Future of Computer Vision

1
Future ofComputer Vision

• Horst Bischof
• Inst. for Computer Graphics and Vision
• Graz University of Technology

2
Motto of the talk

• It is a fantastic time

3
Motto of the talk

• to do computer vision!

4

• WHY?

5
Computer Vision

• At least three goals
• Understand biological visual systems
• Build machines that see
• What are the fundamental processes of seeing

6
Computer Vision

• The systems today are still exceedingly limited
  in their performance ? considerable room for
  improvement

Where are chairs?
How many feet?
Two interpretations?
7
Holy Grails in Vision

• Segmentation
• 2. Correspondence
• Recognition Problem

8
Future of Computer Vision

• Where do the innovations come from?
• 1. Hardware
• 2. Algorithms/Software

9

• HARDWARE

10
Hardware

• First time that HW is no longer a real limitation
  !!
• Processing
• Image Resolution
• Storage
• Internet
• Mobile Devices
• Networks of cameras

11
Processing

• Moores Law still holds!
• Multi-core CPUs
• Highly Parallel ? GPUs ( Software eg. Cuda)
• DEMO

12
Resolution

• Ever growing resolution
• 1975 100 x 100 0.01 MP
• 2008 9216 9216 85 MP
• (BAE)
• UltraCamx 216 MP
• New fantastic opportunities
• ? Computational Cameras

1900 Chicago Alton Railroad Train (photograph a
train), 5000
13
Internet

• Huge repository of images
• Flickr 3.Nov. 2008 3 Billion Photos On-line
• 1 Million added a day
• YouTube 65.000 new Videos a day
• 20 of Internet Traffic
• What can we do with these images?

14
Mobile Vision

• Most of us have a mobile CV device with them
• Small Cameras
• Embedded Systems
• Mobile CV next large application area
• Place Recognition
• Recognizing Tags
• Shopping
• Games
• Augmented Reality

15

• ALGORITHMS

16
(Some) New Developments
17
Bayesian Methods

• Lots of applications
• Computationally heavy
• Easily parallelizable
• ?Energy minimization approaches

Data
Ill-posed
Prior
18
Energy minimization
Level Sets Convex formulations Graph cuts
Continuous Continuous Discrete
Local optima Global optima Global optima
GPU Implementation Memory limitations
Metrication errors
More to come
Apapted f. D Cremers 2007
19
Continous energy functional
Total Variation regularization
Data term

• Data term potentially non-convex ? Global Optimal
  Solution
• Defines domain of application
• Denoising
• Segmentation
• Stereo

Pock et.al
20
Vision Learning

• Combining Computer Vision with ML
• ? Huge Success
• We have good/stable features
• SIFT
• Boundary fragments
• If enough data learning works
• SVM
• Boosting