Structural Description Approach to Human Object Recognition

1
Structural Description Approach to Human Object
Recognition

• Brian J. Stankiewicz

2
Structural Descriptions

• Specify features and their relations
• Usually specified with respect to 3D objects.
• E.g., cars, airplanes, flowers, etc.

3
Structural Descriptions

• Biedermans Recognition By Components (RBC)
• Used simple primitives called Geons
• Specified Geons and their inter-relations
• Other Structural Descriptions
• Marr Nishihara
• Dickinson, Pentland, and Rosenfeld

4
Structural Descriptions

• Primitives are set of geometric parts
• Decompose object into simple volumetric parts
• Describe the parts and their inter-relations
• Cone-above
• Brick below

5
Structural Descriptions

• Biedermans Recognition by Components

• Small set of primitives can specify large set of
  objects

6
Structural Descriptions

• Parts are described on a series of non-accidental
  shape attributes
• Cross-section
• Constant
• Expanding
• Expand Contract

7
Structural Descriptions

• But how do we determine the appropriate shape
  attributes from image?
• Use non-accidental imageproperties.
• E.g., Parallel inimage, parallel onobject.

8
Structural Descriptions
Hummel Biederman, 1992
9
Empirical Studies

• Want to understand underlying representation and
  process mediating shape perception
• Develop studies that test explicit assumptions in
  both approaches
• Preview Discuss Biederman Cooper priming
  studies

10
Logic of Priming Studies

• Want to know what information is made explicit
  about an image
• E.g., Store image features or simple volumetric
  parts?
• Do object classification study on images

11
Logic of Priming Studies

• Priming study procedure
• Prime Phase
• Present subject with stimulus (I.e., a picture)
  and have subjects process the stimulus (e.g.,
  identify the object)
• Probe Phase
• After a period of time, either present the
  identical stimulus again, some variant of the
  stimulus (e.g., different contours) or an
  unprimed (not previously seen) stimulus.

12
Logic of Priming Studies
Prime Phase
13
Logic of Priming Studies
Prime Phase
14
Logic of Priming Studies
Prime Phase
15
Logic of Priming Studies
Prime Phase
16
Logic of Priming Studies
Pause
17
Logic of Priming Studies
Probe Phase
18
Logic of Priming Studies
Probe Phase
19
Logic of Priming Studies
Probe Phase
20
Logic of Priming Studies
Probe Phase
21
Logic of Priming Studies
Probe Phase
22
Logic of Priming Studies
Probe Phase
23
Logic of Priming Studies

• Measure response time to stimuli presented in
  probe phase.
• Compare response time of primed images to
  unprimed images.
• Priming RT(Unprimed) - RT(Primed)
• If identical image priming is equal to variant
  stimulus then visual system does not store
  information about variation.

24
Identical Image Identical Shape Identical Concept
RT(ID)
Identical Probe
Different Image Identical Shape Identical Concept
RT(LR)
Variant of Prime (Left-Right Reflected)
Different Image Different Shape Different Concept
RT(Unprimed)
Prime Image
Unprimed Baseline
Different Image Different Shape Identical Concept
Different Exemplar
25
If Orientation is not encoded.
Prime Image
Response Time
Different Exemplar
Variant of Prime (Left-Right Reflected)
Identical Probe
Unprimed Baseline
26
If Orientation is encoded.
Prime Image
Response Time
Different Exemplar
Variant of Prime (Left-Right Reflected)
Identical Probe
Unprimed Baseline
27
Parts versus Image Features
28
Parts Versus Image Features

• One significant difference between View-based
  models (Tarr Pinker) and Structural
  Descriptions (Biederman) is the set of primitives
  used to represent shape.
• View-based uses image features
• Pixels, edges, vertices, etc.
• Structural descriptions use volumetric parts
• E..g, Cones, Cylinders, Bricks, etc.
• Use image features to extract simple volumes
• Image features are not stored in memory.
• Part description stored in memory

29
Parts Versus Image Features
Image Feature Deleted A
Image Feature Deleted B
Same Name Different Exemplar
Intact
30
Parts Versus Image Features
Prime
Identical
Probe Images
Complement
Different Exemplar
31
Parts Versus Image Features
Response time to identical image
Significantly faster thanunprimed recognition
32
Parts Versus Image Features
Response time to image withnon-overlapping image
features.
33
Parts Versus Image Features
Recognition times for identicalimages or
complement imagesshows identical priming
advantage. Suggests that the priming is notin
the image features. Is the advantage in priming
dueto saying the same name twice? No. Objects
that have the samename but are different
exemplarsare slower than identical and
complement. Some of the priming is visual
34
Parts vs. Image Features

• Are the data conclusive?
• Suggests that priming is not in the simple 2D
  features.
• But is the priming in the parts?
• Not conclusive
• How would you go about testing?

35
Parts Versus Image Features

• Second experiment
• Same as first, but complementary images differ in
  the set of parts.

36
Parts Versus Image Features
37
Parts Versus Image Features
Prediction if priming is in Parts?
Prediction if priming is in something other than
parts or features?
Different Exemplar
Unprimed
Identical
Parts Complement
38
Biederman Cooper Exp. 2 Results

• Priming for complementary parts is same as
  control.
• Suggests that priming is in the parts.
• Combined with Exp. 1 suggests that priming is not
  in edges, but in parts.

39
Parts Versus Image Features

• Is the priming in the parts?
• Yes.
• A significant response time advantage for
  identical over different exemplar.
• No advantage for part complement.
• If priming were not in parts we would expect the
  same advantage for identical and part complement

40
Summary

• Priming studies
• Reactivation of previously stored representation
  should be processed faster
• Use logic to see what information is made
  explicit
• Compare identical image to variant response times
  (relative to unprimed, and different exemplar
  controls).

41
Summary

• Used Priming study to determine the use of image
  features versus parts.
• Used complementary images
• Complementary image features showed no difference
  between identical and complement
• Complementary part objects showed a reliable
  difference between identical and complementary
  images.

42
Discussion

• Does this support Biedermans RBC approach?
• Why/why not?
• Does this demonstrate that we dont do
  multiple-views plus alignment or mental rotation?
• Why/why not?