2D to 3D (Marr vs. Biederman)
Why 3D: Otherwise we'd have to store several views per object
Neither theory specifies how how the matching is done
Both explanations have limitatinos and cant explain all shapes and how we discriminate within categorical differences
No conclusive evidence for object centered description: --> Probably we also depend on viewpoint centered description
In within category disinctions a wealth of info would be lost - Two collies would look the same
Context and surface detail are probably important as well i.e. colour helps recognition and distinction
Probably more than a sinlge way how objects are recognized
Biederman
HOW?
36 Geons= Shapes
Determining Invariant= non accidental properties
Segmenting by concavities
Choosing by comparing non accidental properties
Advandage: Geons immediately turned into 3D
EVIDENCE
Relying on conavity to seperate geons
Biedernan: Deleting non-accidental prop. led to problems in recognition
Recognition Priming: But only when viewpoints < 135 degr. and no geons were hidden
Critique
Generally: Even if all geons could be identified we have problems
Problems with certain shapes: "Wheel" from front
Bulthoff and Edelmann: Complex objects in Novel viewpoints problematic even when 3D description possible
Marr
locate primary axis
locate components by
Marrs 3 Assuptions
Prozesses: What vision has to accomplish
Algorithms: How to operationalize processes
Brain: implementation
Marrs 2D
1. Greyscale description
Blurring to eleiminate lighting effects
2. Raw primal sketch: Detection of light intensity changes includeing edges, boundaries, blobs, Bars
=Edge extraction
Gestalt Grouping Principles
(Assigning Tokens)
3. Full Primal Sktech
4. 2,5 D Sketch: Integrate distance vectors from motor texture cues, stereopis
Viewpoint dependent
Evidence
Depth cues are processed seperately
Critique
Ignores Colour
Ignores context
context Importatnt
Marrs 3D
How?
Build a cannoical coordinate frame (wire frame)
Viewpoint independent
Assumption all is cylindrical
1st. Starting point silloutte from viewpoint depeentent object description
He calls the 2D sillouette a contour generator for to the objects (3D) contour
to convert to 3D make assumtions: Closeness, Planarity
2. Locate areas of sharp concavity to divide parts
Concavities are important contour points (lips, forehead, chin)
3. Determine principal Axis for all parts
relation Axis sizes important for distingusihing gorillas from humans
Matching 3D object with catalog
Hierarchical comparison: Bcasic Shape (Biped vs. quadruped), number , organization of limbs, next human or ape
Evidence
Lawson and Humphreys: Rotation no Effect unless Axis forshortened
Warrington and Taylorgical: Neurological Patients problems w. untypical view (critical features hidden)
Humphreys & Riddoch: Similar Patients: Critical features no/ Forshortened Axis yes
Thats why axis is important for 3D
Critique
Does not explain how / if its implemented in the brain
Recognition as pasive
Catalogue not explained
Limited shapes like animals
Planar assumption: Problems with Octagon vs. Cube
Ignores Role of Context facitlitates Recognition
Ignores Role of Colour factilitates Recogntion
Texture not important
Classification
how withing category?
How naming?
How sematic classification?
Differences
Biedermann
locate key features / thus geons (non aacidental properties) from contour
Why?: All geons have certain non-variant properties
curvilinarity
parallelity
Problem with frontal whee
Evidence: (Biederman) Problems if non accidental are disrupted
36 Geons
cylinders
Wedges
spheres
blocks
Marr
Locate Axis from Contour Genrator
Locate Components
By find inward anglex (concavities)
Evidence: Problems when axis wasd obscured (Humphreys & Riddoch)
Only cylinders
Common
Both start with 2D reprensentation
Both match 3D rep
Problems
Not well specified how 2D to 3D
Having trouble within category: How Faces?
Both perceptual classification, not this is an aminal
Both describe between category description
Neurological and Eperimental Evidence: General Problem with novel viewpoints
Conclusion: Recogntion not completely reliant upon object -centered regogn. also viewpoint dependent regog.
Object Recongnition: 5 Stages: Humphreys & Bruce
Early visual Processing: Raw primal to Full Primal
Viewpoint Depenent Object Descriptions: 2,5D
Structural Classification
Semantic classification
Naming
Why is 3D a progress
Template Aproach
One single template cannnot match all variations
Feature Detection
A strucural description would not recog a rotated "L"