high level vision

Question 1

why is object recognition hard?

Answer 1

because objects with similar functions come in different shapes and sizes and are viewed from different perspectives

Question 2

what is representation?

Answer 2

the ability to match object to what we have stored in memory

Question 3

what are the steps of perceptual organisation?

Answer 3

• identify edges and uniform regions
• group figure regions together
• fill in missing edges and surfaces

Question 4

what does bierdman’s recognition by components theory suggest?

Answer 4

• restricted the set of 3D volumes to represent part shape - geons
• geons are viewpoint invariant (can see lines and vertices despite rotations in depth)
• objects are broken into geons
• object representations are assemblies of geons and their spatial relations

Question 5

what is the support for RBC?

Answer 5

removal of contours defining concavities affect object recognition

Question 6

what is the alternative to RBC?

Answer 6

we may just do it as they appear from a particular view and we have a stored representation with unfamiliar views needing transformation

Question 7

what is recognition best for?

Answer 7

• learnt viewpoints
• viewed within 40 degrees of trained viewpoints
• could also interpolate between learned views

Question 8

how does visual hierarchy impact representations?

Answer 8

representations of shape in V4 is richer than V1 but inferno is richer in V1 than V4 as receptive field size increases and neurons become more selective

Question 9

so how do we recognise objects?

Answer 9

no convincing theories on the debate between structural descriptions and view based representations

Question 10

are cell’s grandmother?

Answer 10

no, probability of finding one cells that responds to one object is very small

Question 11

how many cells respond to seeing one person?

Answer 11

each cells probably responds to 50-150 people

Question 12

what are the advantages of sparse and invariant coding?

Answer 12

• carries a large amount of info
• energy efficient
• allow a large storage capacity
• tolerance to degradations

Question 13

what are deep neural networks?

Answer 13

multi level networks that can be trained to recognise objects

Question 14

what can deep neural networks do overtime?

Answer 14

recognise new instances of the object that it has never been trained on when shown many instances of an object with feedback

Question 15

what is visual agnosia?

Answer 15

impairment in deriving the meaning of a visually presented stimulus

Question 16

how can you recognise objects if you have visual agnosia?

Answer 16

via sensory modalities

Question 17

what is prosopagnosia?

Answer 17

the selective disruption of the perception of faces, one’s own face as well as those of others, which are seen but not recognized as faces belonging to a particular owner

Question 18

what are the levels of recognition?

Answer 18

superordinate, entry level and subordinate

Question 19

what is the superordinate level of recognition?

Answer 19

a more general term for an object

Question 20

what is the entry level of recognition?

Answer 20

the label that comes to mind most quickly

Question 21

what is the subordinate level of recognition?

Answer 21

a more specific term for the object ie. a name

Question 22

where are prosopagnosia lesions located and how do they happen?

Answer 22

located in ventral occipitotemporal cortex

and can result from unilateral right hemisphere regions

Question 23

what is specialisation?

Answer 23

cells respond selectively to specific stimulus characteristics

Question 24

what is modularity?

Answer 24

cells that respond to similar stimulus characteristics are clustered into specific brain regions

Question 25

what area is there evidence for modularity?

Answer 25

face identity processing