Monday, November 13, 2006

Automagic tagging - not yet

For a while I've been advocating, and enjoying, the benefits of tagging. Whether it's publicly using tools like flickr, zooomr and del.icio.us, or privately on our piccy collection using Picasa or, across a wider range of documents, with the new tagging capabilities built into Vista. Using free-form tags to add metadata to objects - particularly piccys where the volume of images and content can be large - is quicker than using a more rigid taxonomy, but it's still a time consuming process. That's why I'm always on the look out for an automatic mechanism to tag, or as a first step provide assistance for manual tagging, our piccy collection at home.

For a while a service called Riya looked promising. Riya enables you to tag piccy's and then, using some clever vision-recognition algorithms, provides tags automatically for future images based on the analysis of the images you initially tagged. Unfortunately it appears the focus behind Riya has switched away automatic tagging, last week the team behind Riya launched like.com, an online visual-based fashion shopping site. According to Scoble:

Why not keep working on face detection? Because they learned through user testing that they’d never be able to make it good enough. They found that by focusing on visual image searches they can get a much more satisfied user base.

Another online application that aims to do similar things to Riya is Alipr. Provide an image (or a URI to an image) and Alipr suggests appropriate descriptive keywords from a vocabulary of 332 English language words.

A writeup in MITs Technology Review suggests that on a test of 5,411 images on flickr, Alipr provided at least one keyword that matched a manually assigned flickr tag for that image 98% of the time. A quick test using public images from my flickr stream was less successful.

The limited success of Riya and Alipr suggest that, for some time at least, we'll need to wait on improvements in vision recognition algorithms before we get much help in tagging our image memories. In the meantime though, we can sit back with a certain sense of satisfaction that all the manual effort in tagging images has the added bonus of creating some nice large datasets to help improve those algorithms...

...and with that thought, back to the tag-fest ;-)

Tags: | | |

0 Comments:

Post a Comment

<< Home