Lecture 3 - Object Recognition Flashcards
Describe 2D pattern matching – template, feature & structural theories. Describe 3D object recognition – Marr & Nishihara (1978) & Biederman (1987; 1989) Consider viewpoint invariance and viewpoint dependence. Describe models of object recognition – stage model & cascade model. Explain cases of agnosia
Template Theories of 2D Pattern Matching
multiple templates are held by memory
compares stimuli to templates in memory for one with GREATEST OVERLAP until found match
problem of imperfect matches
cant account for the FLEXIBILITY of pattern recognition
comparison would require identical orientation/size/position of template to stimuli
Prototype theories of 2D pattern matching
modification of template matching
- flexible templates
- we possess AVERAGE of each characteristic - but how do we have e.g. average cat figure?
Evidence for prototype theories of 2D pattern matching:
FRANKS AND BANSFORD
- presented objects based on prototypes
- prototype was not shown
- but PS still confident they’ve seen prototype
Feature theories of 2D pattern matching
- pattern consists of a set of features or attributed
- e.g. A = two lines and connecting cross bar
- but surely we need to know relationship
Structural descriptions of 2D pattern matching:
- need to know features AND relationship of features
- how each bisecting line supports the other.
3D object recognition:
- interpret input to visual system as coherent structures
- must be processed to give a description
- determine what is one object (segregate from the background and other objects) to generate structural description and match this too familiar objects in memory
Marr’s computational approach -
explains this 3-D object recognition.
- primal sketch (light intensity, edges, contours and blobs)
- 2 1/2 D sketch (depth, motion, shading, observer-centred)
- 3-D representation (object and relationship, observer-independent)
Marr and Nishihara (1978)
- objects are made up by cylinders
- structural relationships are expressed by hierarchical organisations of cylinders
- each cylinder has axis and way in which others are joined are expressed as cooridnates.
Biderman (1987&1989)
- objects composed of basic shapes - GEONS which are geometrical icons
- viewpoint invariant theory - recognition using:
- 36 different volumetric shapes
- structural relations between parts
properties of GEONs:
- Curvature - points on a curve
- parallel - set of points in a parallel
- Co-termination - edges terminating in a common point
- symmetry
- co-linearity - points on a straight line
Biederman experiment into geons:
- deleted edges at points where easily reinstated or difficult to determine
- stimuli presented for the varying time period (100 to 750msec) with 25/45/65% contour removes
- PS were slow and inaccurate with deleted edges making object non-recognisable - good at recognisable.
Support for Biederman (1987)
- Vogels, Biederman Bar and LORINCZ: found cortical neurosn in monkeys sensitive to GEONS
ADVANTAGES AND DISADVANTAGES OF BIEDERMAN
AD:
- PARSIMONIOUS - a small set of primitive shapes
DIS:
- structure is not always key to recognition (peach/nectarine)
- which geons
- within-category discrimination
- context? no top-down element
Viewpoint dependent theory
- changes in viewpoint reduce the speed and accuracy of object recognition.
- viewpoint dependent more important for within-category discrimination
Vanrie et al 2002
- when are we viewpoint-dependent or view-point indepdennt
viewpoint dependent = complex within catgeory decisions
viewpoint invariant = easy categorical decisions
both in different contexts?