detected spikes every 12 raws lines as evidence of the PDAF pixels and i take it..

So a first question is what about columns ?

Or as a summary what is the matrix figures of AF pixels over the FF frame ? 12*12 , 12*12*1.5 or other ? This would drive to 220 000 up to 300 000 PDAF pixels .... As the D6 scene recognition probably ?

My question is coming from a reverse analysis :

-As a general ( lazy ?) understanding the AF of the Z9 is performed through the video bus and calculated ( with or without IA) using the View finder matrix ( XVGA definition) . With or without IA is not in the scope at this technology level :

The question is consequently : how a so accurate and repeatable focus accuracy at long distance can be achieved using a so low definition matrix ?

Per example when using 800 mm the tick sharp accuracy has to be inside the 3 cm DOF at 10 m. Pretty hard to achieve but it works nicely on the Z9.

My guess is that the final focus accuracy is given not by the AF video bus or viewer matrix but by a final ajustment back using the real AF pixels somewhere.

As Expeed has different input channels for Video, Raw and AF wires, my guess is that Exspeed 7 is not only using the Video bus and viewer matrix for the final positionning of the lens by uses in parallel the far better PDAF accuracy.

As far as i could understand on the Z9 only the Pin Point Focus is able to use elementary AF pixels . All other AF modes use groups of AF pixels including the Single AF point or 3D.

Any knowledge and thoughts warly welcome.

Claude.T

 

[Stavrosf]

Z9 has 45 million pixels and every 12th row is used for AF. That makes around 3.5 to 4 million pixels. Each logical AF point has about 25 rows of 240 pixels so the camera uses around 6000 pixels for each point (the numbers are taken from Thom Hogan guide for Z6/Z7 - Z7 is also 45 mpix so the numbers are similar if not identical). The fact that there are rows skipped means that the horizontal items are identified easier than the vertical (eg there are not crossed sensors like the dslrs).

 

[Alberto Tanikawa]

One could argue the Z8/Z9 have 45.7 million color pixels to sample colors from to feed 3D tracking

 

[Restoc]

Z9 has 45 million pixels and every 12th row is used for AF. That makes around 3.5 to 4 million pixels. Each logical AF point has about 25 rows of 240 pixels so the camera uses around 6000 pixels for each point (the numbers are taken from Thom Hogan guide for Z6/Z7 - Z7 is also 45 mpix so the numbers are similar if not identical).

Thanks.

z6 z7 sensor have a very different photosites reading architecture and the full stacking of Z9 /Z8 makes a great difference. May be better to avoid analogies.

You introduce a wording of "logiczal AF point" is it a recognized information sourced by Nikon or measurement of experts such Jim Kasson or an arthmetic deduction. or adjacent source as Sony ?v

The fact that there are rows skipped means that the horizontal items are identified easier than the vertical (eg there are not crossed sensors like the dslrs).

Well... everybody can read from Nikon in the Z9 handbook (and verify easily) that the row lines parallel to the longest edges generally the horizontal edge, is the weakest Af direction to avoid for thin subjects or textures.

 

[Restoc]

Z9 has 45 million pixels and every 12th row is used for AF. That makes around 3.5 to 4 million pixels.

Where somebody says or show that a row of photosites is a continuous line of AF photosites only ?

I checked through Sony litterature, as well they never speak or show continuous lines of photosites.

Af photosites are more probably a 12*12 or a 12*18 matrix for keeping the 1.5 ratio.

So the best total amount of photosites given by a 12*12 matrix would be only 318 000 AF photosites or 216 000 for a 12*18. Not 3 millions.

I welcome a solid information coming from Sony or Nikon : This is useful to understand how the AF accuracy in focus distance and consequently DoF can be achieved consistently beyond 500 mm on small parts of the frame. Very useful to know for birders and more.

 

[Stavrosf]

My source was Thom Hogan book for Z6 and Z7 for everything that you asked (but I have already written that didn't I). The stacked part of sensor of Z8/Z9 as far as I have read does not change any of the numbers regarding the AF points. Of course the algorithms, the speed etc of the AF is the difference (even the implementation taking in account the dual stream on Z8/Z9). The info I have shared was for Z7. On Z6 there are indeed different numbers. I haven't bought a Z8/Z9 so I do not have the respective book if the exact details exist there.

--
stavrosf
https://www.fourfourianakis.eu

 

[Leonard Shepherd]

One could argue the Z8/Z9 have 45.7 million color pixels to sample colors from to feed 3D tracking

On this sub-thread detail, while the camera has 45.7 pixels, each pixel is covered by a single colour filter - implying 11.4 each for R and B and 22.8 (as there are 2 G) for green.

Digital systems have been reported sampling from up to about 32 surrounding pixels to calculate an exact colour etc.

Probably on Nikon knows how many sampling points current Nikon ML uses.

 

[chambeshi]

"....the total number of pixels on the sensor is up from 46.89 million on the previous design to a whopping 52.37 million for the Z9 sensor. This is pretty unusual among cameras, and no-one outside of Nikon seems to be quite sure what all the extra pixels are for."

*****

"The Z9 had 493 discrete AF points laid out in a grid of 29 x 17 which can be manually selected. The outside row of points is not used for the auto modes, meaning 405 are available for use in the various auto-selection AF modes.

As the image sensor is used for autofocus, the processor can analyse the image content and the opportunity to focus on subjects rather than just points is available for mirrorless cameras. This is used to good effect in the Z9 – not only can you choose to focus on people’s faces or eyes, but it will also detect animals and vehicles for locking and tracking focus."

*****

https://blog.reikanfocal.com/2022/03/reikan-focal-and-the-nikon-z9/

 

[chambeshi]

Nikon Z9 Disassembly And Teardown – Kolari Vision 2025

By Alex Mirchuk

kolarivision.com

 

[SVPhotography]

Z9 has 45 million pixels and every 12th row is used for AF. That makes around 3.5 to 4 million pixels.

Where somebody says or show that a row of photosites is a continuous line of AF photosites only ?

I checked through Sony litterature, as well they never speak or show continuous lines of photosites.

Af photosites are more probably a 12*12 or a 12*18 matrix for keeping the 1.5 ratio.

So the best total amount of photosites given by a 12*12 matrix would be only 318 000 AF photosites or 216 000 for a 12*18. Not 3 millions.

I welcome a solid information coming from Sony or Nikon : This is useful to understand how the AF accuracy in focus distance and consequently DoF can be achieved consistently beyond 500 mm on small parts of the frame. Very useful to know for birders and more.

For CDAF - it uses every row that isn't converted for PDAF

For PDAF - it uses the converted photosites - according to byThom:

"All of Nikon's mirrorless cameras use the old Nikon 1 system: rows of focus-aware photosites, and those rows are spread apart by 12 pixels. That means that focus information on the mirrorless cameras is mostly discriminating on the long axis. The problem is this: if you try to focus on part of a subject that doesn't have much long axis detail but has short axis detail, the Z cameras don't focus as well, if at all. The top DSLRs don't have this issue. "

Hence, there are not cross-type AF points like in the DSLR days. I believe Canon has started to put cross-type PDAF points on the R1 focus system for much improved low light and vertical focus acquisition and tracking.

 

[Bob A L]

Wow!!

 

[Restoc]

"....the total number of pixels on the sensor is up from 46.89 million on the previous design to a whopping 52.37 million for the Z9 sensor. This is pretty unusual among cameras, and no-one outside of Nikon seems to be quite sure what all the extra pixels are for."

*****

"The Z9 had 493 discrete AF points laid out in a grid of 29 x 17 which can be manually selected. The outside row of points is not used for the auto modes, meaning 405 are available for use in the various auto-selection AF modes.

As the image sensor is used for autofocus, the processor can analyse the image content and the opportunity to focus on subjects rather than just points is available for mirrorless cameras. This is used to good effect in the Z9 – not only can you choose to focus on people’s faces or eyes, but it will also detect animals and vehicles for locking and tracking focus."

*****

https://blog.reikanfocal.com/2022/03/reikan-focal-and-the-nikon-z9/

Be careful !

You are making the confusion beetween AF photosites and the Focus Point ( 493) wording of the Users Reference guide. A Focus Point as used in the reference guide is the result of grouping photosites except for the Pin Point CDAF. Even the Single Focus point is not corresponding to only one AF photosite but the result of a computing over several photistes.

Understanding the number of PDAF photosites is useful for ( maths or techno fans) for calculating the mean accuracy achivevable by a Lens at a given distance and playing with very shallow DoF for long lenses. For sure only Long and very long focal users can give importance to these figures.

Understanding Focus points ( 493) and Autofocus modes driving Focus points is a wording dedicated to use the camera : how to use the suitable mode for a given subject moving or still or subject types. Focus Point is the wording useful at Users and the wording used in the handbook. So it's a general use level but brings nothing for OP topic.

Nikon wording in various documentation makes veru well the difference beetween AF photosites and AF Focus points according what they are speaking about : technology or camera handling.

 

[Restoc]

One could argue the Z8/Z9 have 45.7 million color pixels to sample colors from to feed 3D tracking

On this sub-thread detail, while the camera has 45.7 pixels, each pixel is covered by a single colour filter - implying 11.4 each for R and B and 22.8 (as there are 2 G) for green.

Digital systems have been reported sampling from up to about 32 surrounding pixels to calculate an exact colour etc.

Probably on Nikon knows how many sampling points current Nikon ML uses.

We have no evidence that color information is used for PDAF. especially before demosaicking !! Just imposible.

The only information, coming from Sony SC is that PDAF photosites ( not pixels) are only under the green color ( higher sensitivity under visble spectra). So they would work only in Luminance.

May be color is used ( but need confirmation) somewhere from the low resolution Viewer image or more precisely the buffer storing video and jpegs but not the raw data, when using IA. But this is far from the PDAF level necessary to initiate computing for long range focusing.

 

[Horshack]

I discussed a potential explanation here and here.

 

[Restoc]

Thanks for the links.

If we condider only the path from On Sensor AF photosites up to the final focus value transmitted to the lens, for focusing at the right place and distance, a lot of information are really missing to understand what and how AF phase information is used and processed ( with or without IA.

My guess is that IA helps only to find out roughly one face or one eye but due to the lower resolution of the video/EVF data stream or intermediate image nececcsary for IA is not in position to deliver the lens comand accuracy that we can measure with long lenses at long distance . It looks like phase AF signal ingested in Exspeed is called back after AI ( or any AF zone) has given a bulk area where to focus.

The paths and logic beetween paths and cores inside Exspeed for achieving the expérienced accuracy of the system looks far more complex to understand and .. no way for us to extract more details from Nikon litterature !

PS Thanks for all your very clear videos on other topics.


Claude.T

 

[mosswings]

Thanks for the links.

If we condider only the path from On Sensor AF photosites up to the final focus value transmitted to the lens, for focusing at the right place and distance, a lot of information are really missing to understand what and how AF phase information is used and processed ( with or without IA.

My guess is that IA helps only to find out roughly one face or one eye but due to the lower resolution of the video/EVF data stream or intermediate image nececcsary for IA is not in position to deliver the lens comand accuracy that we can measure with long lenses at long distance . It looks like phase AF signal ingested in Exspeed is called back after AI ( or any AF zone) has given a bulk area where to focus.

The paths and logic beetween paths and cores inside Exspeed for achieving the expérienced accuracy of the system looks far more complex to understand and .. no way for us to extract more details from Nikon litterature !

PS Thanks for all your very clear videos on other topics.

Claude.T

I think you've tumbled to what's going on - you have two separate systems - a targeting system that doesn't have to have pixel-level resolution and a focusing system that provides phase information. The targeting system works on large areas of the frame, whereas the focusing system determines proper focus therefrom by comparing displacement of pairs of lines of contrast in the immediate vicinity of what the targeting system thinks is the subject. As long as the target can be placed sufficiently within the focusing system's view the focusing system will have enough contrast line length to resolve the location of maximum contrast and map this into a focusing command for the lens.
There have been numerous reports of final focus not being precisely what the EXIF data claimed. This isn't exactly an example of what I'm talking about, but it does demonstrate that the relationship between targeting and focusing is not as rigid as we'd like. The target center is not precisely colocated with what the focusing system uses to determine subject distance but it's quite close.
Modern AF systems are very complex. When tracking, the system is also focusing, because any improvement in focus helps the targeting system identify the subject.

There are some other assumptions you initially made that I think aren't warranted - that focusing is done at the same resolution as the EVF displays. I don't think that's the case. Eventually the GPU outputs an XVGA stream to the EVF, but it's been working with the full resolution of the sensor up to that point - and before that point is where all the targeting and focusing is done. The GPU may have to be briefly diverted from its targeting/focusing duties to render the EVF display (if we're not talking about a Z8 or Z9) but it starts with the same data.

 

[jthomas39]

Thanks for the links.

If we condider only the path from On Sensor AF photosites up to the final focus value transmitted to the lens, for focusing at the right place and distance, a lot of information are really missing to understand what and how AF phase information is used and processed ( with or without IA.

My guess is that IA helps only to find out roughly one face or one eye but due to the lower resolution of the video/EVF data stream or intermediate image nececcsary for IA is not in position to deliver the lens comand accuracy that we can measure with long lenses at long distance . It looks like phase AF signal ingested in Exspeed is called back after AI ( or any AF zone) has given a bulk area where to focus.

The paths and logic beetween paths and cores inside Exspeed for achieving the expérienced accuracy of the system looks far more complex to understand and .. no way for us to extract more details from Nikon litterature !

PS Thanks for all your very clear videos on other topics.

Claude.T

Can you explain more on why long telephoto focusing is different than the shallow depth of field in a 35 or 50mm f/1.2 lens? I'm not sure about why it's not a similar PDAF method in both cases?

(And I don't know this term "IA"?)

 

[mosswings]

Thanks for the links.

If we condider only the path from On Sensor AF photosites up to the final focus value transmitted to the lens, for focusing at the right place and distance, a lot of information are really missing to understand what and how AF phase information is used and processed ( with or without IA.

My guess is that IA helps only to find out roughly one face or one eye but due to the lower resolution of the video/EVF data stream or intermediate image nececcsary for IA is not in position to deliver the lens comand accuracy that we can measure with long lenses at long distance . It looks like phase AF signal ingested in Exspeed is called back after AI ( or any AF zone) has given a bulk area where to focus.

The paths and logic beetween paths and cores inside Exspeed for achieving the expérienced accuracy of the system looks far more complex to understand and .. no way for us to extract more details from Nikon litterature !

PS Thanks for all your very clear videos on other topics.

Claude.T

Can you explain more on why long telephoto focusing is different than the shallow depth of field in a 35 or 50mm f/1.2 lens? I'm not sure about why it's not a similar PDAF method in both cases?

(And I don't know this term "IA"?)

I believe the author means “AI”.

 

[Restoc]

Thanks . I agree that SW upper layers of the AF on-board computing couldn't be based on the lower XVGA resolution . This implies that a high or semi high res image is buffered somewhere on the paths at disposal of Exspeed core in charge of AF for scene recognition and that a separate more accurate Phase information comes boack for achieving the final Lens accurate positionning independently of what we see in the EVF and what we get in final Exifs on the files and of Focus Point display in CNX Studio.

The image of " scene Intelligence" ( somewhere on the EVF stream/pipeline) is reconstructed a time after complete capture (on board demosaicking for EVL/LV and IA processus). This could explain why some differeces beetween Phase and SR can be observed.

 

[Restoc]

Different not on math point of view . But with long range we are just more facing permanently that a little out of focus exactely on the eye of a bird destroys immediately the picture . Just a question of amlifying the discrepancies : a small out of focus on the crux of the picture is just amplified and far more visible for the human eye-brain system. And in addition we canot call the birds or animals to sit and come near so we need to crop and improve the visible issues on the files.