Object Perception Theories Flashcards
Wk5 L2 - Wk6 L1 Set 8
Limitations of Gestalt Theories
- qualitative/descriptive: not based on normative systems, provides descriptions of how the perceptual system functions, but does not say WHY or HOW
- is not quantitative
Theory of Relatability
more quanitative
Theory of Non-accidental features
junctions
Theory of Geons
???
The visual system connects spatially separated visible
areas using what two processes:
contour interpolation
surface interpolation
Contour interpolation
if there is a missing edge, the brain must fill it in
Surface interpolation
if there is a part of an object missing due to occlusion, the brain must fill in the surface to perceive a complete object
What are the steps to the contour interpolation process?
1) Begins with the locating of contour junctions
2) Interpolated edges begin and end at these junctions
3) Contour interpolation follows a smoothness constraint, known as contour relatability
Theory of relatability
What extent of curviness or angle is tolerated by the visual system in the process of interpolation
(same object vs different objects)
Too large an angle –> un-relatable
Acceptable angle –> relatable
NOTE: relatability is related to the gestalt idea of “good
continuation”
What are the steps to surface interpolation?
1) The contour interpolation process depends on oriented edges leading into junctions
2) The surface process complements contour interpolation
3) Surface properties “spread” under occlusion within real and interpolated boundaries
4) This process depends crucially on matches of color, lightness, and texture
Whether we form a unit or not depends on…
surface properties AND contour properties
*example: color differences violates relatability, sharp angle violates relatability
Shape perception is invariant to…
Scale invariance (size) Orientation invariance
Template representation of shapes
*any pattern that would fit the template ‘x’
would be recognized as ‘x’.
* template is specific to orientation in
plane, orientation in depth, size, location,
etc.
Problem with template representation
you would need billions of templates for each object, because it would always require the same activation of the same neurons for each template (x-invariant)
example: could not recognize the letter A in different positions, scales, orientations, viewpoints, etc
Volumetric/Structural representation of shapes
- ideal way of representing objects, because it is 3D and thus takes care of the invariances
Examples of volumetric theories
Polyhedra
Superquadratics
Generalized cylinders
Geons
Recognition by Components (RBC) is better known as
Geon Theory
Geon Theory
the visual system represent objects by their geometric icons (geons) and the spatial relationships between them, metric changes are not important (width/height, etc)
- strengths: provides invariance, by combining geons you could create thousands of shapes
- visual system finds non-accidental features, then extracts structural (or geon) description
How many geons are there in the brain?
About 36
What is another word for a geon?
volumetric primitive
Evidence for geon theory
If the non-accidental features (e.g., corners) are available, geons can be recovered and the object can be recognized.
If the same amount of contour is available but the non-accidental features are missing, the geons cannot be recovered and object recognition will be more difficult.
Invariance in geon theory
- Scale (size) invariant
- Position invariant
- Rotation (view-point) invariant
Problems with geon theory
- not good for describing natural/non-manmade objects
- not good for within-category recognition
- ## human representation is view-based, and not invariant
What is another more simple adaptation of a structural or geon theory?
concavities
Examples of View-point Object Representation Theories
- Templates
- Graphs
- Feature sets
Elastic Graph Representation
each node represents a local feature (hypercolumn in v1), edges are connections between the hypercolumns (synaptic)
local features combined with spatial relationships between them considered a graph
Invariance in Elastic Graph
- position invariant
- somewhat orientation invariant
- somewhat scale invariant
- viewpoint robust
Feature-set Representation
- selective neurons are filters sensitive to specific features
- the retinal image gets processed in parallel
- if the right set of signals gets activated, signals an object classifier and we recognize it in different views
Strengths and weaknesses of feature-set representation
strength: very fast recognition
weakness: ignores spatial relationships, so system could be easily fooled to falsely recognize something
