Faces in iPhoto

I’ve been playing around with the face recognition feature in the newest version of iPhoto, and I’m pretty impressed with how well it works. And with how well it fails. I went through the process of confirming/teaching iPhoto to recognize me on photos where it had assigned a high probability that a photo was me, but needed my confirmation to be sure.

I know most face recognition systems start with identifying the size and shape of and spatial relationship among eyes, nose, mouth, etc. What I found to be interesting was the false positives the system delivered. When a picture it thought was me wasn’t me, probably 75% of the time it was an immediate family member: one of my parents, one of my brothers or my son. So clearly there is some amount of facial geometry that members of my biological family share. I wonder if other iPhoto users out there have noticed this? Hmm, maybe iPhoto could conceivably be used as a poor-man’s paternity test, or it might lead someone to discover they were adopted…

What was also interesting was that for the remaining false-positive identifications of me in my photo library, most of the rest were confined to four of my closest male friends: Sam Chambers, Scott Derringer, Martin Reinfried and Jason Mendelson. I’m not sure what to make of that. Are we friends because we look like one another in some subtle way that isn’t obvious to me? Or is it just statistically-likely noise since I have relatively more photos of them in my collection than other folks who appear from time to time?

  • Jud Valeski

    I've noticed a few things:
    – recognition of children, who are physically developing rapidly, is pretty darn good. confirming my son as a true baby, to infant, to toddler, to his 6 yr. old self now, is surprisingly good, despite the obvious evolution of his face. from this I deduce that the spacial relationship (as you point out) is actually fairly, relatively, consistent as the body/face grows.
    – using the app over time helps. suggests that duration from one photo to the next (post confirmation) weights the algorithm.
    – adults… it works phenomenally well.
    – works really well across resolutions and blurry/fuzzy shots.

    overall I'm blown away, though I suspect they're just pushing those with very high confidence/probability to the top of the heap. out of one's entire corpus of photos, as end users, we'll be impressed if the accuracy reaches 100%, even though <1% of one's photos are actually represented in the "match" set.