Jay McGavren's Journal

2008-09-26

Picasa Web Albums facial recognition - edge cases...

Finally! Picasa Web Albums rolled out its facial recognition feature. No equivalent for the desktop Picasa (yet), sadly.

It’s fast, and accurate. Some interesting bits:

-Most of the time, it has no trouble distinguishing between my pre-beard and post-beard face. -It thinks my wife and my-wife-with-sunglasses are two separate people. -It thinks my-wife-with-sunglasses and my-son-with-sunglasses are the same person. :) -Pictures of a rapidly-growing kid (age 3-5) seem to make it uncertain, though its “most likely” match is usually correct. -Based on only one picture of my brother at age 26, it just picked out a picture of him at age 11! (It wasn’t certain, but he was the most-likely match.) -A fake sneer or frown tripped it up (though again, the most-likely match was correct). -A subject losing (or gaining) a considerable amount of weight makes it uncertain. -The lighting on a face has to be REALLY poor to trip it up. -Most-likely match was right even on a face that was ~20% blocked. -Good guess on a face viewed through smoke (from a yellow smoke bomb). -Two different women with similar makeup, facial expression and pose were sorted next to each other. (And the second was mis-identified as the first, probably because she usually wears glasses and wasn’t in this shot.) -The best guess is usually of the same sex, but when it’s tripped up, members of the opposite sex may outrank the actual subject for second- and third-best guesses. -Subjects of the same (or similar) ethnicity tend to get grouped together. -Sticking your tongue out is a good way to trip it up. -Odd angles (like up from the floor) confuse it (badly). Profile-views aren’t as bad, though. -(Obviously) cartoon faces occasionally find their way into the list. -A short-lived goatee only throws it off once; after that the best guess is correct for further goatee pictures. -Bangs partially covering the eyes throw it off.

I know some of these are “duh” items, but this has just made me appreciate what a hard problem facial recognition is. Especially in the area of photos, where you’re dealing with changes in age, weight, hair, expression, poor lighting, sensor noise, and a host of other problems.

Now, how can I sort the photos back on my hard drive, since Google and I have done all this work?

comments powered by Disqus