Opinions of iPhoto 09 facial recognition accuracy?

[andiwm2003]

main problem for me is that faces does not recognize about 20% of the pics with my GF in it. but i want to have all pics with my GF in an album. so i'd rather have a list with more false positives that i can correct manually than a list that is missing 20% of pictures that i will never find (unless i look through my entire library of 8000 pictures).

the pictures iphoto recognizes are 90% correct, though.

the problem is if i want to organize the entire library by faces it can't be that 5% of faces are not detected and 20% of faces of a certain person are not recognized. that defies the pupose.

i hope for an update soon but i don't see how this can be done. because when they update the face detection algorithm then you would have to rescan the library and teach all faces again from scratch since it would be different characteristics that iphoto uses to recognize a person. so i guess we are stuck with this feature as it is.

 

[chrisli75]

Faces worked great for me. It worked 90% of the time.

The only things I hate is that if someone is wearing glasses it does not work and if it's in a dim lit area it does not work.

 

[petvas]

Just look at the pictures. iPhoto works great....

 

[hacksaw-C87]

The more distinctive facial features are the more accurate I've found iphoto to be. My own face is distinctively narrow and the iphoto has been incredibly accurate at my facial recognition. However my best friend has frequently been confused with almost everyone in my album as his features are more indistinctive. I'm sure Apple with update the software used for facial recognition over time- it was never going to work perfectly straight off the bat. Also, the higher the photo quality in your album the better the recognition works. My old 5MP images have rendered relatively poor facial recognition results but my 13MP camera has produced immaculate ones. There are a lot of variables to consider.

 

[JW008]

Another complaint from many people is having to "confirm" photos one by one. You can make the thumbnails smaller on your screen (lower right corner of iPhoto), then drag a box over the photos that you want to confirm. You can hold ALT/OPTION and do the same for those to "not confirm" quickly. Users do not need to confirm one by one, as many people posting here have mentioned having issues about. I wonder if this hint is mentioned by Apple in their tutorial videos online?

I'd figured out the batch confirm, but not the batch not-confirm. I was batch confirming everyone and then clicking on people a second time to not-confirm. That's really helpful, thanks!

 

[urkel]

Okay, after several grueling hours of Face tagging with iPhoto then I can sum up the experience in one sentence:

"It's good enough to be impressed with, but not good enough to be relied on."

Face recognition is one of the most exciting things in photo management, but even at an impressive 60-70% accuracy then you're still stuck with 30-40% missed or mislabeled tags. And in my 20k photo library that means days and days of telling the software "No, Cousin Jeff is the ugliest guy in the family doesn't look anything like me" or "Hey, that's a horse!". Then when you think you finally finished tagging, the refresh search pulls up another 400 pictures to de-label.

Don't get me wrong, I think the feature is great. But it just demands too much input from the end user so it's not going to be a seamless addition to your workflow. (But the results are definitely cool when mixed with those great new slideshows. Why aren't those offered as screensavers?)

 

[r0k]

I have not tried using any face recognition software before iPhoto09. I was fairly impressed with this feature. It helped me organize photos by person rather than by date. I started tagging with my immediate family. It was very time consuming. There were a LOT of false positives, mostly toward the end of the photos I was being asked to confirm. I found that clicking done rather than wading through dozens of false positives resulted in a fresh set where the first 50-100 photos were over 90% right. Once I started hitting less than 80% right, I would click done and start over. If I decided to confirm some new photos, I was not taken to the photos I wanted, I was taken back to the middle of ones I'd confirmed hours ago.

Face recognition does work. Face recognition does have bugs. Face recognition needs an update to incorporate missing features like multiple tagging, 3 way toggle yes/no/notsure. (Yes, I've got some pretty awful scans).

There were some grainy poorly lit scanned photos that I wasn't sure about but I had already clicked once so I had to click again to deselect them. What about a third click to get back to "I'm not sure?" Then there was the lack of ability to tag multiple photos. This just plain blows. I should be able to draw a big lasso around 20 or 30 photos and only have to click once. The same is true for places, I should be able to tag multiple photos in a single album instead of doing them one by one.


I should mention that the photos I was tagging span close to 70 years. I tagged infants with the same name as the teenagers and the adults of the same person. I'm sure this threw off the software. The software should add an "age" setting so that if I tag an infant with the same tag as the adult, the software doesn't have to rely on the file date of a photo taken in 1941 but scanned in 2008.

My goal in getting iLife09 was to build selective albums that featured my wife and kids. iLife09 has the feel of something rushed to market without sufficient testing. But I needed this feature now, not after 6 more months of development so for me Apple released this at the right time. Despite annoyances, the software performed well enough to accomplish the task I set for it. I'm glad I bought it, but I'm also glad I'm dealing with Apple who have been known for frequent software updates.

 

[0098386]

Ok so I'm dissapointed in iPhotos facial recognition. It detects there are faces (sometimes, sometimes rocks and grass have faces ) but it took me tagging about half the photos of my girlfriend before it actually found some by itself... It couldn't detect one of my friends at all.

However I am quite impressed by the geotagging function. I don't have a GPS camera but I've gone through most of my events and put location data in. I like looking at the resulting map too much

 

[InkMaster]

Ok so I'm dissapointed in iPhotos facial recognition. It detects there are faces (sometimes, sometimes rocks and grass have faces ) but it took me tagging about half the photos of my girlfriend before it actually found some by itself... It couldn't detect one of my friends at all.

However I am quite impressed by the geotagging function. I don't have a GPS camera but I've gone through most of my events and put location data in. I like looking at the resulting map too much

Its not broken, its just finding Jesus in all those things

But on a more serious note; for some reason it thinks that every single rim on a car is a face... eh?

 

[Rivix]

Its a hit and miss for me so far. I wish I didn't have to confirm names so much.

 

[adegutis]

iPhoto 09 identifies penguins (Tux maybe?)

I got some interesting and mixed results.

http://www.degutis.com/blog/iphoto09-amazing-face-recognition/

It does seem to get smarter about the name tags, at least until you exit and re-launch it again, at which point it seems to have to learn once again.

Al

 

[ohcomeon]

Keywords

Maybe I'm doing something different to most others on this forum, but I just upgraded to iphoto 09 this morning and all my old keywords are still in place. I can even add more if I want by presseing command-K.

I'm happy with it. Even the face recognition works fairly well (not 100% yet but certainly not as bad as some people are making out)

 

[andiwm2003]

well, after a few nights playing with faces here is my verdict: not good enough to organize my library

if i want to organize my libray of 7500 pics for 20 people then i could do it easily by hand in 15 evening sessions with 99.9% accuracy.

with faces i can do 2 people per evening (training faces and correcting errors and looking for missed pics). so it would take only 10 evenings but with only 70%-80% of pics detected. 20-30% of pics would be overlooked completely.

so that is simply not good enough to organize my library. i'll use faces to help organizing my library but it's not a serious organization tool on its own. maybe after an update.

my 2 cents

 

[kepardue]

I'm kind of surprised at how bad the experiences of Faces are here. While it's not perfect, it generally did a commendable job scanning my library and hitting about 2200 of the 2800 pictures of my now 10 month old son. I had previously gone through and tagged all of the pictures of him using Keywords, so I just used that as my filter to go in and fill in the blanks.

Yes, that's a huge difference between 2200 and 2800, but you've also got to consider that it's impossible to tag pictures where the face is turned partially or completely away from the camera and so on. It sure did make going through those pictures a heck of a lot easier, and even found some that I had missed.

That said, at one point it did think that a desk fan sitting on my desk at work was my son. And I promise, I have a normal looking baby.

 

[Schtumple]

I've found it to be very reliable, it's picked up virtually every picture with a people in, and subsequently worked out who's who, yeah there's been some errors, but that's to be expected, it's not human

 

[KristenM]

I only installed it last night but after playing with it most of the day today, I am very impressed. I wasn't expecting perfection. I'm realistic to know this isn't a NSA piece of software. I know I'm horrible about tagging photos so I figured this had to better than nothing.

Some of the most impressive things I have seen it do
-finding 95% of the pictures of my niece from all angles ranging from the day she was born until today (~8 months old)
-correctly identifying ~80% of pictures of me ranging from birth to now (29 yrs)
-successfully telling the difference between my twin cousins who, while not identical, looked very similar when they were little.
-making me realize that I'm definitely not adopted. I was always convinced I looked like no one in my family, until I just saw all these zoomed in faces. God we all look the same!

And my favorite of all...
-identifying a picture of a scary dog with teeth sticking out that I saw while in Buenos Aires as my advisor. I'll give it 5 stars just for that alone

 

[ChrisA]

i hope for an update soon but i don't see how this can be done. because when they update the face detection algorithm then you would have to rescan the library and teach all faces again from scratch since it would be different characteristics that iphoto uses to recognize a person. so i guess we are stuck with this feature as it is.

The part where you say "i don't see how this can be done" may be true but that does not imply it can't be done.

If iPhoto keeps records of which phots were used for a "training set" then it can use those photos, not features extracted from those photos to train a completely unrelated algorithm. Yes you are correct that a total re-scan would be required. But the rescan has a large advantage of having a large training set of manually inspected images.

The way these facial recognition systems work is in two steps. The first is to asign a set of numbers, derived from mearurements to each images. These numbers are typically ratios like say, the width between eye pupils divided by the width of the head or the hight divided by the width of the head. The more of these the better.

Then if there are N ratios it plots each image as a point on a N-dimensional graph. Then points that are within a "short" distance of each other are considered to be the same person. Trainng the system defines or re-defines the definition of "short".

So you can see there are many places Apple could improves this withou t making it a total re-write. One simple example might be if the routine that computes the width of a face were made more robust so that it could ignore fluffy hair. This would have the effect of adjusting the location of all those points on the graph and pushing some point in and some out of the defintion of "close" and when this happened the software would have to ask the user if it is correct

The new software would be written to never change the tag on any photo that was manually tagged. So you would never have to go through much of the manual work again

The other thing is that after yu have tagged your entire collection, from then on, you only have to new new photos.

 

[dwd3885]

iPhoto facial recognition works great for me. As soon as I tagged myself in a few pictures, the next 20 rows of pictures of me were actually pictures of me. All I had to do was confirm and I'm done. Works incredibly well. I just have so many photos that it takes longer to confirm everything and tag everybody pictured in the photos.

But iPhoto '09 has nearly swayed me into solely using my MBP instead of managing my photos in Windows Live Photo Gallery.