L3 High-level perception

• Everyday perceptual information:

  • Identifying objects/ faces
  • Reading words
  • Understanding speech

  This processing is rapid and efficient.
• What is template theory?


  When we recognise something, we match it up with the closest instance (template) of things stored in our memory. Pattern recognition is based on global similarity match between sensory input and templates stored in memory.
• What is the feature matching theory?

  Pattern recognition is based on identification of features in the visual array. These are fragments or elementary components of a larger pattern. A limited number of features can be used to represent a large number of objects.

  
• Low level visual feature analysis:

  • Hubel and Wiesel conducted single cell recordings in the visual cortex of cats
  • Found that specific cells respond only to certain kinds of stimuli - simple feature detectors are wired in
• Recognition by components:

  • Objects can be described in terms of a small set of geometrical parts named geons - about 24
• What are geons?

  Simple 3D shapes e.g. cylinders, cones, wedges, etc. each have 15 sizes and builds
• How could objects be represented by geons?

  Mental representation of an object consists of an array of constituent geons, along with description of spatial relations among them - attachment relations.
• Biederman - evidence for geons

  • If object recognition critically relies on attachments between geons, then deleting info about attachments should make recognition more difficult
  • Study:
  • brief presentation of images with 65% deletion of contour
  • either deletion of middles of segments, or at verticles
  • middles = 70% correctly identified, vertices = 45%
• What are problems with geon theory?

  • It would be difficult to distinguish between objects with similar geon structure e.g. horse and cow, if the theory were true.
  • This would make it hard to recognise faces
  • Works less well for natural objects e.g. puddle
• How does the Thatcher illusion work?

  We process faces using local features e.g. noses, eyes, etc and their configurations. When faces are inverted, only local features are available - which are not very dissimilar between normal and inverted view.