Posted on April 10, 2006 at 3:48 PM in experiments
Note: The below is an archived entry from Earthling, formerly EarthLink's official blog. The blog itself has been decommissioned and is no longer updated, and comments are trackbacks are no longer accepted.
I recently wrote about Riya.com, a new photo sharing and storage site that has face recognition built-in as a main part of its appeal. I've been fascinated with face recognition since I first learned in college that it might be a specialized and discrete type of image processing in the brain. That blew me away. I don't know what the current state of the field is, but at that time there were many researchers who believed it to be distinct from other types of object recognition. Prosopagnosia is the name for the inability to recognize the identity of people based on faces. According to the Agnosia Cognitive Primer:
Patients can identify facial parts, recognize a face as a face but with no recognition of the person. In severe cases, patients can not recognize their own face.
We know about this in part because there have been patients who have lost their ability to recognize faces after suffering a lesion or loss of activity in a very specific part of the brain. Here's a site devoted to research on face recognition, especially as it relates to computer work.
Knowing just enough to be dangerous, I wanted to put Riya.com's face recognition to the test. Since it's in beta and I didn't want to mess up the experience for others, I asked the good folks at Riya for permission before I started to mess around with it. They agreed to let me have at it, with the caveat that I understand that non-traditional uses of it would probably yield unpredictable results.
Read on to hear about the game I like to call "Does it have a face?"
The general idea was to see what Riya recognizes as face-worthy, and what it thinks is faceless. We already know that it's probably pretty good at picking faces out of photographic snapshots, but other things have faces too. I'm sure that for the programmers that created it, deciding what should come up as a face and what shouldn't was a matter of fine-tuning. I wanted to experiment to find out where some of those limits lie. As an aside, I'm only experimenting with images that I have the rights to, that have been published for public consumption, or that are in the public domain. If you recognize lots of images from my Flickr collection, that's because I know I have the right to use them.
The first image I tried was the classic monkey/bunny/cop trust fall image:
No faces.
Riya didn't find any faces here. But I selected each face area and assigned each a name and an entry in my address book, in the hope that I can train Riya to recognize them. So far, I've uploaded 16 more photos with those "faces" in them and Riya hasn't figured out that there are faces there.
Next, I tried my photo of an elephant statue from inside a store in LA:
Success! Face!
Riya saw the face on the elephant. I added a name and an address book entry (Elephant Voluspa McGee) for this person. If I had any more photos of it I'd upload them to see if it recognized them.
Next up was a photo of Chancellor Palpatine from the Star Wars movies. This is an actual photograph of an actual human being.
No face. (photo from here).
I did get Riya to recognize a photo of Palpatine eventually, but I didn't have the rights to it so I deleted it. In the next installment, if I can get permission to upload lots more pictures of Palpatine, I have another trick to train Riya for. If you have any to donate, please drop me a line.
Finally, I found some pictures of Mount Rushmore. These are faces, right? Big, huge, chiseled faces.
One face (photos from the National Park Service Digital Image Archives).
In the interest of full disclosure, I initially uploaded a different Rushmore photo -- one whose ownership I couldn't verify. It was very similar to the one above, and in that case, Riya recognized the huge Teddy Roosevelt face as a face, but none of the others. I deleted that one from Riya. When I uploaded the image above, Riya picked out Washington and Jefferson as having faces. In fact, it even recognized Jefferson's face specifically, because I had labeled it in a previous photo.
I uploaded a bunch more Mount Rushmore photos.
Two faces! But Riya didn't know who was who.
Recognized TJ, but not the others.
No match on Lincoln here....
or here.
Finally, at the end of my first day's experimenting, I had to know:
Does this guy have a face?
Riya says no. I say yes. I'm hoping I can change Riya's mind over time. More experimentation to come.
Comments
Dave,
This is hilarious! I love it!
We totally have to get monument recognition to work for you. ;)
T
Posted by Tara 'Miss Rogue' Hunt | April 11, 2006 6:52 PM