w10 gemini

Question 1

What is the difference between the “viewer-centred” and “object-centred” approach to object recognition?

Answer 1

In the viewer-centred approach, the 3D object is modeled as a set of 2D images, showing different views of the object. In the object-centred approach, a single 3D model is used to describe the object.

Question 2

What is a geon?

Answer 2

Geons are simple three-dimensional shapes such as spheres, cubes, cylinders, cones, or wedges.

Question 3

What is a structural description?

Answer 3

A structural description is a representation of an object in terms of its component geons and their relative locations and sizes.

Question 4

Explain the nearest mean classifier method.

Answer 4

The nearest mean classifier calculates the mean of the feature vectors for all the training examples in each class (the prototype). For a new object, it finds the closest class prototype (using Euclidean distance) and assigns the new object to that class label.

Question 5

Explain the nearest neighbour classifier method.

Answer 5

The nearest neighbour classifier saves the feature vectors for all the training examples. For a new object, it finds the closest training example (using Euclidean distance) and assigns the new object to the same class label as that closest example.

Question 6

Explain the k-nearest neighbour classifier method with k=3.

Answer 6

The k-nearest neighbour classifier with k=3 saves the feature vectors for all the training examples. For a new object, it finds the 3 closest training examples (using Euclidean distance). The new object is assigned to the class label that is the majority among these 3 nearest neighbours.

Question 7

Write down Bayes’ theorem.

Answer 7

p(H|E) = (p(E|H) * p(H)) / p(E)

Question 8

Explain the interpretation of p(H|E) in Bayes’ theorem in relation to a computer vision system.

Answer 8

p(H|E) is the posterior probability that a hypothesis H (e.g., the object is a chair) is true, given the image evidence E. This is what the vision system needs to evaluate to determine the most likely explanation for the image data.

Question 9

Explain the interpretation of p(E|H) in Bayes’ theorem in relation to a computer vision system.

Answer 9

p(E|H) is the likelihood that if hypothesis H were true, the image would contain particular evidence E. This is based on our understanding of how images are formed, such as how surface properties and lighting create certain images.

Question 10

Explain the interpretation of p(H) in Bayes’ theorem in relation to a computer vision system.

Answer 10

p(H) is the prior probability, representing our initial assumptions about the likelihood of a hypothesis being true before seeing any evidence. If a hypothesis is initially improbable, stronger evidence is needed to support it.

Question 11

Explain the interpretation of p(E) in Bayes’ theorem in relation to a computer vision system.

Answer 11

p(E) is the probability of observing the evidence E regardless of whether the hypothesis H is true. If the evidence is very common, it reduces our confidence in inferring a specific hypothesis based on that evidence.

Question 12

In the production line problem, what is the probability of objA?

Answer 12

p(objA) = 0.75

Question 13

In the production line problem, what is the probability of objB?

Answer 13

p(objB) = 0.25

Question 14

In the production line problem, what is the probability of an indistinguishable image given objA at oriA?

Answer 14

p(I|objA) = 0.1

Question 15

In the production line problem, what is the probability of an indistinguishable image given objB at oriB?

Answer 15

p(I|objB) = 0.2

Question 16

Using Bayes’ theorem, if an indistinguishable image is observed, what is the probability it is objA at oriA?

Answer 16

p(objA|I) = (p(I|objA) * p(objA)) / p(I) = k * (0.1 * 0.75) = 0.075k

Question 17

Using Bayes’ theorem, if an indistinguishable image is observed, what is the probability it is objB at oriB?

Answer 17

p(objB|I) = (p(I|objB) * p(objB)) / p(I) = k * (0.2 * 0.25) = 0.05k

Question 18

In the production line problem, if an indistinguishable image is observed, which bin should the robot sort the object into to minimize errors?

Answer 18

The robot should sort the object into the bin for objA because p(objA|I) > p(objB|I), meaning it’s more likely to be objA.

Question 19

What is the viewer-centred approach to object recognition?

Answer 19

The 3D object is modelled as a set of 2D images, showing different views of the object. Recognition occurs by matching the current view to a stored view.

Question 20

What is the object-centred approach to object recognition?

Answer 20

A single 3D model is used to describe the object. Recognition involves decomposing the viewed object into its components (geons) and matching their arrangement to stored models.

Question 21

What are geons designed to be?

Answer 21

Geons are designed to be sufficiently different from each other to be easily discriminated, robust to noise (identifiable even with missing parts), and view-invariant (look similar from most viewpoints).

Question 22

What are the problems with the object-based recognition by components theory?

Answer 22

It can be difficult to decompose an image into geons, it’s difficult to represent many natural objects using geons, and it cannot detect finer details necessary for identifying individuals or similar objects.

Question 23

What is the image-based approach to object recognition?

Answer 23

Each object is represented by storing multiple 2D views (images). Object recognition occurs when a current pattern matches a stored pattern.

Question 24

What is template matching in the context of image-based recognition?

Answer 24

An early form of image-based approach where the current view is directly compared to stored templates. It’s considered too rigid to account for the flexibility of human object recognition.

Question 25

What is the multiple views approach in the context of image-based recognition?

Answer 25

A more recent image-based approach where multiple views of objects are stored. Recognition occurs by matching the current view, and interpolation between stored views allows recognition from novel viewpoints.

Question 26

What are configural theories of object recognition?

Answer 26

Configural theories emphasize the global shape and relationships between features in object recognition.

Question 27

What are featural theories of object recognition?

Answer 27

Featural theories emphasize the individual features of an object in recognition.

Question 28

How do inverted faces relate to configural and featural processing?

Answer 28

Inverted faces are processed featurally, meaning individual features are processed independently, and the relationships between them are ignored. Upright faces are processed configurally, or holistically.

Question 29

What are the three main types of theories for how we categorize objects?

Answer 29

Rules, prototypes, and exemplars.

Question 30

How does the ‘rules’ theory explain object categorization?

Answer 30

Category membership is defined by abstract rules. Anything that satisfies the rule(s) for the category belongs to that category.

Question 31

What are some arguments for and against the ‘rules’ theory?

Answer 31

For: explains over-extension of grammar rules. Against: some members are better examples of a category (graded membership).

Question 32

How does the ‘prototypes’ theory explain object categorization?

Answer 32

We calculate the average (or prototype) of all individual instances from each category. A new stimulus is compared to these stored prototypes and assigned to the category of the nearest one.

Question 33

What are some arguments for and against the ‘prototypes’ theory?

Answer 33

For: explains why prototypical category members are accessed more quickly. Against: variations within a class cannot be represented.

Question 34

How does the ‘exemplars’ theory explain object categorization?

Answer 34

Specific individual instances of each category (‘exemplars’) are stored in memory. A new stimulus is compared to these stored exemplars and assigned to the category of the nearest one.

Question 35

What are some arguments for and against the ‘exemplars’ theory?

Answer 35

For: explains some kinds of mis-categorizations. Against: struggles with the concept of graded membership (some members being better examples).

Question 36

What is supervised learning in the context of classification?

Answer 36

Learning where the class for each data point in the training set is known. New data points are assigned to appropriate classes based on similarity to these labelled training examples.

Question 37

What is unsupervised learning in the context of classification?

Answer 37

Learning where the class for each data point is unknown. All data points are assigned to appropriate classes based on similarity without pre-existing labels.

Question 38

What is the nearest mean classifier (prototype) and when is it suitable?

Answer 38

It calculates the mean feature vector for each class and assigns new data to the class with the nearest mean. It’s suitable only if the data is linearly separable.

Question 39

What is the nearest neighbour classifier (exemplar) and when is it suitable?

Answer 39

It assigns new data to the class of the nearest training example. It’s suitable if the data is non-linearly separable.

Question 40

What is the k-nearest neighbours classifier (exemplar) and when is it suitable?

Answer 40

It assigns new data to the class that is the majority among its k nearest training examples. It’s suitable if the data is non-linearly separable.

Question 41

What are some common similarity measures used in classification?

Answer 41

Sum of Squared Differences (SSD), Euclidean distance, Sum of Absolute Differences (SAD) (Manhattan distance), Cross-correlation, Normalised cross-correlation, Correlation coefficient.

Question 42

Describe the ‘What’ and ‘Where’ pathways in the cortical visual system.

Answer 42

The ‘What’ pathway (ventral stream) goes from V1 to the inferotemporal cortex and is involved in object identity and category information. The ‘Where’ pathway (dorsal stream) goes from V1 to the parietal cortex and is involved in spatial and motion information.

Question 43

How are receptive fields organized hierarchically in the visual cortex?

Answer 43

As you progress along a pathway, neurons’ preferred stimuli get more complex, receptive fields become larger, and there is greater invariance to location.

Question 44

Give examples of receptive fields at different stages (Eye to V1).

Answer 44

Eye -> LGN: Centre-surround Cells. LGN -> V1: Simple Cells (respond to edges/bars at a specific orientation and location). Simple Cells -> Complex Cells (respond to edges/bars of a specific orientation within a small region).

Question 45

What is the trend of receptive fields along the ventral pathway?

Answer 45

Receptive fields become larger, have higher complexity, and higher invariance to location.

Question 46

What are feedforward models of cortical hierarchy?

Answer 46

Models that propose a purely serial, feedforward sequence of cortical information processing, like HMAX and CNN.

Question 47

Explain the HMAX model.

Answer 47

HMAX is a feedforward model that uses alternating layers of simple (S) and complex (C) cells to increase selectivity and invariance of receptive fields. S-cells respond to conjunctions, and C-cells respond to any input in a small neighbourhood.

Question 48

How do S-cells and C-cells operate in HMAX?

Answer 48

S-cells perform a summation (‘and’-like operation) increasing selectivity. C-cells perform a max operation (‘or’-like operation) increasing invariance.

Question 49

How can HMAX be seen as hierarchical template matching?

Answer 49

Features at one stage are built from features at earlier stages, forming increasingly complex templates.

Question 50

What is a Convolutional Neural Network (CNN)?

Answer 50

A hierarchical model similar to HMAX that uses standard image processing techniques: convolution and sub-sampling.

Question 51

How do convolution and sub-sampling relate to HMAX?

Answer 51

Convolution in CNNs is equivalent to the function of S-layers in HMAX (responding to conjunctions). Sub-sampling in CNNs is equivalent to the function of C-layers in HMAX (responding to any input in a small neighbourhood).

Question 52

What are recurrent models of cortical hierarchy?

Answer 52

Models that incorporate feedback connections and lateral connections within cortical regions, allowing for interaction and combination of bottom-up and top-down information.

Question 53

What are the two types of recurrent connections in the cortex?

Answer 53

(1) Lateral connections within a region enabling interaction between neurons in the same population. (2) Feedback connections conveying information from higher cortical regions to primary sensory areas.

Question 54

What are bottom-up processes in perception?

Answer 54

Using the information in the stimulus itself to aid in identification. They are stimulus-driven and discriminative.

Question 55

What are top-down processes in perception?

Answer 55

Using context, previous knowledge, and expectation to aid in identification. They are knowledge-driven and generative.

Question 56

How does Bayesian inference relate to bottom-up and top-down information?

Answer 56

Bayes’ Theorem describes an optimum method of combining bottom-up (likelihood) and top-down (prior) information to form a posterior probability.

Question 57

Explain Bayes’ Theorem using the terms posterior, likelihood, prior, and evidence.

Answer 57

Posterior (p(object|image)) is what we want to know: the probability of a particular object being present given the image. Likelihood (p(image|object)) is what we can calculate: the probability of the particular image being a projection of the particular object. Prior (p(object)) is what we know from prior experience: the probability that the particular object will be present in the environment. Evidence (p(image)) is the probability of observing the image, which can often be ignored as it’s the same for all possible interpretations.

Question 58

Why is vision considered an inverse problem?

Answer 58

Because we know the pixel intensities (outcomes) and want to infer the causes (objects in the scene).

Question 59

Why is vision considered ill-posed?

Answer 59

Because there are usually multiple solutions (multiple causes that could give rise to the same outcomes).

Question 60

How does the visual system compensate for the ill-posed nature of vision?

Answer 60

By using assumptions, constraints, or priors about the nature of the physical world.

Question 61

Give examples of priors used in Bayesian inference in vision.

Answer 61

Texture is circular and homogenous, light comes from above, faces are convex, size is constant, neighbouring features are related, similar features are related, connected features are related, strings of letters form words, knowledge about image content.

Question 62

What is the difference between discriminative and generative methods in the context of Bayesian inference?

Answer 62

Discriminative methods model the posterior probability directly. Generative methods model the likelihood and the prior probability.