BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Semantic Texton Forests for Image Categorization and Segmentation.
  - Jamie Shotton\, Microsoft Research Cambridge &amp\; Darwin College
DTSTART:20080723T130000Z
DTEND:20080723T140000Z
UID:TALK12561@talks.cam.ac.uk
CONTACT:David MacKay
DESCRIPTION:In this talk we'll discuss Semantic Texton Forests\, efficient
  and powerful\nnew low-level features proposed recently at CVPR 2008. Thes
 e are ensembles\nof decision trees that act directly on image pixels\, and
  therefore do not\nneed the expensive computation of filter-bank responses
  or local\ndescriptors. They are extremely fast to both train and test\, e
 specially\ncompared with k-means clustering and nearest-neighbor assignmen
 t of feature\ndescriptors\, and the talk will be motivated with a real-tim
 e demo of object\nsegmentation.  The nodes in the trees provide (i) an imp
 licit hierarchical\nclustering into semantic textons\, and (ii) an explici
 t local classification\nestimate. A bag of semantic textons combines a his
 togram of semantic textons\nover an image region with a region prior categ
 ory distribution. The bag of\nsemantic textons is computed over the whole 
 image for categorization\, and\nover local rectangular regions for segment
 ation. Including both histogram\nand region prior allows our segmentation 
 algorithm to exploit both textural\nand semantic context. Our third contri
 bution is an image-level prior for\nsegmentation that emphasizes those cat
 egories that the automatic\ncategorization believes to be present. We eval
 uate on two datasets including\nthe very challenging VOC 2007 segmentation
  dataset. Our results\nsignificantly improve segmentation accuracy\, and m
 ore importantly\ndrastically increase execution speed.\n\n\n
LOCATION:TCM Seminar Room\, Cavendish Laboratory\, Department of Physics
END:VEVENT
END:VCALENDAR