Recognizing People in Photos Through Private On-Device Machine Learning | Hacker News Reader

Recognizing People in Photos Through Private On-Device Machine Learning

machinelearning.apple.com

3 pointspo8 comments

This is capability is actually really important if you take iPhone photos to commemorate family or group activities.

I've got well over 50,000 photos, and I couldn't possibly identify every person in the photos I've taken without face recognition.

Having the software and the processing power available to analyze an large collection of photos on device rather than in the cloud is a game changer.

I can see that it would be useful in a professional capacity where having 50,000 photos was part of the job. But I must ask, what do you do with those photos? Do you really cull through them and share them or use them again? 50,000 of them? It seems like you've made the case for the memory hole I was talking about.

Of course, any software always had edge cases and niche uses. My complaint primarily was that this seems to fall in the bucket of not-terribly-useful-whiz-bang-stuff that Apple (and, by all means, not only Apple) likes to include so they have something to do PR about. Animoji was definitely in this category and every FAANG company is guilty of this to some degree.

I cannot imagine that, if you surveyed 10,000 users, even one of them would list "on-device facial recognition" as a feature that they need or want from their phones.

50,000 photos aren't that hard to amass if you're old. Digital photography was available to enthusiasts in 1998. If you took ~200 photos a month, you'd have 52,800 (200×12×22).

(I've personally got several times that: it's why I wrote PhotoStructure)

Undoubtedly true. I have 10,000 song files with the same problem!

I'm not a professional photographer. I was a digital photography enthusiast from nearly the beginning.

I kept taking photos on an SLR for a while in parallel with digital photography, then I gradually switched over to all digital. Then I stopped taking photos with a purpose built digital camera and started focusing on the photos I can take with my iPhone.

I worked really hard to drive all the digital images I have into iCloud, whether they originated on an iOS device or not. I have a decent collection of photos that came from Kodak PhotoCDs. I could have a lot more photos available in my iClod collection if I invested in a negative scanner.

My biggest use cases for face recognition are for life milestone events for other people. Weddings, graduations, funerals. People who know me often ask me to look for photos of loved ones if we've been at events together.

I could do more than I am doing with the technology. But what I can do without too much effort sometimes surprises me.

This is so not important or even mildly interesting. ML to identify people whom you presumably already know... on your on phone? Why? First of all, neither you nor anyone else will look at most of those pictures ever again, a side effect of us all having phones and our photo albums around all the time. The photos become throwaways and, effectively, go down the memory hole.

If Apple wants to do some on-device ML (which I'm certainly more in favor of than centralized ML which is just a honeypot for abuse), can't they come up with something actually useful?

I’ve found it super useful, personally, and am really glad both that it’s a feature and that it’s not on the cloud.

They have had a similar feature since iPhoto in maybe 2010, but it has only been on the phone since recently.

I reference photos of people I know all the time as well as use photos to track wiring of projects I work on, etc.

The same infrastructure allows searching for all kinds of different objects as well, which is pretty cool.

Especially interesting to me is to search for “art”. It shows things I’ve tagged, as well as what the model calls art which is often actual art (I’m an artist among other things) but sometimes is weird photo angles and interesting color schemes.

I think it has decent taste, even.

I’m pretty sure this blog article is just about one specific feature. I suspect that they use ML for other things in iOS but for clarity may have focused on one thing here.

That's a terrific use case, albeit super niche. I'm an artist, too. It would be fantastic to have ML help in identifying new paintings or other art (or just ones you don't know) and classifying them as belonging to an artist, movement, school, or period.

I learned this stuff the old-fashioned way, through flash cards, slides, and lots of memorization in art history class. People like you and I can tell if something is more Diebenkorn than Duchamp, but that's a specialized skill that ML could probably duplicate quite well and would be very useful for teaching art and art critique.

Reply on news.ycombinator.com