Does Nikon use face recognition with AF?

Neil-H

Senior Member
Messages
1,407
Reaction score
764
Location
America baby !!!
I have heard this a few times and am no closer to seeing proof one way or another. At the moment I am an owner of two systems, with a D7k and a 70D. I actually really like both systems for many reasons, some overlapping and some very different. The problem is I would prefer to put most/all future investment into a single mount since I can't really use my lenses/ect on both cameras.

As for the two I own, I love my 70D's grip feel, it's more comfortable than the D7k. I also much prefer the LV of the 70D, and I like having an articulating touch screen. On the other hand, the D7k has a much better auto ISO design, which is very useful in my shooting, and it has a few more AF options when adjusting the focus points. I haven't been able to discern much difference in AF tracking ability, although I admit I haven't really "tested" them in a vs shootout.

Im more or less looking to identify what differences tend to be innately linked to the brand so I can better determine which I will focus on, so to speak ;-) . One thing that stood out as a benefit is the face detection that I heard about. They both have it in LV of course, but I shoot VF 98% of the time. The word is that Nikon uses the RGB or metering sensor I guess, similar to how it uses color to track in 3D mode.

Does anybody have any definitive info on this? Is this only on certain bodies, or is this completely untrue?
 
Cameras as old as the D7000 also use it with AF-Area mode "auto" when focusing through the viewfinder

--
My opinions are my own and not those of DPR or its administration. They carry no 'special' value (except to me and Lacie of course)
Actually, only Nikon cameras with a 91,000-pixel RGB metering sensor can detect faces in the scene to focus on them during OVF shooting.

Their cameras with 2,016-pixel RGB sensors can't detect faces, simply b/c the resolution of that secondary sensor is too low. These cameras will, however, put boxes around faces in image review mode or playback b/c at that point it's easy to identify faces using image analysis. In other words, no Nikon APS-C camera will detect faces in OVF shooting.

Can't blame anyone for getting confused over this though - Nikon's websites don't really make this clear!
Im confused then, my D7k's manual says "If a type G or D lens is used, the camera can distinguish human subjects from the background for improved subject detection". This is from the first paragraph on PG 95. Is this RGB somehow able to detect a human body and not the face? It is confusing but I would like to know for sure what is what.
We should probably do an article clarifying this, as it gets confusing with all the products out there. A good rule of thumb is to just remember that you need a relatively high resolution sensor 'seeing' the scene all the time to be able to focus on faces. This is why mirrorless cameras do this almost effortlessly, as they have all the data off the main imaging sensor to do scene/subject analysis in real time. Canon introduced face detection in OVF shooting with the 1D X, and also offer it now in the 7D2 and 5DS/R, and Nikon's have been offering it in their full-frame cameras since the D4 (when they introduced the 91,000-pixel RGB sensor). These cameras will continue to track the faces as well as the subject or camera moves, btw, in AF-C. Nikon DSLRs will do this better than Canon DSLRs, save for the 1D X which almost keeps up. 7D2 tracks faces very well in dual-pixel AF mode (Live View), b/c at this point its using its image sensor. Certain mirrorless will beat both Nikons and Canons at this, but results vary from model to model. This shouldn't be surprising, of course, since the high resolution image sensor theoretically allows for very accurate tracking.

Cheers, Rishi
In the same way that 3d tracking can detect differences in color... Sort of...in the lower resolution exposure sensor cameras. G and D type lenses relay subject to camera distance info - basically mapping the position of the focusing mechanism to a focus plane position through a table. The combination of this with color detection can crudely track a face that's hue-distinct from the background. But the low resolution hampers its performance.

The exposure sensor is not guaranteed to be a regular array in the 2k levels, but if it were, it can only put 2-4 pixels on a focus zone. The 91k sensor can put 50-100. That's a huge improvement in resolution. However, remember that because this is a color sensor, the effective spatial resolution is reduced by about 3x to account for the CFA needed. Thus, The 2k sensor places on average 1 full color sensel in each focus zone, the 91k sensor about 30-35. Even more huge of a difference.
 
Last edited:
Not true according to the folks I spoke with at Nikon and based on my own testing of AF-Area mode "auto". The PDAF, with input from the 2,016 sensor will "recognize" faces and try to prioritize for them. It's not perfect and rudimentary compared to what we see in the latest gen FF bodies.
As far as I understand it, the 2,016-pixel sensors don't have enough resolution to discern faces, though I concede they may have some rudimentary understanding of skin tone to prioritize faces (much like the newest Canon Rebel offerings).

We may just be arguing semantics here - my point was that actual image analysis for facial recognition can't happen with such a low resolution sensor. And our tests seem to confirm this - even with a face in the scene, it's easy to confuse a D7000 to focus on something in front of the face.
That's why I spoke directly with Nikon Reps about it. It's similar to what we see in AF-Area mode "3D". With a flesh tone bias in AF-Area mode "auto". It can say..."hey, that must be a face so lets focus on that area". Again rudimentary compared to the newer bodies with the 91k sensors
Incidentally, who did you speak to at Nikon? You can PM me if you'd like. We've also spoken to Nikon reps about this, and have gotten not-so-straight answers as you suggest you received.
Depends on the scene. There are scenarios were the 2,016 can put enough pixels on a face to recognize it as one, and prioritize AF-Area mode "Auto" for that "face" . You can set up the D7000 (in good conditions for it) and watch AF-Area mode "auto" track ("tend towards" the peoples faces in a scene might be a better way of putting it) faces if they move slow enough.
If you assume the 2,016-pixels are referring to the total number of R+G+B pixels (and it's marketing, so I can't imagine they're referring to 2,016 R + 2,016 G + 2,016 B pixels, else they'd call that a 6,048 pixel sensor), wouldn't that mean an effective resolution of 32 x 21? Or, if you were to assume a Bayer matrix with proper demosiacing (kind of doubt it): 55 x 37? Either way, far too low resolution actually detect faces by looking for facial features (eyes, etc.).

A 91,000-pixel sensor, OTOH, at worst would be ~213 x 142, which, again at worst, would give you this level of pixelation in a scene with a face in it:

Simulation of what Nikon's 91,000-pixel RGB metering sensor might 'see', overlaid with a D750 AF grid. Essentially a 213 x 142 pixel representation of the original scene. Original photo courtesy of Nikon, taken by Cliff Mautner.

Simulation of what Nikon's 91,000-pixel RGB metering sensor might 'see', overlaid with a D750 AF grid. Essentially a 213 x 142 pixel representation of the original scene. Original photo courtesy of Nikon, taken by Cliff Mautner.

So you can see there how Nikon's cameras with 91,000-pixel RGB sensors might easily be able to detect faces. Here's what a 2016-pixel RGB sensor might see, again at worst:

Simulation of what Nikon's 2,016-pixel RGB metering sensor might 'see'. Essentially a 32x21 pixel representation of the original scene. Original photo courtesy of Nikon, taken by Cliff Mautner.
Simulation of what Nikon's 2,016-pixel RGB metering sensor might 'see'. Essentially a 32x21 pixel representation of the original scene. Original photo courtesy of Nikon, taken by Cliff Mautner.

Don't see how that could properly detect faces. Skin tones, maybe, but faces? Doubtful.

Furthermore, if you look at the D7200 product page, it says: "Face-Priority AF available in Live View only"

My point is that I think the original D7000 material was wrong, and they've updated their description. The harder-to-believe story would be that going from the D7000 to the D7200, Nikon decided to get rid of a very valuable feature...

And face-detection is valuable, by the way, just perhaps not in the way some might think. It's most valuable for very fast-paced shooting scenarios with fast, erratic subjects, where one may not even have the time to 'center focus and recompose' (w/ 3D tracking btw, which'll automatically then move over the AF point to stay on your subject). In that case, having the camera automatically pick the largest face and focus on it might be beneficial - when you don't have time to pick the AF point yourself or even focus & recompose. I've often found this to be the case when shooting extremely erratically moving babies, cluttered by other objects. Of course at this point I won't expect perfect focus on the eyes at 35mm f/1.4, so I'll stop down to f/2 to give myself a little DOF.

All this said, we're still waiting to hear back from Nikon what the actual resolution/layout of the sensor is. Because I myself am surprised at how well the 2,016-pixel sensor subject tracks (better than a 7D2, for example), and am skeptical of how it could do so if it were only 'seeing' the scene with an effective 32 x 21 resolution proxy.

-Rishi

P.S. One does wonder if the density of pixels varies across the sensor, though the D3X's 1,005-pixel metering sensor shows no change in density across the surface (and, when they say it's 1,005, they mean only 335 pixels of each color):

Nikon D3X 1,005-pixel RGB metering sensor.

Nikon D3X 1,005-pixel RGB metering sensor.
 
Last edited:
In the same way that 3d tracking can detect differences in color... Sort of...in the lower resolution exposure sensor cameras. G and D type lenses relay subject to camera distance info - basically mapping the position of the focusing mechanism to a focus plane position through a table. The combination of this with color detection can crudely track a face that's hue-distinct from the background. But the low resolution hampers its performance.

The exposure sensor is not guaranteed to be a regular array in the 2k levels, but if it were, it can only put 2-4 pixels on a focus zone. The 91k sensor can put 50-100. That's a huge improvement in resolution. However, remember that because this is a color sensor, the effective spatial resolution is reduced by about 3x to account for the CFA needed. Thus, The 2k sensor places on average 1 full color sensel in each focus zone, the 91k sensor about 30-35. Even more huge of a difference.
See my post further down, but yeah I could buy that it does some sort of hue recognition, but that's not 'face detection', since any skin tone, and anything like that skin tone, might be picked up then.

Interesting about it not being a regular array, although the 1,005-pixel RGB metering sensor in the D3X was a regular array.

And, yes, it's really only 672 pixels of each color on the APS-C offerings from Nikon, and 30,333 pixels of each color from Nikon's current FF offerings... that's of course a huge step up -- see my 'simulated' comparisons in my post further down.

All this said, we're never ceased to be amazed at how good subject tracking is, despite these 'blocky' images the metering sensor 'sees'. The tracking algorithms must be top notch, and are certainly industry-leading.

Some mirrorless offerings are catching up though. Understandably, given how much resolution/data they have to work with.

-Rishi
 
In the same way that 3d tracking can detect differences in color... Sort of...in the lower resolution exposure sensor cameras. G and D type lenses relay subject to camera distance info - basically mapping the position of the focusing mechanism to a focus plane position through a table. The combination of this with color detection can crudely track a face that's hue-distinct from the background. But the low resolution hampers its performance.

The exposure sensor is not guaranteed to be a regular array in the 2k levels, but if it were, it can only put 2-4 pixels on a focus zone. The 91k sensor can put 50-100. That's a huge improvement in resolution. However, remember that because this is a color sensor, the effective spatial resolution is reduced by about 3x to account for the CFA needed. Thus, The 2k sensor places on average 1 full color sensel in each focus zone, the 91k sensor about 30-35. Even more huge of a difference.
See my post further down, but yeah I could buy that it does some sort of hue recognition, but that's not 'face detection', since any skin tone, and anything like that skin tone, might be picked up then.

Interesting about it not being a regular array, although the 1,005-pixel RGB metering sensor in the D3X was a regular array.

And, yes, it's really only 672 pixels of each color on the APS-C offerings from Nikon, and 30,333 pixels of each color from Nikon's current FF offerings... that's of course a huge step up -- see my 'simulated' comparisons in my post further down.

All this said, we're never ceased to be amazed at how good subject tracking is, despite these 'blocky' images the metering sensor 'sees'. The tracking algorithms must be top notch, and are certainly industry-leading.

Some mirrorless offerings are catching up though. Understandably, given how much resolution/data they have to work with.

-Rishi
Right. It's really only hue rexognition, although if the subject is a face that fills most of the frame it'll start behaving better. As your simulation demonstrates, for typical face sizes in many tracking situations the 2k sensors can really only discern color differences that may imply a face, but can't resolve the details of a face, which is what "face detection" should mean, and what is clearly possible with a 91k sensor.
 
Right. It's really only hue rexognition, although if the subject is a face that fills most of the frame it'll start behaving better. As your simulation demonstrates, for typical face sizes in many tracking situations the 2k sensors can really only discern color differences that may imply a face, but can't resolve the details of a face, which is what "face detection" should mean, and what is clearly possible with a 91k sensor.
Yes, and when a face fills most of the frame, that's exactly when it doesn't even matter whether you have 'face detection' or not, b/c it's likely at that point that that face is the closest subject, and here even a 5D Mark III without any image analysis will likely be fine.

And then how would it deal with different skin tones? It seems to me that the 2,016-pixel sensors would have a really, really hard time accounting for the entire set of all skin tones, and then stil differentiating them from other objects.

This is why I consider the D7000 to not really have face detection, and I think it was a misprint... especially when you consider that newer D7000/5000-series go out of their way to specify that Face AF is only available in Live View.

In fact, in a real-world shootout trying to shoot a baby and her mother filling most of the frame with a 35mm equiv. lens, the D7000 really did not reliably focus on either face at all, often getting distracted by a closer shoulder blade or sweater around the baby, for example. Seemed to me to just be automatically selecting the closest subject in most of our tests, but I'd be curious to see others' experiences if they differ vastly.

If, OTOH, I specified the baby or mother's eye as the starting point and used 3D tracking, even that old D7000 did remarkably well. Though it would get confused much, much more easily than a FF Nikon. That low resolution is also why we often see tracking jumping between eyes of a subject on an APS-C Nikon, but not on a FF Nikon - which itself is impressive: the fact that the camera can discern one eye from the other, though this probably has more to do with the fact that the algorithm has a 'search radius' that takes into account what's around the eye (or initial subject) as well.

-Rishi
 
Last edited:
How do you know it can recognize the faces?
Because you can set the camera up on a tripod in AF-Area mode "auto" (AF-C) and watch it actually track faces with the AF array's (focus boxes move over the faces) as people move around the scene. It's pretty obvious that it actually knows what a "face" is and can detect them. It will actually prioritize for "faces" in the scene vs other more contrasty targets. If you had a sample of the camera in your hand, as I do< you could confirm the capability as well.
When you did this test, did you try having subjects well in front of all human faces in the scene, but not so close as to be closer than minimum focus distance? Because when we do this, the D7000-series cameras focus on the closer subject almost every single time.
the fact that AFTER the picture is taken it highlights the faces is not a solid proof.
Agreed
Good, b/c this was a point of confusion for some. The APS-C cameras do put a box around faces in playback, but that's b/c it's easy to use software after the pictures is taken for subject analysis.
Yea I dont have D7000 in hand. its great if your D7000 can recognize faces with viewfinder focusing. My D7200 unfortunately cannot do that even with AF-C "auto". Any complex situation where the face isnt the closest thing to the camera and/or on the center of the scene and its very random if face is selected or not..
That's a shame. Sounds as if the feature may be dumbed-down a bit in your model. I know they took out the "Human Subjects" phrase in the D7200 manual, as it applies to AF-Area mode "auto". You might see if menu a3 has any affect.

I know it still works pretty well on the D7100 (a little better than the D7K due to better resolution of the RGB meter, I think).
Well that doesn't make any sense, seeing as how both Nikon's D7100 and D7200 product pages specifically spell out: "Face-Priority AF available in Live View only". Seems to me more likely that the D7000 product info and manual were simply mistaken than that they were 'dumbed down'... b/c dumbing them down would be, quite frankly, entirely illogical.

-Rishi
 
Last edited:
Not true according to the folks I spoke with at Nikon and based on my own testing of AF-Area mode "auto". The PDAF, with input from the 2,016 sensor will "recognize" faces and try to prioritize for them. It's not perfect and rudimentary compared to what we see in the latest gen FF bodies.
As far as I understand it, the 2,016-pixel sensors don't have enough resolution to discern faces, though I concede they may have some rudimentary understanding of skin tone to prioritize faces (much like the newest Canon Rebel offerings).
Which by definition is a form of facial recognition. It recognizes a face and will try to focus on that. Some of the time :)
We may just be arguing semantics here - my point was that actual image analysis for facial recognition can't happen with such a low resolution sensor. And our tests seem to confirm this - even with a face in the scene, it's easy to confuse a D7000 to focus on something in front of the face.
Yes, it is rudimentary and easy to confuse...but it's still able to recognize faces and focus on them (within the limits the FoV of the focus arrays allow).
That's why I spoke directly with Nikon Reps about it. It's similar to what we see in AF-Area mode "3D". With a flesh tone bias in AF-Area mode "auto". It can say..."hey, that must be a face so lets focus on that area". Again rudimentary compared to the newer bodies with the 91k sensors
Incidentally, who did you speak to at Nikon? You can PM me if you'd like. We've also spoken to Nikon reps about this, and have gotten not-so-straight answers as you suggest you received.
Two of the Nikon USA reps at the 2013 NAB. I honestly do not have their cards now.
Depends on the scene. There are scenarios were the 2,016 can put enough pixels on a face to recognize it as one, and prioritize AF-Area mode "Auto" for that "face" . You can set up the D7000 (in good conditions for it) and watch AF-Area mode "auto" track ("tend towards" the peoples faces in a scene might be a better way of putting it) faces if they move slow enough.
If you assume the 2,016-pixels are referring to the total number of R+G+B pixels (and it's marketing, so I can't imagine they're referring to 2,016 R + 2,016 G + 2,016 B pixels, else they'd call that a 6,048 pixel sensor), wouldn't that mean an effective resolution of 32 x 21? Or, if you were to assume a Bayer matrix with proper demosiacing (kind of doubt it): 55 x 37? Either way, far too low resolution actually detect faces by looking for facial features (eyes, etc.).
I don't think the D7K does facial feature recognition. It's simply capable of recognizing what is likely a face in a scene and prioritizing focus on that.
Here's what a 2016-pixel RGB sensor might see, again at worst:

Simulation of what Nikon's 2,016-pixel RGB metering sensor might 'see'. Essentially a 32x21 pixel representation of the original scene. Original photo courtesy of Nikon, taken by Cliff Mautner.
Simulation of what Nikon's 2,016-pixel RGB metering sensor might 'see'. Essentially a 32x21 pixel representation of the original scene. Original photo courtesy of Nikon, taken by Cliff Mautner.
It's pretty obvious from your image above (the 32x21 pixel representation) that we are likely looking at a face and where we should likely focus if focusing on a face is a priority. The D7K certainly seems capable of doing that especially when combined with subject distance and the large proprietary image database built into the D7K.
Don't see how that could properly detect faces. Skin tones, maybe, but faces? Doubtful.
You honestly say you can't look at the image above and tell that's a face and were to prioritize focus for it?
Furthermore, if you look at the D7200 product page, it says: "Face-Priority AF available in Live View only"
And the manual also says it can distinguish human subjects in AF-Area mode "Auto". In that mode you can actually see it "pick out faces" quite often. I do think that constitutes basic face recognition
My point is that I think the original D7000 material was wrong, and they've updated their description.
I disagree. I think the manual was correct...but the tech wasn't "up to speed" yet so not flaunted. Still worked a lot of the time.
The harder-to-believe story would be that going from the D7000 to the D7200, Nikon decided to get rid of a very valuable feature...
It looks to me the feature is still there. Was in the D7100 and I honestly have not seen anything to suggest AF-Area mode works any differently in the D7200 than it did in the D7K/D7100. I will admit the D7200 manual is a real step backwards though, IMO
And face-detection is valuable, by the way, just perhaps not in the way some might think. It's most valuable for very fast-paced shooting scenarios with fast, erratic subjects, where one may not even have the time to 'center focus and recompose'
Sounds like your describing AF-Area mode "Auto"
In that case, having the camera automatically pick the largest face and focus on it might be beneficial - when you don't have time to pick the AF point yourself or even focus & recompose.
Or what appears to be the closest face in a scene. That's kind of what I end up with AF-Area mode "auto" in the D7K...but it's not all that consistent. A work in progress (made more whole in the D750, D4, etc. )
I've often found this to be the case when shooting extremely erratically moving babies, cluttered by other objects. Of course at this point I won't expect perfect focus on the eyes at 35mm f/1.4, so I'll stop down to f/2 to give myself a little DOF.
Sounds like you're looking for "Eye recognition features"
All this said, we're still waiting to hear back from Nikon what the actual resolution/layout of the sensor is.
Hogan says it's not exactly Bayer....but that's all I've seen from him on it
Because I myself am surprised at how well the 2,016-pixel sensor subject tracks (better than a 7D2, for example), and am skeptical of how it could do so if it were only 'seeing' the scene with an effective 32 x 21 resolution proxy.
It does do quite well....maybe it can indeed recognize a face too :)
--
My opinions are my own and not those of DPR or its administration. They carry no 'special' value (except to me and Lacie of course)
 
How do you know it can recognize the faces?
Because you can set the camera up on a tripod in AF-Area mode "auto" (AF-C) and watch it actually track faces with the AF array's (focus boxes move over the faces) as people move around the scene. It's pretty obvious that it actually knows what a "face" is and can detect them. It will actually prioritize for "faces" in the scene vs other more contrasty targets. If you had a sample of the camera in your hand, as I do< you could confirm the capability as well.
When you did this test,
Not to far back
...did you try having subjects well in front of all human faces in the scene, but not so close as to be closer than minimum focus distance? Because when we do this, the D7000-series cameras focus on the closer subject almost every single time.
Yes, it will almost always end up on closet subject...but if you check what AF-Area mode "Auto" prioritized during focus (using Nikon software) It often selects faces in a scene if they are prominent. That will not always be the selected focus distance though.
the fact that AFTER the picture is taken it highlights the faces is not a solid proof.
Agreed
Good, b/c this was a point of confusion for some. The APS-C cameras do put a box around faces in playback, but that's b/c it's easy to use software after the pictures is taken for subject analysis.
No but you can check using ViewNX2 to see what focus boxes it picked in AF-Area mode "auto". And they will often be ones with a face in the FoV.
Yea I dont have D7000 in hand. its great if your D7000 can recognize faces with viewfinder focusing. My D7200 unfortunately cannot do that even with AF-C "auto". Any complex situation where the face isnt the closest thing to the camera and/or on the center of the scene and its very random if face is selected or not..
That's a shame. Sounds as if the feature may be dumbed-down a bit in your model. I know they took out the "Human Subjects" phrase in the D7200 manual, as it applies to AF-Area mode "auto". You might see if menu a3 has any affect.

I know it still works pretty well on the D7100 (a little better than the D7K due to better resolution of the RGB meter, I think).
Well that doesn't make any sense, seeing as how both Nikon's D7100 and D7200 product pages specifically spell out: "Face-Priority AF available in Live View only".
It also says it can identify "Human Subjects" in AF-Area mode "Auto" . Based on experience with that mode...it seems it recognizes faces and will sometimes prioritize focus on them. Live View appears to take it a step farther and use facial recognition to exclude other subjects (Face-Priority AF). AF-Area mode "Auto" doesn't seem to be setup to exclude all else for a face. AF-Area mode "Auto" does not seem to be designed for Face-Priority AF, but does indeed seem capable of face recognition. May be why it can identify "Human Subjects"
Seems to me more likely that the D7000 product info and manual were simply mistaken than that they were 'dumbed down'... b/c dumbing them down would be, quite frankly, entirely illogical.
I was hinting that his D7200 sample was "dumbed down" as my sample of the D7K seems capable of recognizing a face in AF-Area mode auto. A bit tongue and cheek. I suspect that if we present his D7200 with some faces in AF-Area mode auto...review of the focus box data may show it did recognize a face or two. I really would be surprised if Nikon did indeed change the AF-Area mode "auto" algorithms in the D7200
 
...did you try having subjects well in front of all human faces in the scene, but not so close as to be closer than minimum focus distance? Because when we do this, the D7000-series cameras focus on the closer subject almost every single time.
Yes, it will almost always end up on closet subject...but if you check what AF-Area mode "Auto" prioritized during focus (using Nikon software) It often selects faces in a scene if they are prominent. That will not always be the selected focus distance though.
So you're saying it prioritized the closest subject, not a human face? Please clarify.

Because we have tests of different size faces with objects placed in front of the face (but not closer than MFD) that show the D7000 and APS-C Nikons focusing on the object in front of the face, not the face.

It wouldn't be surprising if it chose the face if it were the closest subject. That's not what's in question though.
Good, b/c this was a point of confusion for some. The APS-C cameras do put a box around faces in playback, but that's b/c it's easy to use software after the pictures is taken for subject analysis.
No but you can check using ViewNX2 to see what focus boxes it picked in AF-Area mode "auto". And they will often be ones with a face in the FoV.
Yes, we know this - we used this very feature to show eye-tracking on the D750 here. But it wasn't actually eye-tracking, btw, it was just tracking whatever we started on, which almost all Nikons are incredibly good at, albeit FF orders of magnitude better than APS-C offerings here.
Well that doesn't make any sense, seeing as how both Nikon's D7100 and D7200 product pages specifically spell out: "Face-Priority AF available in Live View only".
It also says it can identify "Human Subjects" in AF-Area mode "Auto" . Based on experience with that mode...it seems it recognizes faces and will sometimes prioritize focus on them. Live View appears to take it a step farther and use facial recognition to exclude other subjects (Face-Priority AF). AF-Area mode "Auto" doesn't seem to be setup to exclude all else for a face. AF-Area mode "Auto" does not seem to be designed for Face-Priority AF, but does indeed seem capable of face recognition. May be why it can identify "Human Subjects"
Where does it say that? I just performed a search for 'human' on both product pages, and it's nowhere to be found.

Also, let's please leave Live View out of the discussion, to stay on point. Saying that 'AF-area mode 'Auto' doesn't seem to be setup to exclude all else for a face' is just a complete distraction: Nikon FF bodies are set up to *always* focus on faces above everything else, b/c they can detect faces, and it's exactly this that I'm saying APS-C offerings cannot do, b/c they cannot identify faces via subject recognition (looking for facial features to recognize it as a face).
Seems to me more likely that the D7000 product info and manual were simply mistaken than that they were 'dumbed down'... b/c dumbing them down would be, quite frankly, entirely illogical.
I was hinting that his D7200 sample was "dumbed down" as my sample of the D7K seems capable of recognizing a face in AF-Area mode auto. A bit tongue and cheek. I suspect that if we present his D7200 with some faces in AF-Area mode auto...review of the focus box data may show it did recognize a face or two. I really would be surprised if Nikon did indeed change the AF-Area mode "auto" algorithms in the D7200
Detection of faces in review has absolutely nothing to do with whether or not faces are detected in OVF shooting.

-Rishi
 
Seems to me more likely that the D7000 product info and manual were simply mistaken than that they were 'dumbed down'... b/c dumbing them down would be, quite frankly, entirely illogical.
I was hinting that his D7200 sample was "dumbed down" as my sample of the D7K seems capable of recognizing a face in AF-Area mode auto. A bit tongue and cheek. I suspect that if we present his D7200 with some faces in AF-Area mode auto...review of the focus box data may show it did recognize a face or two. I really would be surprised if Nikon did indeed change the AF-Area mode "auto" algorithms in the D7200--
My opinions are my own and not those of DPR or its administration. They carry no 'special' value (except to me and Lacie of course)
I was testing this by putting a couple of faces (my children) on the frame and having objects between myself and the children. Then clicked the focus button and see which AF-points would light up. Like I said, its quite random that it would focus on a face if the face isn't the closest subject. Note that the faces were quite small in my test.

I did not test cases when the human subject is filling most of the frame. Interesting test case could be an upper body shot of naked person with a shoulder in front and face further back but toward camera. ie. person is sideways looking at camera. I would guess that sometimes it can focus on the face but also sometimes picking up the shoulder. the smaller the subject is in the frame, the less it would pick up the face vs shoulder.

Anyway, I think we are more or less on the same page about the technical capabilities of the D7x00 and its more semantics about what "face recognition" means. Also Mako seems to have better luck with his testing than myself (or Rishi?)
 
...did you try having subjects well in front of all human faces in the scene, but not so close as to be closer than minimum focus distance? Because when we do this, the D7000-series cameras focus on the closer subject almost every single time.
Yes, it will almost always end up on closet subject...but if you check what AF-Area mode "Auto" prioritized during focus (using Nikon software) It often selects faces in a scene if they are prominent. That will not always be the selected focus distance though.
So you're saying it prioritized the closest subject, not a human face? Please clarify.
The "what to use as the actual focus solution" algorithm in AF-Area mode "Auto" seems to be complex (some call it hap-hazard). It obviously can recognize a face in a scene and the closest subject. It will often choose/decide to set focus on closest subject (not always) vs a human face...though it was obvious that it recognized a face in the scene. That is very different than the algorithms designed into AF-Area mode "Face-Priority" which is specifically designed to focus first on any faces in the scene.
Because we have tests of different size faces with objects placed in front of the face (but not closer than MFD) that show the D7000 and APS-C Nikons focusing on the object in front of the face, not the face.
Yes, AF-Area mode "Auto" is not designed to exclusively focus on faces. I think that's because the face recognition capabilities in that mode are limited compared to what we see with AF-Area mode "Face-Priority". That does not mean that AF-Area mode "Auto" doesn't use face recognition in it's algorithms/design.
It wouldn't be surprising if it chose the face if it were the closest subject. That's not what's in question though.
I've never said AF-Area mode was setup to focus on faces first (exclusively like AF-Area mode "Face-Priority" ). I've only pointed out that it does indeed seem to have a Face recognition component that can be demonstrated.
Good, b/c this was a point of confusion for some. The APS-C cameras do put a box around faces in playback, but that's b/c it's easy to use software after the pictures is taken for subject analysis.
No but you can check using ViewNX2 to see what focus boxes it picked in AF-Area mode "auto". And they will often be ones with a face in the FoV.
Yes, we know this -
Then you should be able to confirm the D7K does indeed often do a form of face recognition in AF-Area mode "auto"
...we used this very feature to show eye-tracking on the D750 here. But it wasn't actually eye-tracking, btw, it was just tracking whatever we started on, which almost all Nikons are incredibly good at, albeit FF orders of magnitude better than APS-C offerings here.
Well that doesn't make any sense, seeing as how both Nikon's D7100 and D7200 product pages specifically spell out: "Face-Priority AF available in Live View only".
It also says it can identify "Human Subjects" in AF-Area mode "Auto" . Based on experience with that mode...it seems it recognizes faces and will sometimes prioritize focus on them. Live View appears to take it a step farther and use facial recognition to exclude other subjects (Face-Priority AF). AF-Area mode "Auto" doesn't seem to be setup to exclude all else for a face. AF-Area mode "Auto" does not seem to be designed for Face-Priority AF, but does indeed seem capable of face recognition. May be why it can identify "Human Subjects"
Where does it say that?
Human Subjects? Page 95 in the D7K manual and also in the D7100 manual

Also from Nikon:
"The D7000 detects the main subject using all 39 focus points and automatically focuses on it. With a G- or D-type AF NIKKOR lens, the D7000 can distinguish foreground and background using "subject identification" of Scene Recognition System and detects a person's position by recognizing human skin tone for improved subject acquisition"

That would suggest it can indeed recognize a "face" when the human subject is clothed and will affect the focus solution based on the "face" being detected. Not up to the capability of the D750...but certainly a form of facial recognition. Facial recognition facilitated by skin tone and scene recognition...is still simple face recognition
I just performed a search for 'human' on both product pages, and it's nowhere to be found.

Also, let's please leave Live View out of the discussion,
Why? It helps to point out the differences in implementation of Face Recognition used in Af-Area mode "Face-Priority" and Face recognition in AF-Area mode "auto". That can help add clarity to the discussion
...to stay on point.Saying that 'AF-area mode 'Auto' doesn't seem to be setup to exclude all else for a face' is just a complete distraction:
Not really...as it also helps to clarify the form of Face Recognition being implemented.
Nikon FF bodies are set up to *always* focus on faces above everything else, b/c they can detect faces, and it's exactly this that
As you say...maybe we should "let's please leave FF Face Recognition out of the discussion" as we are really talking about how it's implemented in the D7K in this thread.
I'm saying APS-C offerings cannot do, b/c they cannot identify faces via subject recognition (looking for facial features to recognize it as a face).
But it can recognize a face in AF-Area mode Auto by other means and select a focus array to cover it. That is by definition Facial recognition. It's just not using, maybe, facial features to confirm/determine if it's a face or not. Might be though, as in your 2,016 pixel RGB sensor example, it's most certainly obvious we have "eyes" in that round "face looking area". I wouldn't be surprised if that is built into Nikon's image database. It's a pretty big database after all.
Seems to me more likely that the D7000 product info and manual were simply mistaken than that they were 'dumbed down'... b/c dumbing them down would be, quite frankly, entirely illogical.
I was hinting that his D7200 sample was "dumbed down" as my sample of the D7K seems capable of recognizing a face in AF-Area mode auto. A bit tongue and cheek. I suspect that if we present his D7200 with some faces in AF-Area mode auto...review of the focus box data may show it did recognize a face or two. I really would be surprised if Nikon did indeed change the AF-Area mode "auto" algorithms in the D7200
Detection of faces in review has absolutely nothing to do with whether or not faces are detected in OVF shooting.
Why? If the review makes it obvious that faces are being recognized and selected by the camera for attention in AF-Area mode "auto"
 
Last edited:
How do you know it can recognize the faces?
Because you can set the camera up on a tripod in AF-Area mode "auto" (AF-C) and watch it actually track faces with the AF array's (focus boxes move over the faces) as people move around the scene. It's pretty obvious that it actually knows what a "face" is and can detect them. It will actually prioritize for "faces" in the scene vs other more contrasty targets. If you had a sample of the camera in your hand, as I do< you could confirm the capability as well.
When you did this test, did you try having subjects well in front of all human faces in the scene, but not so close as to be closer than minimum focus distance? Because when we do this, the D7000-series cameras focus on the closer subject almost every single time.
I did you one better. I was testing this specifically and I also tested the idea that it only locks onto skin hue. This shot is obviously just a test shot but it fixes a lot of questions. I gave the camera ample time to reacquire focus to the hand, and I made sure her hand was raised before I began focusing. I also made sure the hand was not too close for the lens (I focused on it first to make sure it was within focus range).

The end result, even though my daughters hand was the same skin color as her face, the D7k in 39pt auto mode still locked onto her face. I tried several times in a row and it consistently locked onto her face, ignoring her hand. The shot here was from the first attempt. Of not also is that her face is not considerably large in the frame, meaning it doesn't occupy a large number of RGB pixels.

just FYI the 50mm F1.8G's minimum focus distance is listed as 17.7", the hand was beyond that distance.
just FYI the 50mm F1.8G's minimum focus distance is listed as 17.7", the hand was beyond that distance.
 
Last edited:
I think D7000 and other Nikons with less than 91k pixels on their RGB sensor are smart enough to pick a good focus point often, but I dont think they can actually recognize the faces.

So while D7000 is probably good at picking up the scene I believe D750 / D8x0 / D4 would be even better, especially when it comes to face recognition/portraits :) But you still have to pay the FF-extra $$ for the camera and lenses to get that. I do not know about Canon viewfinder face detection.
The only Canons with OVF face detection (also in Auto-AF area mode) are: 1D X, 7D Mark II, 5DS, 5DS R.

And, yes, the D800, D800E, D810, D750, D4, D4S are all quite good at picking out faces in the scene when in Auto Area AF. We have not found Nikon APS-C cameras to be anywhere near as good, so we're surprised by Neil-H's result below. We'll have to re-test this more thoroughly, so thanks actually for starting this thread, Neil-H.

-Rishi
 
I think D7000 and other Nikons with less than 91k pixels on their RGB sensor are smart enough to pick a good focus point often, but I dont think they can actually recognize the faces.

So while D7000 is probably good at picking up the scene I believe D750 / D8x0 / D4 would be even better, especially when it comes to face recognition/portraits :) But you still have to pay the FF-extra $$ for the camera and lenses to get that. I do not know about Canon viewfinder face detection.
The only Canons with OVF face detection (also in Auto-AF area mode) are: 1D X, 7D Mark II, 5DS, 5DS R.

And, yes, the D800, D800E, D810, D750, D4, D4S are all quite good at picking out faces in the scene when in Auto Area AF. We have not found Nikon APS-C cameras to be anywhere near as good, so we're surprised by Neil-H's result below. We'll have to re-test this more thoroughly, so thanks actually for starting this thread, Neil-H.

-Rishi
My pleasure, I appreciate the time to investigate it further. Face detection was one of IMO the better ML features that I have used, so it's something that could be very useful on a DSLR too. Unless the camera locked to the face for some other reason it seems pretty good at finding the faces. I will also continue to randomly check when I find this kind of scene available. With that said it's cool that they can even do this with the RGB sensors and really hints at what is gonna be possible in the future too.
 
I did you one better. I was testing this specifically and I also tested the idea that it only locks onto skin hue. This shot is obviously just a test shot but it fixes a lot of questions. I gave the camera ample time to reacquire focus to the hand, and I made sure her hand was raised before I began focusing. I also made sure the hand was not too close for the lens (I focused on it first to make sure it was within focus range).

The end result, even though my daughters hand was the same skin color as her face, the D7k in 39pt auto mode still locked onto her face. I tried several times in a row and it consistently locked onto her face, ignoring her hand. The shot here was from the first attempt. Of not also is that her face is not considerably large in the frame, meaning it doesn't occupy a large number of RGB pixels.

just FYI the 50mm F1.8G's minimum focus distance is listed as 17.7", the hand was beyond that distance.
just FYI the 50mm F1.8G's minimum focus distance is listed as 17.7", the hand was beyond that distance.
This is now very interesting, as we did a similar test - even with a non-skin-toned object in the front, where the D7000 did *not* focus on the face.

We'll have to dig further. Unfortunately, we don't have a D7000 anymore, but do have the D7200, which is probably of more interest anyway...

Thanks for sharing this result,

Rishi
 
Human Subjects? Page 95 in the D7K manual and also in the D7100 manual

Also from Nikon:
"The D7000 detects the main subject using all 39 focus points and automatically focuses on it. With a G- or D-type AF NIKKOR lens, the D7000 can distinguish foreground and background using "subject identification" of Scene Recognition System and detects a person's position by recognizing human skin tone for improved subject acquisition"

That would suggest it can indeed recognize a "face" when the human subject is clothed and will affect the focus solution based on the "face" being detected. Not up to the capability of the D750...but certainly a form of facial recognition. Facial recognition facilitated by skin tone and scene recognition...is still simple face recognition

...

But it can recognize a face in AF-Area mode Auto by other means and select a focus array to cover it. That is by definition Facial recognition. It's just not using, maybe, facial features to confirm/determine if it's a face or not. Might be though, as in your 2,016 pixel RGB sensor example, it's most certainly obvious we have "eyes" in that round "face looking area". I wouldn't be surprised if that is built into Nikon's image database. It's a pretty big database after all.
If you search through the entire D7000 manual, there is not one mention of Face-Priority AF, or face recognition, in OVF AF. Which I believe was the original question?

So it seems we're just arguing semantics at this point... my original point still stands: the D7000 product info page said 'Face Priority AF', but then the D7100 and D7200 product pages changed that to 'Face Priority AF in Live View only'.

That, to me, indicates a misprint in the original D7000 product info. I can believe that it prioritizes skin tones - the 'human subjects' (thanks for pointing that out) - like Canon's newest Rebels now do, but that's skin-tone detection, not face-detection.

What's interesting, now, is Neil-H's recent test that shows focus on the face instead of the hand in front of the face. This means we need to go back and more thoroughly investigate this.
Detection of faces in review has absolutely nothing to do with whether or not faces are detected in OVF shooting.
Why? If the review makes it obvious that faces are being recognized and selected by the camera for attention in AF-Area mode "auto"
Because image review uses a completely different, post-shot image processing algorithm that has high-resolution data - not available to the camera at the time of capture. That's why I specifically mentioned this - it's a misconception to think that just b/c a face is boxed in image review, that the camera has the capability of recognizing this face at the time of capture.

-Rishi
 
We'll have to dig further. Unfortunately, we don't have a D7000 anymore, but do have the D7200, which is probably of more interest anyway...

Thanks for sharing this result,

Rishi
Glad to share Rishi. Here is another set I just took a few moments ago. All of them were nearly in succession, I would aim and half press the shutter to lock, then take the shot. In between these 7 shots there were a couple of times it did lock to the hand, but that was when I drifted too far to the left, when I kept the face more in the center it was consistent.

One thing that could be happening is the camera would rather lock onto whatever the 9 cross type sensels see in the center, as the hand is only covered with the non cross types. Or, the camera is indeed consistently seeing the face. When I get too far away it won't lock to the face (im guessing it's too small for that few of pixels to identicy), but when the face comprises a good portion of the center 21 cluster, it seemed to do well.

I don't know how practical that is as somebody could just using the 21pt setting but it's still interesting.



One of my usual 3 test subjects.
One of my usual 3 test subjects.



702d436879c14a06943c44ce118ce357.jpg



ea8c18a385064b5da323997502e3545b.jpg



1985c24958a54d8c99f79e44edcee9d4.jpg



0ff6c8c9810b47239c233efde9d42205.jpg



7deca93269f34cebafb596c2e023e7ee.jpg



eef56331265c4085ba9f1e465cc9586c.jpg
 
Thanks again, this is very interesting. Our tests did have non-central faces, maybe that does have something to do with it. Certainly warrants a thorough re-investigation. If it's just limited to faces in the central area, that's of limited value though, since you generally want to use face-detect AF when you *don't* have enough time to pick the focus point yourself (which generally means your subject position in the frame is erratic).

Did you perchance try this with things reversed - face on the left, hand in the center?

Thanks, Rishi
 
Detection of faces in review has absolutely nothing to do with whether or not faces are detected in OVF shooting.
Why? If the review makes it obvious that faces are being recognized and selected by the camera for attention in AF-Area mode "auto"
Because image review uses a completely different, post-shot image processing algorithm that has high-resolution data - not available to the camera at the time of capture. That's why I specifically mentioned this - it's a misconception to think that just b/c a face is boxed in image review, that the camera has the capability of recognizing this face at the time of capture.
I've been reading this thread with some interest from the sidelines. I believe there may be a slight miscomunication in the above.

If I'm not mistaken, when Mako2011 said "review of the focus box data may show it did recognize a face or two", he was referring to verifying the selected focus point in PP, using e.g. ViewNX2's "Show focus point". Perhaps the use of the word "review" may have induced some confusion.

Cf. what he had said earlier in the same post: "No but you can check using ViewNX2 to see what focus boxes it picked in AF-Area mode «auto». And they will often be ones with a face in the FoV."

Anyway, just trying to help (and genuinely curious :-) )
 

Keyboard shortcuts

Back
Top