Department of Computer Science and
Engineering - University of Bologna
Introduction
http://didattica.arces.unibo.it/
What is Computer Vision ?
Computer Vision deals with extraction of information from images
Computer
Input Image Output Information
Vision
and is therefore different from Image Processing
Image
Input Image Output Image
Processing
Computer Vision & Image Processing
quite often, though, Image Processing helps Computer Vision:
Image Computer
Input Image Output Image
Processing Vision
Output Information
Object Classes,
Position and Orientation,
Measurements
Why studying this stuff ?
Computer
Vision
Countless applications
in many diverse fields !
ENG BIO
Past.
the block world, Roberts, MIT, 1965
..from labs to the
factory
Cognex DataMan, 1982
first industrial OCR system
Through successes and failures, computer vision has
eventually reached maturity. The technology is now as
effective, reliable and affordable as to be deployed in
countless industrial applications.
Present.
massively into the factory
76,4%
Inspection,Gauging and Guidance
A case study: tile inspection
Old systems for detecting broken
tiles were mostly based on photocells.
The two borders orthogonal to the
motion direction cannot be inspected.
Small rotations can yield false
positives.
Poor flexibility.
A computer vision approach turns out
significantly more effective !
PCB inspection
Computer vision systems are widely deployed to check PCB
wires. Moreover, component placement inspection also
relies on vision systems.
More recently, 3D vision systems have been
introduced in the electronic manufacturing
process to carry out Solder Paste Inspection
of SMT components.
Robot Guidance
Computer vision is key to endow industrial robots with an
otherwise unattainable degree of flexibility.
The ability to detect and localize a few objects allows the robot
to adapt its operation to a wide range of unknown arrangements
(i.e. type of objects as well as their positions and poses).
The Present.
Turnkey systems designed for a
Application specific application, which is either
Specific Vision widespread throughout the industry or
Systems used in a particular sector.
The user can configure the system
Configurable without writing or changing
the source code. Typically PC-based,
Complexity & Cost
Vision Systems
flexible multi-camera support.
Cameras with embedded computational
Smart intelligence (e.g. a CPU,DSP or FPGA),
Cameras which can be programmed or
configured.
Turnkey products combing a sensor and a
Vision processor. Equipped with specific
Sensors application software geared at a widespread
use (e.g. presence/absence, code reading)
Present.
+14.1% vs +8.2%
23,6%
(10% in 2011)
increasingly into everyday life
Everyday life applications
Everyday life applications
Key advances: Features !
Local Invariant Features the detector/descriptor paradigm allows finding
and matching salient regions across images invariantly to viewpoint
changes.
Roberts, MIT, 1965
SIFT, Lowe, UBC, 2004
and Depth Sensing
Affordable Depth Cameras key to controlling machines by gestures and/or
body movements (Natural User Interfaces).
First generation Kinect, Asus Xtion:
structured light approach based on
PrimeSenses technology. now for
Apple only !
Second generation Kinect
(Xbox One), Intel RealSense :
TOF depth sensing technology
NUI - Kinect Body Tracking
31 Body Parts
Training on synthetic depth images ( 1 million)
Computer Vision & Machine
Learning
Low-Level Cues Decision Rules
Features
Image Design Design Information 2000
Features
Image Design Learn Information 2012
Features
Image Learn Learn Information
Deep Learning aka End-to-End Learning
Deep Learning
Truly data-centric
paradigm:
Big and good data
Value shift from
domain knowledge to
data
Deep Learning
Top performance in the most difficult computer vision tasks, often better
than human performance
YOLO watching Skyfall
DeepFace by FAIR
Amazing new tasks !
.and Future
Computer Vision goes 3D
.and Future
3D will foster Robotics
.and Future
Linked/Augmented Life
Devices based on geo-localization,
inertial sensors and computer vision
may link seamlessly digital content to
physical objects
Concluding Remarks
Computer vision is a combination of signal processing, machine
learning and math (e.g. probability and statistics, linear algebra,
optimization theory, projective geometry, differential geometry of
curves and surfaces, PDEs)
Major trends:
Computer vision is key to Digital Manufacturing (aka Industry 4.0).
Computer Vision is more and more going 3D.
3D Computer Vision will foster Robotics,
Computer Vision will increasingly pervade our everyday lives.
Computer Vision will be more and more embedded into products.
Computer vision will enable novel Augmented Reality applications.
