1. History of Computer Vision Graz, May 2008
Early contacts …
Computer Vision
Towards a Simple Brute-Force Utility?
Wilhelm Burger
FH OÖ / Campus Hagenberg – Digital Media
wilbur@ieee.org
Graz, May 2008
Kretztechnik
ca. 1980
2
ultrasound imaging … much hardware, little software …
3 4
Honeywell Systems
University of Utah (CS Dept., 1985) and Research Ctr.
Minneapolis (1986)
•Graphics???
•Evans & Sutherland
•Internet!
•VLSI-Processing
•2D Motion
Symbolics 3670
•Bir Bhanu, Tom Henderson 5 6
W. Burger, Hagenberg 1

2. History of Computer Vision Graz, May 2008
Driving the ALV … Motion flow fields
Autonompus Land Vehicle (ALV)
DARPA
7 8
Manual feature tracking … Computing ego-motion
9 10
Rules and Inference… Multiple hypotheses
11 12
W. Burger, Hagenberg 2

3. History of Computer Vision Graz, May 2008
Consistent worlds … Primitive motion simulation
13 14
Modeling ambiguity … In Retrospect …
“Strategic” projects
strong competition for money
vague specifications, assumptions, benchmarks …
unrealistic expectations
very little test data (imagery, ground truth)
highly creative branding (“smart”, “brilliant”, …)
Deficient technology
data capture (video!)
processing power
Brittleness everywhere
Ad hoc techniques
poor demos, inflated and unrelated results
15 16
What has changed since? What did not change …
Remarkable progress in last decade The visionaire’s toolbox remains limited
Success in specific applications histograms
voting
Major progress in
dimensionality reduction
3D reconstruction
random sampling
object detection/recognition
Much improved hardware Segmentation is still popular!
Public Awareness Development of platforms did not advance
High-level Vision?
17 18
W. Burger, Hagenberg 3

4. History of Computer Vision Graz, May 2008
Media-Related Applications … Statements I love …
image database annotation/retrieval
cinematography
“THE AI of this game was implemented …”
computer animation,
level-creation for 3D games,
“I am doing THE Computer Vision of this
virtual studios, mixed reality project”
video analysis
sports applications “Oh, and it MUST work on a mobile phone …”
smile detection ☺ (in real time)
Web (stitching, Google Earth, PhotoSynth etc.)
+ some bogus
19 20
Is Computer Vision solved? Research motives
Is CV more than Reasons for investigating a problem
inverted graphics? Everybody else does it
Nobody else does it
How “intelligent”
It is unsolved
should CV be? It is solved but you don’t know
What about You need to publish
semantics? …
Reasons for giving up on a problem
It is solved
It is too hard to solve
It is irrelevant (nobody cares)
You ran out of money
…
21 22
Evolution of problems Is CV a typical engineering problem?
Once solved, interest in a particular problem naturally declines …
The basic idea is
Degree simple …
Interest Success
t
Source: Ridiculous Research Inc. (2008)
Renewed Interest
23 24
W. Burger, Hagenberg 4

5. History of Computer Vision Graz, May 2008
Does “Brute Force”
Lead Anywhere?
Computing power IS essential
… but only needs
makes costly processes eventually feasible
refinement
not just a cheap excuse
von Neumann machines can do it
new technologies: GPUs, multi-core CPUs
But …
Software/platform development is neglected
Too much small-scale development
Environments needed for stable, continuous,
asynchronous, distributed, reconfigurable, debugable, …
operation.
25 26
Semantics – The Holy Grale? Trying to summarize
Automatic Speech Semantics is a completely open issue.
Translation Understanding
Don‘t try to emulate biological systems.
? Standard technology is fine (and evolving)
Large-scale/high-level frameworks deserve
renewed attention
Semantic
Web
Computer
Vision Computer Science education is not enough
Source: Ridiculous Research Inc. (2009)
27 28
Some final thoughts …
Interesting applications in Human-Computer
Interaction (all forms of “assistance”)
Learning
Too limited view of learning (classif. of feature vectors)? Thank you!
Small training sets (single exemplars)! And hold on to your dreams!
Breakdown and restructuring of concepts
Store and use “irrelevant” data
Simulators
Long-term training & testing on large data sets
Ground truth (almost) for free
29 30
W. Burger, Hagenberg 5