Seminar: Rich Zemel, "Learning to label complex images" (5/11/09)

Learning to label complex images
Rich Zemel
University of Toronto
http://www.cs.toronto.edu/~zemel/

Image parsing entails the interpretation of visual images, forming  descriptions of the scene depicted in the image.  One step towards  this goal involves assigning to each pixel or region of an image one  of a predefined set of labels (e.g., road, tree, sky).  This image  labeling task is an increasingly popular problem in the machine  learning and machine vision communities.  In this talk I will  describe the most successful approaches to this goal, which involve  including contextual information in the process.  For example, a  local image patch may be ambiguous as to whether it contains a  hippopotamus or a polar bear; however, a nearby patch containing  snow and ice can help resolve this ambiguity.  We have developed  methods for representing and learning such contextual information,  and demonstrated their utility in labeling complex real-world  images.  I will also describe recent approaches that can benefit  from training images with partial or noisy labels.