Tuesday, April 17, 2007

Hope for image searches

Current image searches give only fair results, but researchers are working on the problem. Here's a couple of articles in MIT Technology Review:
  • Better, more accurate image search
  • Software learns to tag photos. The test used Flickr photos to assign up to 15 tags per photos. ALIPR (Automatic Linguistic Indexing of Pictures) uses statistical methods to analyze a photo one pixel at a time and determine what tags apply. While the software is still in development, it assigned at least one correct tag 98% of the time. Not perfect, but pretty cool. I drool in anticipation of an improved version.

No comments: