computational vision lecture 1: overview + biological vision jeremy wyatt
Post on 21-Dec-2015
218 views
TRANSCRIPT
Computational Vision
Lecture 1: Overview + Biological Vision
Jeremy Wyatt
What you should be able to do
Make informed choices about which sort of algorithms to apply to solve specific problems.
Use standard vision libraries or software to construct working vision systems.
Apply algorithms to simplified problems by hand.
Discuss the advantages and drawbacks of different methods, explaining their working.
Schedule
1 lecture a week, Mondays @ 2pm, Muirhead
1 lab/lecture a week, Thursdays @ 12pm (Robot Lab or Chem Eng)
I am currently away on Monday Oct 3 and Monday Nov 14, so there will be no lectures on those days
Syllabus
Lectures
1. Biological Vision2. Edge detection3. Hough transforms4. Motion/Depth5. Recognising objects6. Recognising events7. Recognising faces8. Visual attention
Labs
1. Matlab tutorials
2. Edge detection
3. Hough transforms
4. Face recognition
5. Object recognition
Assessment
70% 1.5 hour unseen exam in May/June
30% 3 page experimental write-up of one of your labs (in pairs)
(due Dec 7 12 noon)
Biological Vision
Light and image formation Retinal Processing Colour Visual Pathway Striate Cortex
Visible spectrum Humans perceive electromagnetic radiation
with wavelengths 380-760nm (1 nm = 10-9 m)
0.1nm 10nm 1000nm
Image Formation f is the focal length (in metres)
is the power of the lens (in dioptres)
Human eye has power ~59 dioptres
f
Image planeLensLight rays
1
f
150 0.02dioptres f m
f
Image Formation Most of the refractive power of the human eye comes from the air-
cornea boundary(49 of 59 dioptres)
As an object moves closer the power of the lens must increase to accommodate
So if the object is infinitely far away
But if it is 1m away the lens must change shape to produce a sharp image
u v
1 1 1
f u v
1 1 150
0.02dioptres
f
1 1 151
1 0.02dioptres
f
As an object moves in world how does it move across the image plane?
If the image plane is curved then as gets larger this becomes a worse and worse approximation
Image Formation
h
u
v
i
tan( )h i
u v
Retinal Processing 120m rods, 6m cones
Retinal Processing Amacrine and horizontal cells integrate
receptor outputs
More rods connect to each ganglion cells: less acuity, but greater sensitivity
Ganglions have receptive fields
Types of Ganglion cell Centre surround cells
ON area
OFF area
OFF area
ON area
Light spotTime
LightON Cell OFF Cell
Perceptual effects
These ON cells fire most
Grid of ON cell receptive fields
Colour Two theories/systems Trichromatic (Young-Helmholtz)
Explains– How we discriminate wavelengths 2nm in difference– How we can match a mixture of wavelengths to a single colour– Some types of colour blindness
Colour
Trichromatic theory can’t explain colour blending
?
?
Bluey green
Orange
Greeny red?
Yellowy blue?
Opponent Colour Theory
Ganglion ON cells sensitive to outputs of cones
ON
OFF
Opponent colour theory
Excitatory Inhibitory
Red on Green off Yellow on
After images
Visual pathway
The striate cortex Composed of hyper-columns Within each are columns of cells tuned to
features of a particular orientation
Summary
Image formation Very early visual processing Filling in and perceptual effects Colour perception Eye-cortex mapping
Reading
Vicki Bruce, Visual Perception, pp1-60 Neil Carlson, Physiology of Behavior,
pp142-157