Ho John Lee's Weblog

Ojos is the working name for a startup that’s building a photo site with automatic tagging through face and text recognition. Here’s a sample photo of a Treo, annotated with the words their technology can recognize.

From Rob Hof on his blog at Business Week:

Munjal Shah, onetime cofounder of the auction services firm Andale, finally let slip on his new blog what he’s been working on since leaving last year. As he writes: “I am co-founding a company because I found I had 31,246 photos all named DSC0009.jpg.” In other words, his startup, tentatively named Ojos (Spanish for “eyes”), is creating a new way to search and organize photos.

Over at Munjal Shah’s new blog, he elaborates further:

I think Flickr’s tag based system is just super (in fact I love it), but I wanted all of my photos on there, I wanted them all tagged, and I didn’t want to spend hundreds of hours doing it. So being the lazy engineers that we are, we thought maybe we can at least auto-tag some of the faces and names. Folks can fix mistakes we make but it will still be less than than tagging in the first place and in the end you will have a tagged library of photos.

Looks like they’re going to be applying face and/or scene similarity and text recognition to help organize the rapidly growing collections of digital photos being generated through the mass consumer adoption of digital cameras and online photo services such as Flickr.

I don’t think there’s room in the market for another freestanding web site, even one based on better face and text tagging. At the same time, the autotagging capability can’t be tested, demonstrated, or evolved without a live data set and community of users.

This feels like it should become or at least expose a web service at some point down the road. It could then be used with any photo hosting service or web site to reach a wider set of users than just one site. It might also help distribute the computational load of calculating the regions of interest, feature vectors, and resizing, by pushing the task out to the clients in many cases as part of the upload process. Computing resources continue to become cheaper and faster, but there are a few bandwidth bottlenecks along the way, so why not let the desktop chew on it a while and send up the precomputed metadata, along with the (possibly smaller) image.

Auto tagging, combined with a community of users that helps “clean up” the relevancy of the applied tags, might also work well for labeling photos of celebrities and well known places.

I’ve wished for something like this a few times in the past, so I’m hopeful that this team will come up with a useful service and look forward to trying it out when they make something available.

Update 08-28-2005 23:28 – Posted some additional comments at Munjal’s site. Briefly – I think it’s becoming interesting to do image content-based retrieval in conjunction with tagging and other user behavior. I should write up some notes on group search and tagging.