[bwDraco]

Inspired by "https://www.pentaxforums.com/forums/172-pentax-k-3/271999-k-3-multisegment-me...-smart-me.html".

From my own testing, I can confirm face detection exists when shooting through the viewfinder. If the autofocus system is set to Auto 27 or Auto 9, and the Real-Time Scene Analysis System detects faces in the frame that are within the AF area, it will use the focus point closest to the subject's eye. The camera will also adjust the exposure accordingly. The detected face(s) is reported in the Exif data, and up to 5 faces can be detected in viewfinder shooting (checked using ExifTool).

Questions, thoughts, or comments?

—DragonLord

---

Last edited by bwDraco; 07-10-2015 at 10:50 AM.

[rawr]

I always knew the K-3 PDAF recorded the presence of faces in the EXIF, but I was unaware how it worked to actually refine the AF. Thank you for doing this investigation.

[Larrymc]

Inspired by "https://www.pentaxforums.com/forums/172-pentax-k-3/271999-k-3-multisegment-me...-smart-me.html".

From my own testing, I can confirm face detection exists when shooting through the viewfinder. If the autofocus system is set to Auto 27 or Auto 9, and the Real-Time Scene Analysis System detects faces in the frame that are within the AF area, it will use the focus point closest to the subject's eye. The camera will also adjust the exposure accordingly. The detected face(s) is reported in the Exif data, and up to 5 faces can be detected in viewfinder shooting.

Questions, thoughts, or comments?

—DragonLord

Good to know information........many thanks for the research!

[Anthen]

Awesome to know

[JinDesu]

Whoah, so it'll consistently choose the focus point nearest the eye? So not only is it face detection, but also facial feature detection.

[aleonx3]

What if the DOF is not enough to cover all the faces detected? Will it put more weight to the face closest to the camera in the focusing algorithm? And I suppose as you indicated, it will work only if you put the camera in multisegment metering, not center-weigh metering.

By the way, awesome find, thank you!!

[wissink]

This is with phase detect? Really? Holy smokes.

[danny09]

Thanks but any proof to back your claims?

[bwDraco]

Did some more testing. The camera needs enough light to reliably detect faces, and lens speed plays a major role. I could not get reliable face detection using my DA 18-135mm, but I got consistent results with my DA 50mm. In Auto 27, the camera is willing to choose one of the far left or right AF points of the array if it is closest to an eye, something that does not occur during non-portrait shooting unless there is nothing in the main AF grid. These tests are based on a portrait of myself, so I don't have images to show, but you should be able to replicate them yourself fairly easily. I have not been able to test multiple faces.

Keep in mind that you'll want to use the Auto 27 mode to make the most of this feature.

—DragonLord

[IchabodCrane]

Did some more testing. The camera needs enough light to reliably detect faces, and lens speed plays a major role. I could not get reliable face detection using my DA 18-135mm, but I got consistent results with my DA 50mm. In Auto 27, the camera is willing to choose one of the far left or right AF points of the array if it is closest to an eye, something that does not occur during non-portrait shooting unless there is nothing in the main AF grid. These tests are based on a portrait of myself, so I don't have images to show, but you should be able to replicate them yourself fairly easily. I have not been able to test multiple faces.

Keep in mind that you'll want to use the Auto 27 mode to make the most of this feature.

—DragonLord

Thanks for the info. Did you test what happens if you move the camera after it acquired initial lock on the eye?

[stevebrot]

Thanks but any proof to back your claims?

There are a number of tags and two families of subtags in the makernotes section of the exif that the authors of the ExifTool utility have associated with face detect for Pentax cameras*. They are not unique to the K-3 and position in the exif is not consistent between models. It is not clear which are applicable to current model dSLR cameras or whether their current function is actually related to face detection. The claims regarding face detection are based on the values of one or more of those tags. Here is a sampling:

• FaceInfo/FacesDetected (int)
• FaceInfo/FacePosition (int, int)
• AFPointSelected/Face Detect AF (int)
• FaceDetect (int, int)
• FaceDetectFrameSize (int, int)
• FacePos/*
• FaceSize/*

Note that the values above are all numbers and not text. Any text is supplied by ExifTool. A careful reading of the thread referenced in the OP is useful.

https://www.pentaxforums.com/forums/172-pentax-k-3/271999-k-3-multisegment-me...-smart-me.html

I took part in that thread, but at this point am unconvinced that the camera is detecting faces in any consistent or intelligent manner. I am also high dubious that the camera will isolate the eye as focus point.


Steve

* Note that the tag names are essentially arbitrary and assigned according to context. Other exif readers may label the tags as something else.

ExifTool Tag Names
ExitTool Pentax Tag Names

[bwDraco]

In AF.C, the camera tries to track the eye. On AF.S and AF.A, the focus is locked and does not track. My test target isn't the best, but it does seem to work as expected. As I said, it's very lighting-dependent, and point selection needs to be on Auto for this to work.

—DragonLord

---------- Post added 07-10-15 at 02:13 PM ----------

This is with phase detect? Really? Holy smokes.

According to Ricoh, this is how the RTSAS works. The 86,000-pixel RGB metering sensor introduced on the K-3 is what makes this possible (source):



Face detection is an undocumented feature, but I can confirm it exists on my K-3 II.

—DragonLord

[rawr]

My test target isn't the best, but it does seem to work as expected.

Does the face detection in your tests need to work in colour, or does it handle black and white? The black and white vs colour question relates to whether the RGB sensor uses skin tone colour detection as one input into face detection, or just relies purely on face shape/geometry detection algorithms.

[danny09]

Great explanation. But again please show us at least a dozen exif-intact pictures that support your claims. As Steve mentioned above, the data bytes reserved in EXIF could also be misrepresented by exif reader softwares. So please help us understand your theory by putting up examples.

[clockworkrat]

Show us your face!