object recognition

Question 1

what is the goal of our object recognition system?

Answer 1

must be able to deal with different representations of the same object

Question 2

what does the goal of the object recognition system apply to?

Answer 2

structures differences

differences in pose, distance, lighting, position, viewpoint

degraded images - occlusion, noise, distortion, filtering

Question 3

what is the two-stream model?

Answer 3

post V1, information is transmitted via two pathways - ventral and dorsal stream

more complex networks link to frontal lobe and other areas

Question 4

what is the ventral stream?

Answer 4

what pathway

from V1 to V2 then V4 and IT cortex

associated with object recognition and memory

Question 5

what is the dorsal stream?

Answer 5

where and how pathway

from V1 to V2 and finally V5/MT

associated with motion, location, saccadic control

Question 6

what is object agnosia?

Answer 6

damage to ventral stream can create a deficiency in object recognition

Question 7

what is the problem with models of object recognition?

Answer 7

all models share a common assumption

the senses register the presence and an internal representation of the stimulus is generated (perceptual representation)

object is recognised when there is a match between the perceptual representation and some stored representation of the same object

object recognition requires the interaction of perception and memory

Question 8

what did David Marr assert about models of object recognition?

Answer 8

necessarily includes a computational approach

should have three levels of analysis - computational (what is the system doing and why?), algorithmic (which processes, rules and algorithms are used to solve the problem?), implementational (how are these processes implemented by the system?)

Question 9

what are template-matching models?

Answer 9

simplest model for object recognition

assumes we have templates - perceptual representations stored in our long-term memory

template is available for any object

when an object appears in the receptive field of this detector that matches that template, it signals

Question 10

what is the problem with template-matching models?

Answer 10

work well but letters must be in exactly expected location and orientation

for this to work, would need a detector for every possible orientation, scale and font - requiring an impossibly large brain

Question 11

what are feature detection models?

Answer 11

Selfridge’s Pandemonium model (1959)

a built up template model

described in terms of demons with different jobs

model fits well with Hubel & Wiesel’s work which suggests feature detecting neurons

Question 12

what are feature demons in the feature detection model?

Answer 12

look at image and simply write down how many examples of their feature they see

Question 13

what are cognitive demons in the feature detection model?

Answer 13

shout if they think that the combination of features applies to their letter

more confident they are, louder they shout

Question 14

what are decision demons in the feature detection model?

Answer 14

listens to the cognitive demons and decides who is shouting the loudest, providing that as the perceived letter

Question 15

how are early implementations of the feature detection model crude?

Answer 15

feature demons still using templates

still no information on configuration

still can’t distinguish between different versions

Question 16

what are the controversies in object recognition?

Answer 16

if asked to describe shape of coin, might say its a disc - this is always true, its a feature of the object

physical shape of coin is viewpoint-independent

perceptual representation changes and therefore is viewer-dependent

Question 17

what is Marr & Nishihara’s (1978) structural description model?

Answer 17

believed goal of model is to describe to object unambiguously therefore system must be invariant to transformations in viewpoint, illumination, etc which means the system must know which properties are invariant under transformation and how other properties may vary

coordinate system is object centred, negating the problem of transformation variance

Question 18

what are the primitives in Marr & Nishihara’s (1978) structural description model?

Answer 18

primitive are basic units of information in its representation

volumetric approach - volumes only require axis and size info - maintains specificity without requiring too much storage space

see everything in cylinders

Question 19

how is information organised into an object description in Marr & Nishihara’s (1978) structural description model?

Answer 19

viewer centred - input image, edge image , 2.5D sketch
object centred - 3D model

at 4th stage, system creates viewpoint-independent representation of the object by describing its volumetric structure, 3D model enables consistent recognition of the object from any angle, ensuring the object can be identified no matter how it is viewed

object is described in terms of its axes and the volumes around them - description is modular and hierarchical meaning the object can be described at many scales, allowing for identity matching and discrimination

Question 20

what is the input image in Marr & Nishihara’s (1978) structural description model?

Answer 20

retinal image

intensity and wavelength of light at each point

Question 21

what is the edge image in Marr & Nishihara’s (1978) structural description model?

Answer 21

zero crossing, blobs, edges, bars, ends, curves, boundaries

Question 22

what is the 2.5D sketch in Marr & Nishihara’s (1978) structural description model?

Answer 22

surfaces with local orientations and discontinuities in depth

Question 23

what is the 3D model in Marr & Nishihara’s (1978) structural description model?

Answer 23

composed of 3D “primitive” volumes, organised hierarchically by scale

Question 25

what is the recognition - "model" store - in Marr & Nishihara's (1978) structural description model?

Answer 25

in our brain, have full catalogue of 3D representations of objects even if your object perception doesn't exactly match anything in your model store, find closest match have sufficient information on the object from the image and your memory to help you interact with it (image manipulation)

Question 26

what is Biedermann's (1987) structural description model?

Answer 26

model also called recognition by components (RBC) proposed a set of primitive volumes into which objects are decomposed (not just cylinders) physiological evidence presented from neuronal recordings

Question 27

what are non-accidental properties in Biedermann's (1987) structural description model?

Answer 27

collinearity (straightness of lines) curvature symmetry parallelism co-termination

Question 28

what are geons in Biedermann's (1987) structural description model?

Answer 28

volumes are geons (geometric ions) - many features of these geons remain in 2D and 3D space (non-accidental properties) estimated that there are <=36 geons therefore 36^2 = 1296 pairs of geons, which can be attached in different ways and of different relative sizes - proposed that there are ~75,000 possible 2-geon objects

Question 29

what are experimental evidence did Biedermann's (1987) present for structural description model?

Answer 29

for these geons in human object recognition right column is missing key (2D) geons, while centre column is missing an equal amount of other contour info object recognition was much poorer in missing-geon condition (solid lines), particularly when presented for a short time

Question 30

how does partial occlusion affect Biedermann's (1987) structural description model?

Answer 30

changes objects from flat to 3D visual system will generate hypothesis about how the object's contours may continue behind the occluder - cognitive processes that allow people to infer the full shape of an object even though it's partially hidden geons theory struggles to explain how humans perceive and recognise objects that are partially obscured as it doesn't fully account for this inferential processing

Question 31

what are the pros of structural description models?

Answer 31

invariance is well-explained recognition relies on description rather than matching graded representations cope with discrimination and generalisation evidence that structural information matters to humans and neurons

Question 32

what are the cones of structural description models?

Answer 32

extracting model parameters can be hard in real images structural description is difficult for some objects driven by theoretical desirability rather than behavioural or physiological evidence

Question 33

what are view-dependent models?

Answer 33

don't need to catalogue of fixed 3D models but rather a catalogue of shape descriptions that match view-dependent characteristics of objects brute force association finding of a canonical perspective supports the idea of an experience-based catalogue uses a viewer-based coordinate system weighted approach between layers

Question 34

what is a canonical perspective in view-dependent models?

Answer 34

what comes to mind when we think of an object view that seems to maximise the information use for recognition

Question 35

what are primitives in view-dependent models?

Answer 35

sub-regions of the image - "abstract" features that might consist of lines, curves, textures, colour, shading feature-sensitive units combine into each other in a weighted way, getting more complication size and position invariant (because IT neurons have big receptive fields) these feed into view-tuned object recognition cells recognition by matching input to closest stored view

Question 36

what is a view layer in view-dependent models?

Answer 36

units code a particular view of the object

Question 37

what is an object layer in view-dependent models?

Answer 37

units respond to a particular object view-independent

Question 38

what is the evidence for view-dependent models?

Answer 38

human object recognition is not always viewpoint invariant viewing sphere - practiced recognising objects from specific viewpoints, tested at novel viewpoints - interpolation (between previous viewpoints) easiest, extrapolation (beyond previous viewpoints but in same exists) medium difficulty, orthogonal axis (from a completely new viewpoint) hardest behavioural, physiological evidence exists

Question 39

what are the pros of view-dependent models?

Answer 39

straight forward minimises transformations that must be performed newer models are based directly on what we know of physiology abstract features are recombinable good behavioural, physiological and simulation-based evidence

Question 40

what are cons of view-dependent models?

Answer 40

humans often show quite good generalisation across viewpoints even for novel objects still more memory intensive than e.g. geon model