1B50 – Percepts and Concepts Daniel J Hulme. Outline Cognitive Vision –Why do we want computers to see? –Why can’t computers see? –Introducing percepts.

1B50 – Percepts and Concepts

Daniel J Hulme

Outline

• Cognitive Vision– Why do we want computers to see?– Why can’t computers see?– Introducing percepts and concepts

• Visual System– The Eye and Brain– Early visual processes– Edge Detection

• Percepts and Concepts– Late Visual Processes– Concepts

Lecture 1: Reminder• Cognitive Science: scientific study of intelligence• Intelligence: …. (something to do with brains?)

• Vision is an integral part (and catalyst for the evolution) of the brain

• Ambiguity and the Distal and Proximal stimulus

• Using experience to construct (perceive) one form from a potentially infinite amount of possible forms

Lecture 2: Reminder• The significance of retinal structure

– Rods and Cones distribution

• Receptive Fields and Neural Nets

• Early visual process: Edge Detection

• Convolution between an image and a kernel

Fovea

1

2

3

1

2

3

0

0

Rods & Cones

∑

Sum Inputs

∑

∑

∑

ActivationFunction

Ganglion Cells

∑

∑

∑

∑

Optic Nerve

Activation

Horizontal & Bipolar Cells

Weighting & Join Inputs

Light Source

Stimuli Detectors

1

20

15

10

2

15

0

0

> 50

> 50

> 50

> 50

> 50

> 50

> 50

> 50

x 1

x 5

x 10

Periphery

1

2

3

1

2

3

0

0

Rods & Cones

Sum Inputs

ActivationFunction

Ganglion Cells Optic Nerve

Activation

Horizontal & Bipolar Cells

Weighting & Join Inputs

Light Source

Stimuli Detectors

1

20

15

10

2

15

0

0

> 50∑

63

x 1

x 5

x 10

Fovea

10

20

30

10

20

30

0

0

Rods & Cones

∑

Sum Inputs

∑

∑

∑

ActivationFunction

Ganglion Cells

∑

∑

∑

∑

Optic Nerve

Activation

Horizontal & Bipolar Cells

Weighting & Join Inputs

Light Source

Stimuli Detectors

10

200

150

100

20

150

0

0

> 50

> 50

> 50

> 50

> 50

> 50

> 50

> 50

x 1

x 5

x 10

Periphery

Rods & Cones

Sum Inputs

ActivationFunction

Ganglion Cells Optic Nerve

Activation

Horizontal & Bipolar Cells

Weighting & Join Inputs

Light Source

Stimuli Detectors

10

200

150

100

20

150

0

0

> 50∑

630

x 1

x 5

x 10

10

20

30

10

20

30

0

0

Receptive Fields• Receptive field – the photoreceptors that affect

the ganglion cell

• One photo-receptive cell (rod or cone) may be a member of several receptive fields

• Tile the retina surface

• Always circular in shape

• On center, off surround Off center, on surround

• Edge (contour) sensitive

• Receptive fields are modeled by Difference of Gaussians

Primary Visual Cortex

• Groups of neurons process information about:– Form of objects– Contrast of objects– Location of objects– Movement of objects– Color of objects

Visual Cortex Cells Response

• Lines or edges with certain orientation or size

• Angles or corners

• Movement in one direction, but not another direction

• Two-thirds of vision research involves these types of cells

• It is thought that more complex cells actually respond to specific faces, etc

Vertical Receptive Field

Overlapping and Orientation

Recognising Objects• It is not completely know how we perceive solidity/planes

• Gestalt ‘grouping’ school of thought:

– proximity - how elements tend to be grouped together depending on their closeness

– similarity - how items that are similar in some way tend to be grouped together

– closure - how items are grouped together if they tend to complete a pattern

– continuity - how items are organized into figures according to symmetry, regularity, and smoothness

Electrophysiology

Stereopsis - Stereo (binocular) vision

• Allows us to approximate distance of objects up to a few meters away

• Point matching procedure is used to calculate disparity (use template matching)

• Binocular disparity relates to depth

Monocular Disparity• Monocular cues are cues to depth that

are effective when viewed with only one eye.

• Interposition: When one object overlaps or partly blocks our view of another object, we judge the covered object as being farther away from us

• Atmospheric Perspective: The air contains microscopic particles of dust and moisture that make distant objects look hazy or blurry

• Texture Gradient: A texture gradient arises whenever we view a surface from a slant, rather than directly from above.

• Linear Perspective: Linear perspective refers to the fact that parallel lines, such as railroad tracks, appear to converge with distance

• Size Cues: Consider the size of an object's retinal image relative to other objects when estimating its distance.

• Height Cues: We perceive points nearer to the horizon as more distant than points that are farther away from the horizon

• Motion Parallax: Motion parallax appears when objects at different distances from you appear to move at different rates when you are in motion

Motion

• Object displacement usually correlates to depth. I.e. objects moving towards us usually expand

• Visual system correlates image points from one moment to the next

• Evidence of short range and long range motion detectors

Continuity

Continuity

Concepts

• One cannot fully explain perception without showing that the beliefs it produces tends to be true

• The benefit of perception is to yield true beliefs – even if this means generating ‘incorrect’ perceptions

• Observable and Hidden Variables

• Uggs Valley

Closing remarks

• Cognitive Science as a science

• Sub-symbolic vs Symbolic

• Classical AI vs Modern AI

• Bayesian approach

• Computational issues

• How to solve the problem…

Questions