Object Perception

Perception is easy for us, difficult for computers
Ambiguity, assigning surfaces to objects, seperating objects in scene, problems like occlusion, transparency, etc.

STRUCTURALISM -VS- THE GESTALT APPROACH

Structuralism defined objects into separate points called sensations
Similar to newsprint picture; all points, no true lines
Gestalt - Max Wertheimer: "The whole is different (or more) than the sum of its parts."
Apparent motion and illusory contours
In both these cases, the stimulus is not there, but mind 'fills in the details'
Is this experience, expectation, or something inherent in the wiring of the brain?

Gestalt Laws of Perceptual Organization

1. Pragnanz
2. Similarity
3. Good continuation
4. Proximity / Nearness (compare with similarity)
5. Common fate
6. Familiarity / Meaningfullness
7. Figure / Ground segregation: figure tends to be seen in front, with convex shape, symmetry, vertically or horizontally orientented, background is uniform & low in detail

Experiment with Face/Vase figure - Ss asked: "Face or Vase - "Is line straight or tilted?"
- Ss were 3X more accurate if line projected on the "figure" at that time (in the Ss perception (either the face or vase); showed figure gets more detailed processing
Global versus Local Processing - large, low detail -vs- small, high detail
Criticism of Gestalt - explains visual phenomenon ex-post-facto "after-the-fact"; cannot always predict perception, not many experiments done to confirm "Laws"
Which laws predominate?; What perceptual interpretation is simpler than another?
Orientation and context as other "Laws"
Orientation - same shape (Similarity) but grouped by orientation of shape
Perceptual processing: perception is an end product of a sequence of mental operations; these operations may occur quickly & beyond conscious awareness

CONSTRUCTIVIST APPROACH

Helmholtz's (19th cent.) Likelihood Principle: we will perceive the most likely object to have produced a particular pattern of light and dark on the retina
A large # of possible objects/orientations can create the retinal image
Gregory (1973) Hypothesis testing: sensory stimulation provides data for hypotheses concerning the state of the external world
Kulpë (1904) - colored letters / words presented to Ss, told them to report letter position, then asked about the colors = Ss couldn't report accurately; the info was there but was not attended to, a selection process must take place
We also actively scan scenes
Fixations & eye movements - patterns and time spent looking can be recorded
Infants tend to look at eyes and center of face, older babies start to look at whole face (outline and detail)
Attention can have a large effect on what we perceive
Visual Attention: Wurtz, Goldberg, & Robinson (1970's)
Superior colliculus and V1 cells fired more intensely when monkey atttended, due to arousal and initiation of eye movement
Saccade: high velocity eye movements used to orient to a stimulus
Posterior parietal cells' enhancement is independent of eye movement or other behavior of animal in relation to the stimulus

Perceiving Objects in Stages - at least 2 sequences usually recognized
Pre-Attentive Stage: object broken down into elementary features, primitives
Focused Attention Stage: combined and re-integrated

Ann Treisman (1987,1993) Feature Integration Theory (FIT)
Stages

1. Primitives identified; Pre-attentive
2. Combined focused attention
3. 3-D object perceived
4. Compare to memory
5. 5) Match? = object identified

Identifying primitives with Visual Search task
"Pop-out" Boundaries - curvature, tilt, color, line-ends, and movement
closed areas, contrast, brightness
Illusory Conjunction - "free-floating" properties - before attention, primitives (features) are not associated with particular object
Ss errors recorded in brief (1/5 sec) presentations
Conjunction Search - Serial -vs- Parallel Search
In Serial processing, as # of distractors increase, so does the reaction time; found in conjunction searches, requires attention

Biederman (1987) Recognition by Components (RBC)
- primitives are Geons
principle of componential recovery: if geons are still identifiable, object can be recognized - if geon intersections are obscured, you won't be able to identify the object
Plane example: 80% of Ss correct with 3 geons, 90% with 4 geons

COMPUTATIONAL APPROACH

David Marr - Vision (1982) - highly influential book written prior to his death from leukemia at 34 yrs
Interested in computer vision and algorithms to implement in computers/robotics
Mathematical approach to stages of perception (Bottom-Up approach)
Computer needs experience for Top-Down processing (which we have naturally / through interaction with world)
Takes into account shadows and other components of natural scenes
Marr's Stages

1. Raw Primal Sketch
2. 2 1/2-D Sketch
3. 3-D Representation

Image on retina --> identify edges & primitives --> group primitives & process --> perceive 3-D object

The raw sketch happens before perception
True edges -vs- shadows
abrupt, high contrast gradual intensity changes
Marr's Primitives:

1. Blobs (closed loops)
2. Segments of edges
3. Bars (open passages)
4. Terminations (ends of edge segments)