WRT: Writer Response Theory

Photo search service Riya (nee Ojos) is making a splash with two kinds of image recognition service - face recognition and text recognition. Both allow the recognized element to be automatically framed and tagged within the picture, and subsequently the tagging drives various kinds of category navigation and metadata aggregation.

The company has a lot of buzz based on its face recognition feature - which has garnered good reviews so far, and even led to speculation that the alpha webservice will kill competitors in the photo-sharing community such as flickr.

For WRT purposes I’m more interested in the text recognition than the face recognition - of course, there are interesting parallels between the two. Recognition software of all kinds (face, speech, spam, etc.) often relies on training. This is the case with Riya as well - you upload images with faces and text, and it guesses at the correct answer. You then confirm or correct. While some aspects of training are personal to the user (e.g. some voice recognition training doesn’t transfer well) , others improve as data aggregates from multiple users (e.g. there is broad consensus on most spam). A large community of people all pouring training data into a single system could potentially produce very interesting results, especially in areas like the automatic recognition of brand logos (which appear widely in many images and incorporate unusual typography).

Riya’s features don’t obviate the basis of tagging - it can’t tell that a photo is from a vacation, that it is set in a restaurant, or that you rate the wine bottle in the photo “four-star.” However recognition, in addition to being a killer app for family album uploading, can also bring to text-annotated photos something that is already available in del.icio.us - recommended tags as a baseline for tagging a given resource.