Samia Bouchafa

1

• Samia Bouchafa
• Bertrand Zavidovique

Vito Di Gesù Cesare Valenti
IEF University of Orsay France
DMA University of Palermo Italy
2
Symmetry and perception
3
(No Transcript)
4
Computing Symmetry
5
Edge Based Computation
Symmetry Axial Transform (SAT) (Blum, Nagel,
1978)
Smoothed Local Symmetry (SLS) (Brady, Asada,
1984)
Affine transformations and symmetry (Mukhergee,
Zisserman, Brady, Chan, Cipolla, 1995)
Partial occlusion (Sato, Cipolla, 1997)
String oriented approach (Atallah, 1985),
(Bruckstein and Shaked, 1995)
6
Gray Levels Approaches
Texture analysis and symmetry measures
(Cheterikov and Haralick, 1995)
Measures based on the Radoms transform (Kiryati
and Gofman, 1996)
Context free attentional operators(Reisfeld,
Wolfson and Yeshurun 1995) 
7
Symmetry TransformDi Gesù, Valenti, 1994
8
Discrete Symmetry Transform
9
Points of interest
10
Pyramid-DST(Di Gesù,Valenti 1996)
Discrete Fourier Transform of D0 and
then
11
(No Transcript)
12
(No Transcript)
13
Tracking problems
14
Face analysis
Applicationssecurity systems, criminology.
physical access control, man-machine interactions
15
Expression analysis
Neutral, Sadness, Disgust, Happiness, Fear,
Anger, Surprise
16
Object recognition systems Chella, Di Gesu,
Infantino, Intravaia, Valenti 1997

• Object Recognition Using Multiple Views
• 3D shape reconstruction from image sequences

17
Iterated Object TransformDi Gesù, Zavidovique,
2002
The IOT computes the symmetry transform, T, on
steadily intensity reduced versions of the input
image
18
Contrast change and level lines

• Contrast change definition
• Non-decreasing funtion g
• Level set
• Contrast change impact
• some level sets disappearance
• no geometric deformation
• Motion impact ( noise)
• some new level sets appearance
• Geometric deformation
• level lines crossing

19
Detection criteria

• How can we reconstruct the scene S ?
• Possibilities for each line
• 1. The line is present
• no detection
• 2. The line is not present
• Doubt
• Is the reference complete ?
• Is the background uniform ?
• 3. The line crosses another one
• detection

Week Detection
Strong Detection
20
Motion detection algorithm
21
Level line characterization

• Two possibilities

Global characterization surface, other moments
of inertia, etc.
Associated level line
Our choice local characterization
- Point detection - No level lines occlusions
management
22
The result of the detection algorithm that is
insensible towards contrast changes.
The original sequence presents some contrast
changes due the automatic gain control of the
camera and to natural scene illumination changes.
In the sequence, only points affected by motion
are displayed
23
(No Transcript)
24
Fast Marching Methods and Level Set Methods are
numerical techniques which can follow the
evolution of interfaces. These interfaces can
develop sharp corners, break apart, and merge
together. The techniques have a wide range of
applications, including problems in fluid
mechanics, combustion, manufacturing of computer
chips, computer animation, image processing,
structure of snowflakes, and the shape of soap
bubbles. These are two fundamentally different
approaches to the problem of tracking moving
interfaces, yet they share a common theory and
numerical methodology.
25
Edge Based Computation
Symmetry Axial Transform (SAT) (Blum, Nagel,
1978)
26
Smoothed Local Symmetry (SLS) (Brady, Asada,
1984)
27
DST
Input
Edge based operator Yeshurun
28
Face analysis and algorithmsCardaci, Di Gesu,
Intravaia, 1998

• The algorithm is based on an attentive
  architecture.
• local and global symmetry operators
• Reisfeld, Wolfson,Yeshurun (1995) Di Gesù,
  Valenti, Strinati, (1997)
• graph theoretical algorithms Zhan (1972)
• facial anatomy (model driven) Russel, (1994)

29
(No Transcript)
30
Gelstat clustering (GC)
A relational graph (FG) is then built from the
retrieved FC
Structural information are represented by a
simple Internal Model (IM) based on psycho-visual
correlation between components of face Chen,
Yachida (1996)

31
Results
A sequence with global contrast changes
Séquence initiale
32
Results
The same crossing junction but different lighting
conditions
33
Applications
Road environment Vehicle/pedestrian detection
and counting Subway environment Stationnary
objects/human detection
34
Comparisons
Level lines
Reference sequence
Six months before
Grey levels
Gradients orientation
Laplacian
35
Comparisons
Gradient orientations
Problems with stability and thresholding !