Pcb - Vision Lecture 2

Front 1

What is perceptual organization? What does it help us with?

Back 1

Grouping

The organization of perceptual elements into larger units - Gestalt principles

We constantly see fragmented objects but are still able to extract objects from the background & integrate them into something that is useful

- Helps us navigate the environment

Front 2

What are gestalt principles? And what are they?

Back 2

Grouping rules

- Connectedness, common regions, trade-offs between different rules

1. Proximity

- Changing position of identical objects can cause automatic perceptual grouping

2. Similarity

- Assume a smoothness constraint

- Group similar objects

e.g. Colour, orientation, size

3. Common fate

e.g. 2 objects moving in same direction

4. Symmetry, parallelism

5. Continuity / Closure

Front 3

How is competition between different gestalt principles controlled?

Back 3

One option will win

Usually modulated by top-down control. Our experiences shape how we group objects (e.g. grouping of facial features into a face)

Front 4

Who studied figure-ground segmentation?

Back 4

E. Rubin, 1921

Front 5

What is figure-ground segmentation?

Back 5

Figure = Object

Ground = Surrounding area (e.g. background)

We complete surroundedness as a cue

e.g. Rubin's reversible face-vase figure

It is not just the sum of visual input but segmentation of the scene

Front 6

What could properties of figure vs ground be?

Back 6

Figure

1. Thing-like

2. Close to observer

3. Bounded by contour

4. Shape is defined by contour

Ground

1. Not thing-like

2. Further from observer

3. Extends behind contour

4. No shape at contour

Front 7

What could cues of figure ground organization be?

Back 7

1. Complete surroundedness

2. Size (figure tends to be smaller than background)

3. Orientation preference

4. Contrast

5. Symmetry

6. Convexity (curvature)

7. Parallelism

Front 8

What is visual (amodal) completion?

Back 8

When visual system completes occluded objects

- Usually automatically

Front 9

What are factors influencing amodal completion?

Back 9

1. Occlusion

2. Figural familiarity

3. Figural simplicity / "minimum principle"

4. Familiarity

Constrains the possible solutions to find the most plausible due to usually familiarity

Front 10

What is example of modal completion?

Back 10

Illusory contours

Front 11

What is the involvement of V1?

Back 11

Helps figure/ground segmentation

- via Horizontal connections, excitatory pyramidal cells & inhibitory interneurons

- Visual input is filtered through the receptive fields to the excitatory cells

- If mismatch of oriented elements, would expect this to activate a confused representation due to different orientation preferences

- BUT, it highlights edges as transition from one object to another occurs

Front 12

How would you simulate connectedness across neurons?

What does this result in?

Back 12

1. Self-excitation (sustaining the "vote" for a specific percept)

2. Inhibition (suppress other neuron's "vote")

3. Local excitation (supports the neighbours, assuming that they like similar regions, they will choose the same "vote")

Results in some areas being more active than others.

Tells the visual system there is something "special" about the changes in orientation.

This is the seed of figure-ground segmentation

Front 13

Who tested edge detection & region filling in V1? How?

Back 13

Self et al., 2013

1. Monkeys look to a particular region where stimulus is displayed.

It is positioned so that the receptive field "sees" the center of the stimulus, edge of the stimulus or only the background

2. Record different layers of V1 (input, lateral connections, feedback connections etc.) & measure the timing of responses after stimulus onset

Front 14

What did Self et al., 2013 find in edge detection & region filling in V1? What does this suggest?

Back 14

Horizontal connections implement edge detection

(occured ~70ms after onset, with stronger activity in upper layer 4 & superficial areas)

Feedback connections implement region filling

(~100ms after onset, input arrives in layers 1, 2 & 5)

Suggests that some kind of object processing is already occurring at level of V1

Front 15

What else could V1 be used for? Who studied this?

Back 15

"Tracing" a visual route

Roelfsema

Front 16

How was tracing of a visual route by V1 studied? What were the results?

Back 16

1. Monkeys shown 2 objects with 2 curves to trace

2. Tell monkey to trace the first curve

3. Put a stimulus in a position such that the receptive field falls on the curve

4. Then tell monkey to trace the other curve, but leave receptive fields on 1st curve

Results:

1. Subtle, but measurable changes in how V1 will represent attended vs unattended object

2. Input is quickly transformed to take into account other factors

Front 17

How does V1 & V2 respond to illusory contours? How was this tested? What does this suggest?

Back 17

1. Stimulus moved across receptive field

2. Identical responses in V2 found for similar viewpoints, even if box covering stimulus had gaps on either side

3. Differences outside of the receptive field affect the response of the neurones. (e.g. if only half of the box around stimulus)

4. 40% of V2 neurons responded to illusory contours

5. End-stopped V1 cells induce "filling in" of missing contours

Suggests that "filling-in" gives rise to proto-objects

Front 18

What is the hypothesis about V1 & V2 illusory contours?

Back 18

There are neurons in V1 that are tuned to corners/ends of objects (end-stopped neurons)

Perhaps the integration of information across multiple neurons from V1 results in a "filling in" process in V2

Front 19

What is role of V4 in illusory contours? Who studied this and how?

Back 19

Cox et al., 2013

Test:

Activity of neurons with receptive fields on "blanks" of an illusory square surrounded by "pacman circles" or receptive fields on the "inducers" was analyzed

- Control was a flipped pacman stimuli (loss of illusory square)

Results:

1. If the receptive field was centered on the illusory surface, there was a greater V4 response than if there was no illusory square. (Respond to the illusion of a square)

2. Stronger representation of illusory square than when receptive field focused on the inducers. (Stronger for illusory stimulus than physical stimulus)

Front 20

What are potential problems of the V4 illusory contour study?

Back 20

Eye movements (attention drawn to physical stimuli while fixating on center) could result in higher activity

Front 21

What is the different between Luminance and Lightness?

Back 21

Luminance = Amount of light that reaches the eye

Is a physical quantity

Lightness = Perceived reflectance

is a psychological quantity

Perception is often independent of external illumination

Front 22

What is lightness constancy?

Back 22

1. Physical amount of light (luminance) doesn't necessarily determine what you see

- Percept is determined by relative reflectance

Front 23

What is simultaneous contrast?

Back 23

Physically identical strips of light look different based on its context/surroundings.

- Perceived lightness varies with background luminance

- For example, despite same reflectance, the same brightness of an object can be perceived differently depending on the background luminance

Front 24

What does the visual system care more about? Local contrast or absolute luminance?

Back 24

Local contrast

Front 25

What is contrast gain control?

Back 25

1. Contrast transfer characteristics adapt to the mean contrast in the input

- Shift dynamic range of representation based on illuminance

- i.e. bring down responses in a very bright environment in order to still get informative responses (avoid saturation)

2. 10^10-fold range of luminance levels is observed in natural scenes

3. Therefore it's better to respond more to local contrast than absolute luminance

Front 26

What is shunting inhibition? How may it occur and what may it be useful for?

Back 26

Silent inhibition as a mechanism for contrast gain control

Probably occurs via inhibitory interneurons and the conductance variation of inhibitory channels

Inhibitory neurons are in:

Perigeniculate nucleus (PGN) are feedback inhibited

Lateral geniculate nucleus (LGN) are feedforward inhibited by Retina

Front 27

What is the normalization model?

Back 27

Conductance of interneurons in modulated by many other cortical cells and firing rate is modulated by rectification?

Front 28

What is relationship between mean firing rate of PGN/LGN vs contrast?

Back 28

As contrast increases, mean firing rate increases & plateaus at around 3 contrast units

Front 29

How does context affect perception of things? What was the experiment?

Back 29

Even with similar inputs, context can change our perception

Edge type influences illumination match

(one target square affected by a shadow, and a nearby 'matching array' of 16 gray levels to match the gray-level of the target square)

e.g. if there are 2 contexts with matched local illuminance.

Border dividing 2 fields of illumination was made to appear either as a reflectance edge or an illumination edge by either concealing or revealing the larger context (e.g. showing a shadow producer or not)

1. Illumination edge

- Difference seems to arise from illumination

- Responses close to ratio match

2. Reflectance edge

- Difference looks like it arises from different material/colours

- Response is closer to luminance match

Thus, edge type influences lightness matching

- Luminance difference is ignored for illumination edge

Front 30

What else influences lumination perception? Who studied this?

Back 30

3D Structures - Adelson

3D cube with checkered pattern. Whether you look at edge with change of grey-scale or face of the cube with similar change affects your perception.

Local form cues modify transformation of luminances into perceived lightness

-> Features modulate perception of local brightness

Front 31

What 2 things is luminance perception explained by?

Back 31

1. Early visual circuitry

2. "inferential control" by higher-level knowledge about context

Front 32

What is size constancy? Why may this occur?

Back 32

The same perceived size despite different retinal image sizes

- Because our environment is 3D

- Must infer the 3D image from the 2D image on the retina & use depth cues to interpret relative sizes

Front 33

Who measured perceived size and how? What was the result?

Back 33

Holway & Boring, 1941

1. Subject sat with 2 walls (1 in front, 1 to the side, same visual angles)

2. Circles drawn on the walls of different sizes

3. Depth cues were either present (illumination of walls) or not

Results:

- When depth cues are present, correct size is inferred.

- When not, visual system relies on visual angle only and thus both stimuli perceived as same size irregardless of distance

- To perceive the stimuli as the same size, as the distance increases, the stimulus must be shrunk

Front 34

What are some size constancy theories?

Back 34

1. Compensation using "perceived distance"

2. Emmert's Law: Perceived size for constant retinal size depends on perceived distance

3. Proportionality hypothesis: Relative size is important

4. Texture gradients set an absolute spatial scale

5. Relative height

- Attribute size of objects based on how far away you think it is

- Further away, the bigger the visual angle

- Thus infer size based on visual angle the object subtends

- Height is relative to the horizon

- Angle ratio roughly equal to height ratio

Front 35

What are some illusions that show size constancy?

Back 35

1. Hallway illusion / Ames Room

- Perspective information suggests depth

- Stimuli are the same size, but depth cue suggests some stimuli are further. So the ones "further" away are perceived as larger

- Stimuli are only perceived the same size if the "further" ones are made smaller. Though then in reality, they aren't same size (different sized retinal image)

2. Ponzo illusion

- 2 horizontal lines on a vertical 'road' seem different sizes, but actually are the same

- No real depth perception, but depth system is still activated

3. Moon illusion

- Full moon near the horizon looks huge, whereas if its in the sky, looks smaller

- When there is context (e.g. mountains), you compare relative size of the moon & scaling factors & therefore perceive moon as huge

- This is because the perceived path of the moon is much closer to you than actual path (sky seems flattened)

- Use angle from horizon to determine size of object

Front 36

What could underly some perceptual constancies in terms of receptive fields?

Back 36

View-invariant neurons

e.g. Large view-invariant receptive fields & view-invariant object recognition

Front 37

What are the major sources of information for depth perception?

Back 37

1. Oculomotor

- Accomodation

- Convergence

2. Visual

- Binocular vs Monocular

- If monocular; static cues vs motion parallax

- If static cues; Interposition, size & perspective

Front 38

What are oculomotor cues for depth? When do they work? How strong are these cues?

Back 38

1. Monocular = Accomodation

2. Binocular = Convergence

3. Proprioceptive & motor command cues:

- Copy of eye muscle commands

- Measures of strain from organ tendons

Work for short distances (<5m)

Are weak cues

Front 39

What are visual cues for depth?

Back 39

1. Shading

- Where the shading is (above or below) affects our depth perception

- Usually assume illumination comes from above because of the sun (prior assumption).

- Therefore if shading comes from below, we assume the object is sticking out towards us.

- If shading from above, assume object is concave and going away from us

2. Perspective

- Linear (e.g. lines converging to a 'vanishing point')

- Non-linear (aerial) effects (e.g. atmospheric effects of haze --> What isn't in the haze is closer)

3. Texture gradients

4. Motion parallax

- Objects further away appear to move slower relative to you than those closer

- Based on relative motion

- Arises with & without ego-motion

Front 40

How does stereovision help with vision?

Back 40

1. Puts mathematical constraints on visual system

2. Images of left & right eye are slightly different

3. Disparity = spatial shift of corresponding points on the retina

- Crossed = nearer

- Uncrossed = further away from horopter

- Horopter = surface with disparity 0 on the retina (all points are mapped on corresponding points on retina)

4. When eyes converge, position of image on retina will correspond. When they diverge, it will not

- Distance between the eyes constrains this

Front 41

How was stereovision measured and by who?

Back 41

via a stereoscope

Wheatstone 1838

Front 42

What is the correspondance problem? What is another name for it?

Back 42

An ill-posed problem with no unique solution.

Marr-Pogio algorithm

"Are 2 pixels of two pictures corresponding to each other?"

Front 43

What is the Marr-Pogio algorithm? Where in the brain does this occur?

Back 43

1. Each neuron is interested in relative difference between image 1 & 2 composed of black & white pixels

- Black pixel can come from any point of the image.

- Other image also has many light sources

- Each neuron corresponds to a certain depth

- Some neurons vote for relative image disparity

2. Neurons compete through lateral interactions.

- Those that vote for same depth, support each other (self-excitation) and compete (inhibit) with neurons that support the other depth

3. Result is that the most likely depth emerges as the winner & neural networks converge on how far the image is

4. Occurs in V1 & V2 to sense image disparity

Front 44

What is correspondance problem of motion based on?

Back 44

Time-to-time rather than space-to-space

Front 45

What are the neural mechanisms of image disparity?

Back 45

Disparity selective neurons in V1 & V2 (some also in V3, MT & MST) are selective to specific disparities but not others (disparity-tuned)

- Some will be excited if 2 spots are at slightly different locations

- Others excited more if 2 spots at same location (zero disparity)

Neurons have different tuning characteristics (e.g. some are tuned for far, some for near, some for zero disparity)