Tips For Improving Face Detection Accuracy

These are some rules I have slowly discovered that have helped improve the accuracy of iPhotos' face detection.

a. Don't force iPhoto to recognize *+every single+* instance of an individuals face. (If you need to create a location containing every instance of an individual, then you need to setup a Smart Folder using the persons' name as a Description or Keyword. "Faces" is not the section to do this. See a suggestion below for using "Faces" to do this.)
b. Avoid confirming really poor quality examples of a persons face (low light, pixelated, off in the distance, etc)
c. Avoid confirming only partially visible faces (hair covering half the face, head turned away to a profile, extreme angles too high or too low, etc)
d. Avoid too many instances in which the person's eyes are closed. (Obviously, sunglasses are a real problem here as well)
e. Avoid instances where the persons head is tilted 45 degrees or more, unless the software recognizes the face on it's own.
f. Avoid instances where the face has areas of heavy shadow, and the light source is at right angles to the face. (Outdoors with Sun directly overhead or near horizon.) If face is in an "even" shadow that seems to be OK.
g. Clearly, the more full frontal, clearly lit views of an individuals' face you can include, the better the face detection will be.
h. If you do use the "Add Missing Face" button, then restrict the selection box to just around the eyebrows, nose and mouth. Avoid selecting the whole head since the face detection software concentrates on the features in the center of the face. No need to include the hair as it does not appear to be a factor.

*Gathering Every Instance of a Persons' Face*
Here is how you could use "Faces" to gather every photo of an individual, yet maintain its' face detection accuracy. Mind you, this may only be practical with relatively small libraries, not libraries with 10's of thousands of photos.

1. First, go ahead and name *+every instance+* of an individuals' face (Even poor quality examples).
2. Then, within the "Faces" corkboard section, open that persons' album and "Select All" of their photos (Do this before the Faces scan is complete or you will also select all of the suggested faces under "So-and_So may also be in the photos below")
3. Next, create "New Album from selection..." and give it the name of the person. (Alternatively, you could "Batch Change..." all the selected photos with a "Description" using the persons' name and then use a "Smart Album" to gather all of the persons' photos. This could be handy for adding future photos of the person automatically to the Smart Album.)
4. Once that is complete, now "Reject" all poor quality, or partially visible examples of the persons' face within Faces. This will improve the software's future accuracy at detecting the persons' face in new photos added to the library.

If, at a future date, you add another large group of photos containing that persons' face, you can repeat steps 1 & 2 and simply drag all the photos to the persons' Album. Or "Batch Change..." the Description again to automatically include the newly added photos to the Smart Album you created earlier.

In the end I have discovered that the Faces section of iPhoto is really more a "tool" for creating Albums, rather than a destination for viewing photos.

Hope this helps. Feel free to add any suggestions you have discovered for improving the face detection accuracy.

Cheers

MacBook Pro 2.4GHz (Early 2008), Mac OS X (10.5.6), MacBook, PowerBook G4, Time Capsule, AppleTV, AXBS, 2 iPhones

Posted on Feb 3, 2009 1:41 PM

Reply
6 replies

Feb 3, 2009 2:48 PM in response to Glenn Carter

--> 4. Once that is complete, now "Reject" all poor quality, or partially visible examples of the persons' face within Faces. This will improve the software's future accuracy at detecting the persons' face in new photos added to the library.

How do you know that is will improve accuracy instead of decreasing it?

I would bet that iPhoto has some way of knowing and keeping track of the face-detection quality of face image. If so, it already knows to pay little attention to the "poor quality" examples. It's hard to believe that removing accurate guidance, even if it is of marginal quality, is the right thing to do.

Feb 3, 2009 3:05 PM in response to SBlackstock

I understand what you are saying. But here is how I look at it. The Face detection process appears to take all the faces you have positively identified as one individual and then averages out all the common attributes. I'm betting it's coming up with some kind of mathematical formula that becomes distinctive for that face. So every photo you "Confirm" is being used and added to this formula. Every poor quality face you "Confirm" merely dilutes the softwares' definition of who this person is.

Obviously, the clearer the faces you give the software to analyze, the more accurate it will be. But the more fuzzy and indistinct the images you confirm, the more generic that individuals formula becomes. In that case the software will see more similarities in the faces of other people that are not your target individual.

In other words, If you line up 10 very good/clear examples of a persons face, the more accurate the suggestions you will get from the software for matches. But the more poor/unclear examples of a persons face you confirm, the more generic that face becomes and the less accurate the suggestions the software provides.

Does that make sense?

Cheers

Mar 10, 2009 3:20 PM in response to Glenn Carter

Glenn Carter wrote:
. . .
h. If you do use the "Add Missing Face" button, then restrict the selection box to just around the eyebrows, nose and mouth. Avoid selecting the whole head since the face detection software concentrates on the features in the center of the face. No need to include the hair as it does not appear to be a factor.
. . .


Apparently you shouldn't worry too much about what you select for the face area when using "Add Missing Face"... From Apple's knowledgebase at http://support.apple.com/kb/HT3427:

Note: Only faces that have been automatically detected in your photo library will help iPhoto match faces. Manually identified faces will not aid iPhoto's face-matching ability.


So, unless they later come out with an update that actually uses that data, any effort you make in selecting especially accurate face regions is going to waste. (Though I'll admit I'm still putting in the effort in the hope that the data will someday be used for something in a future version of iPhoto...)

Mar 15, 2009 8:10 AM in response to dilvishTD

Based on that note from the previous link, I assume if iPhoto tags an unknown face of someone at a bad angle and I don't want it to go into iPhoto's catalog for that person's facial recognition - I could delete the unknown face and then add my own "missing face". That way the person is ID'd or associated with that picture but the poor angle does not go against their future recognition.

Agreed?

This thread has been closed by the system or the community team. You may vote for any posts you find helpful, or search the Community for additional answers.

Tips For Improving Face Detection Accuracy

Welcome to Apple Support Community
A forum where Apple customers help each other with their products. Get started with your Apple Account.