PSY280 - 5. Object Recognition Flashcards
Spatial Rule
stimuli presented close together in space tend to be integrated
hear a sound + see something at same time - link it to single event
if at diff locations, don’t see it as single event
Temporal Rule
stimuli presented close together in time tend to be integrated
no integration when stimuli presented at vastly diff times
hear creak + door opens, doesn’t make sense unless simultaneous
Inverse effectiveness
effect of integration is more pronounced with weaker
stimuli
bigger enhancement of perception when dealing with degraded intensity
Multisensory Integration improves
detection
response times: faster to respond when presented in 2 modalities
accuracy: various judgements of stimulus when more than 1 presented
fundamental attributes
defer to vision for judgments about space, to hearing for judgments about time
to identify 2 objects - have to be separated in space
tend to assume auditory stimulus at the location of visual event
Middle vision
stage of processing that comes after basic features & before object recognition or scene
house recognition: mix of complex processes that allows us to recognize it
this is the object, other is the background
Where are the edges?
How can we tell that car & house separate objects?
How can we tell that snowman & house are separate objects?
How can we tell that the windows & the house are part of
the same object?
Where are the edges?
decide where 1 object ends + another 1 behind cell identifies lines
we identify objects because continuous + connected
connectedness can’t answer snowman because they’re connected
we know windows + house are same objects, yet they’re connected
some objects overlap
Connectedness won’t work
Connectedness isn’t even necessary
continuous edges aren’t necessary for our system to perceive edges
visual system fills in gap
we perceive more contrast than actually there
Even if connectedness did
work, what are the
important edges?
we tend to see objects in contexts
our visual system not challenge to find boundaries even if there are multiple edges
illusory contours
don’t even need edges to perceive segregation
contour that is perceived even though nothing changes from 1 side of the contour to the other
illusory contours perception without sensation
Structuralism
perceptions result from summation of many elementary sensations
Gestalt
wilhelm wudnt - complex objects/perceptions can be understood by analyzing components
Structuralism
components = elementary sensations
couldn’t explain illusory contours or apparent motion
elementary sensations of apparent motion that structuralism couldn’t explain
Structuralism
flashing stationary light moving on + off giving illusion of movement
more + more perceptual phenomenon that structuralism couldn’t explain
Gestalt
whole is different than sum of it’s parts
direct response to structuralism
heuristics used to guide interpretation of retinal image
sometimes >1 principle at work - difficult to predict which one is going to win all else being equal
Gestalt
Gestalt principles are like individuals with opinions
about how the information should be organized
under diff conditions, diff principles/rules might have diff opinions
Gestalt
discuss how to organize sensory input into coherent sensory experience
perception by committee - a consensus does quickly emerge that allows us to settle on a single interpretation of a visual scene
gestalt principles
- simplicity: assume simplest structure - simpler to see it as
- similarity: tend to group together stars + circles
- proximity: 3 groups of 2 lines
gestalt principles
- common fate: movement in the same direction
5. closure: things close shaped tend to be grouped together/close objects that are incomplete
gestalt principles
- familiarity: 45 means something to us - then we see us
- good continuation
- similarity: all orange dots, proximity: close together
gestalt principles: good continuation vs. similarity
When 2 principles operating, it’s hard to predict which one takes precedence.
general assumption that smooth lines belong together + lines tend to follow smoothest path
not strict rules, just heuristics
but if good continuation vs similarity pitted against each other, similarity wins
similarity has it’s limits
similarity has limits: works for colour, shape, size, orientation
but combo of these features doesn’t work too well
perception by committee
diff rules can come together + our perceptual experience isn’t always clear about which one wins
middle vision = specialists, diff in opinion about what sensory input might mean
how to organize info through committee
Pandemonium Model
Despite the potential for disagreement, there are some basic committee rules - example of how to implement computer vision model
Image demons like scanners: get sensory info
passes to first level of analysis: feature demons like specific features - cheer when preferred feature is present
Pandemonium Model
cognitive demons like specific letters: listen to feature demons
R cognitive demons making lots of noise, D demon + P demon is also making some noise
Decision demon listen to cognitive demons to see which one is cheesing the loudest
Every image is, in theory, ambiguous
multiple interpretations
every image is ambiguous
but committee comes to consensus quickly
using rules we can come to solution easily + rapidly
house rules: honour physics + avoid accidents
RULE 1: honour physics
come to implicitly understand that objects occlude light
assumption white square is occluding 4 discs behind it
visual system know about this + other physical properties + use them to organize visual info
RULE 2: avoid accidents
could be generated by quadrilaterals at diff depths + angles, just standing at one place that sees this
discards accidental viewpoints in consensus meeting
so unlikely that perceptual system doesn’t even consider it
An accidental viewpoint
viewpoint relative to an object that results in
perception of an accidental or rarely encountered property
What is the figure & what is the (back)ground?
needs to parse everything to separate objects + background
earlier gestalt rules can help us separate, but can’t help us differentiate which is background + object
What is the figure & what is the (back)ground?
some specific Gestalt rules that just apply to figure-ground
assignment (figure-ground segregation).
process of determining that some regions of an image belong to figure (object) + other regions are part of the background
Border ownership
associated with the figure
face/vase figure - border ownership changes based on how you interpret the image
figure & ground
surroundedness: likely surrounded region is figure
parallelism: contours parallel groups into a figure
size: smaller figure as object, larger object as background
figure & ground
relative location: lower region is object
symmetry: identify objects or features based on symmetry
meaningfulness: objects that are meaningful to us tend to be identified as objects
naive template theory
visual system recognizes objects by matching neural
representation of image with a stored representation of same “shape” in the brain
naive template theory
matching pixels - templates involve an array of spot detectors
if stimulus falls on filled spot detectors, match template
too many templates to be involved
recognition by components
structural description of an object in terms of nature of its constituent
parts & the relationships between those parts
recognition by components
parallel contours: lines always parallel to each other
exploits the concept of non-accidental properties: an edge feature that is not dependent on exact viewpoint.
recognition by components
always able to detect edge feature no matter what orientation except for 1 viewpoint: coin curves except when dead on
recognition by components
t-junctions (occlusion): occur when 1 surface occludes another
y- & arrow junctions (corners): signal corners
Geons
geometric ions - basic units of objects
36 geons, each composed of unique set of NAPs
elements that involve structure descriptions
combined non accidental properties in novel ways to represent 37 unique geons recognizable from any viewpoint
Geons
viewpoint-invariance: can be identified when viewed from most viewpoints nonaccidental properties (edges) axis straight or curved: skewer on constant/expanded: parallel?
Principle of componential recovery
If we can see the geons, we can identify the object
too many of nonaccidental properties are obscured meaning you can’t identify geons
if you move a little bit, you can identify it
need access to nonaccidental properties, to get geons, to get object
Geons
only need to see a few geons to identify an object
give them 3/9 78%, 6/9 90%
flexible templates
recognition is sometimes holistic
recognition is not completely viewpoint invariant.
experts on identifying greebles would do poorly if upside down
but rbc does get us far in object recognition
contextual modulation
evidence of Gestalt rules in the brain, both through neuroimaging & at the level of the neuron
contextual modulation: change based on context
contextual modulation
neuron’s response is influenced by stimulation outside it’s
receptive field
same stimulus in diff contexts
providing context of good continuation - neuron likes it more + is more active
contextual modulation: good continuation
likes a continuous line outside its receptive field
changing response based on stimulation outside receptive field
contextual modulation: figure-ground segmentation
neuron recognizes figure ground,
top: bar is in object, more response
bottom: bar is in background, less response
neural activity correlates with sensory experience
different sensory input
different neural activity
hard to tell if cat sees them as separate stimulus
does neural activity correlates with perceptual experience?
You have to be able to keep the sensory information identical, but somehow generate different perceptual
experiences.
does neural activity correlates with perceptual experience?
have to make task really difficult
show them image in peripheral vision
perception outlasts sensory stimulation
once removed still capable of receiving info unless you mask it by presenting another stimulus
does neural activity correlates with perceptual experience?
yes: All of these curves represent responses to
the presentation of Harrison Ford’s face.
had to identify what they saw: harrison ford, a diff face, or diff picture
does neural activity correlates with perceptual experience?
correct identification, face but incorrect identification
face but did not see face: all cases harrison ford was stimulus
fmri signal is huge when correct identification
identification is baseline when saw pattern
middle when saw face
binocular rivalry
You can also force 2 experiences
with the same stimuli
1 image to 1 eye, other image to another eye: back + forth because of rivalry
at fairly regular intervals - without permission
does neural activity correlates with perceptual experience?
same sensory input diff perceptual experience + diff neural signature
green face over red house
face comes in left eye, house on right eye
switches back + forth
does neural activity correlates with perceptual experience?
identify when they see house + when they see face by pressing button
fmri activity greater activation in ffa when seeing face, when house greater ppa activation than ffa
ppa when house + ffa when face