EE863: Computer Vision

1
EE863 Computer Vision

• We will stress fundamentals of machine vision
  science
• Develop an engineering approach to specifying,
  designing, and building vision systems
• Represented as a repeated, purposive exercise in
  perceptual organization and data abstraction

2
Things You Should Know

• Basic programming (C, C, Matlab, )
• Calculus and analytic geometry
• Basic probability and random processes
• Basic linear signals and systems
• Basic linear algebra
• You will learn some basic graph theory

3
What is Computer Vision?

• Marr To know what is where by looking.
• Information processing task
• Determine whats in cameras environment, and
  where, how posed
• Not just a process, because processes require
  representations
• A sequence of transformations through a series of
  representations, of increasing abstraction
• Appears to demand complete reconstruction of the
  physical scene

4
WICV?
Marrs definitition creates an unnecessary burden

• If we just want to build useful, task-specific
  systems complete reconstruction is overkill
• Systems that extract and process useful
  information from visual input, and make money,
  are always designed in the context of a larger
  problem, or task
• Even in biology, we do not find vision for
  visions sake

5
Vision in Context
The richest sense

• Earthworms, snails, and other critters
• Predators versus prey
• Race car drivers versus the submarine service
• The role of human and other biological vision
  systems in the study of machine vision

6
WICV?

• Horn Part of a sense-decide-act loop
• Action possibilities
• Robotics
• Fill in a database
• Prepare a report
• Draw a map
• The central issue of machine vision is the
  generation of a symbolic description (of the
  scene) from one or more images
• Meaningful Just rich enough

7
WICV, Not?

• Image processing
• Pattern recognition
• Imaging
• Graphics
• Multimedia

8
Why is CV Difficult?

• We have a very difficult, inverse problem to
  solve that is inherently mathematically
  ill-posed.
• Volkswagens versus Volvos
• Looking through a peephole
• We are all Mozarts of vision
• An infinity of solutions...

9
Applications
10
Perception is not a mere passive recording of
information impressed upon my sensory organs by
the environment. Rather, it consists of an
active construction by means of which sensory
data are selected, analyzed, and integrated with
properties not directly noticeable but only
hypothesized, deduced, or anticipated, according
to available information and intellectual
capabilities. Gaetano Kanizsa
Vision is not accomplished in the eye. Computer
vision is not accomplished in the camera.
11
So, How Do We Do This Thing?

• Apply world knowledge and heuristics
• Multiple images to produce more geometric
  constraints (constraints are our friends)
• Active sensors, in some applications
• Bear in mind These problems typically entail
  very large data sets

12
System Design Methodology
Six steps to success in computer vision

• Image formation
• Conditioning
• Labeling
• Grouping
• Extracting (attributing or characterizing)
• Matching

13
Conditioning

• Noise suppression
• Distortion mitigation
• Model Our image (or whatever) is an
  informative pattern corrupted by disturbances
• Impairment models
• Filtering, model fitting, response selection

14
Labeling
Here, we are still close to image and signal
processing

• Detection of elemental structures
• Assign into a small number of meaningful classes
• Primitive events such as
• Edges
• Corners
• Blobs
• Labeled pixels

15
Grouping

• AKA Perceptual organization
• Sweep together those elemental events that seem
  to be related in some way
• Spatial coherence criteria typical
• This is a key conceptual step
• Step out of the pixels and build a symbolic
  description of the image (content)

16
Extracting

• AKA Attributing or characterizing
• Assign attributes, or a list of properties, to
  each of the higher order groups
• Geometric
• Photometric
• Temporal
• Domain-dependent

17
Matching

• AKA Recognizing or classifying
• Assign a collection of features to some concept
  already known by the system
• A viewed instance of something (such as a
  chair)
• Another way around Function-Based

18
Lacing it Together

• The cycle of conditioning,, matching may be
  repeated multiple times in the analysis of a
  complex scene, at least conceptually
• May involve matching, etc., across multiple views
  (stereo, motion sequences)
• Some steps may be implicit in some cases

19
Control Issues

• Bottom up
• Top down
• Mixtures
• Bayesian networks
• Use the problem domain Take what the problem
  gives you