BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Optimal Tag Sets for Automatic Image Annotation - Sean Moran (Edin
 burgh)
DTSTART:20091204T110000Z
DTEND:20091204T120000Z
UID:TALK21772@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:Automatic Image Annotation seeks to assign relevant words (e.g
 . ``jungle''\,\n``boat''\, ``trees'') to images that describe the actual c
 ontent found in the\nimages without intermediate manual labelling. Current
  approaches are largely\nbased on categorization\, and treat the tags inde
 pendently\, so an annotation\n(jungle\,trees) is just as plausible as (jun
 gle\,snow). In this talk I will\nintroduce a new form of the Continuous Re
 levance Model (the BS-CRM) to\ncapture the correlation between keywords an
 d apply a priority beam search\nalgorithm to find a near optimal set of mu
 tually correlated keywords for an\nimage. This novel approach provides a f
 ormal and consistent method for\nfinding an optimal set of tags for an ima
 ge by considering multiple\nhypotheses for the identity of the keyword set
  via the beam search\nalgorithm. Furthermore by limiting the width of the 
 beam\, one is able to\navoid the combinatorial explosion associated with e
 numerating and evaluating\nall possible keyword sets for an image. This ap
 proach also makes the\ncontribution of examining the performance gains for
  the CRM and BS-CRM\nmodels under both Gaussian and Laplacian kernels for 
 the representation of\nthe image feature distributions. Extensive evaluati
 on demonstrates the\neffectiveness of the approach in refining the set of 
 keywords assigned to\nimages.\n
LOCATION:Engineering Department\, CBL Room 438
END:VEVENT
END:VCALENDAR