Pattern and object recognition

Object Recognition

• We face more complex problems than pattern recognition with 3D objects in real-world scenes, which leads to the question of which memory representations can support reliable recognition.

Marr's Theory of Object Recognition

• Marr's theory is a complete computational theory that is mostly bottom-up and modular.
• It involves independent processes producing different representations.
• Key notions in Marr's approach include:
  • Generalized cones
  • Key processes in Marr's approach include:
    • Image segmentation and derivation of axes
    • Determining generalized cones
    • Matching with 3D templates in memory
• Stages in Marr's theory of object recognition include:
  • Image segmentation
  • Derivation of major axes
  • Determining generalized cones
  • Matching with 3D templates in memory

Biederman's Recognition by Components Theory

• Biederman's theory involves a library of 36 primitive 3D shapes, known as "geons", which are defined by non-accidental properties such as collinearity and symmetry.
• Structural descriptions are created by combining geons, which can generate an infinite number of objects.
• Key processes in Biederman's theory include:
  • Edge extraction
  • Detection of non-accidental properties
  • Parsing of regions of concavity
  • Determination of components
  • Matching of components to object representation
• Objects are represented by structural models with details of spatial arrangement, size, etc.

Pattern Recognition

• Simple recognition mechanisms, such as Tinbergen's 1951 study, show that aggressive behavior of male stickleback can be triggered by a red belly patch, but a detailed model of a rival fish is ineffective.
• Recognition of complex shape can be reduced to one key feature.
• More sophisticated visual recognition processes require matching complex stimulus configurations with internal representations.
• Pattern recognition of 2D shapes is a simplified problem of object recognition, and is an applied problem motivated by machine vision.
• Main approaches to pattern recognition include:
  • Template theories
  • Feature theories
  • Structural theories
• Template theories involve templates of patterns in memory, which are compared with stimulus patterns, and if they match, the stimulus is recognized.
• Problems with template theories include that matching would fail even with minor differences between stimulus and template.

View-Point Dependent and View-Point Invariant Approaches to Object Recognition

• View-point invariant approaches predict equal ease of recognition at most viewpoints, and represent objects as 3D representations.
• View-point dependent approaches predict easier recognition when objects are seen at familiar orientations, and represent objects as a collection of views.
• Evidence of view-point invariant representations includes:
  • Biederman and Gerhardstein's 1993 study, which showed that participants can name images of familiar objects as quickly as possible, and that presentation of to-be-named objects is positively primed by objects presented in different orientations.
• Evidence of view-point dependent representations includes:
  • Tarr and Bulthoff's 1995 study, and Gauthier and Tarr's 2002 study, which showed that participants familiarized with unfamiliar objects such as "Greebles" in given orientations are slower in identifying them when presented in different orientations.