Η ομιλία περιλαμβάνει εφαρμογές τεχνικών αναγνώρισης προτύπων και μηχανικής μάθησης σε ανάλυση πολυμέσων και κοινωνικών δικτύων. Πιο συγκεκριμένα, θα παρουσιαστούν τεχνικές και εφαρμογές κατάτμησης εικόνων με χρήση Κ-Μέσων και επεκτάσεων, χρήση Support Vector Ma-chines για μάθηση εννοιών σε εικόνες καθώς και τεχνικές ανάλυσης γράφων από κοινωνικά δίκτυα. Θα παρουσιαστούν σχετικές εφαρμογές που αξιοποιούν τα αποτελέσματα της ανάλυσης, όπως αναζήτηση πολυμέσων και εφαρμογή για τουρισμό και ενημέρωση από κοινωνικά δίκτυα. Θα αναφερθούν τρέχοντα ερευνητικά προβλήματα και περιοχές.
9. Αιγφξηζκνο Κ-Μέζσλ κε πεξηνξηζκφ
ζπλεθηηθφηεηαο
Component
Labeling
Παξάκεηξνη θαλνληθνπνίεζεο
l1
l2 p - Sk
Dk (p) = 2 I(p) - Ik + 2 A
sI
sS
Ak
Φωπικό κένηπο
IL
I L ,k
2
Ia
I a ,k
2
Ib
I b ,k
2
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
9
11. Τπνινγηζκφο Τθήο
• Δθαξκνγή θίιηξνπ
• Γηα ππνινγηζκφ πθήο: Discrete Wavelet Frames (DWF)
D p,sk
Jp
I p
I ss skk
J s
2
M
p S sk
Mk
1
T p
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
Ts sk
11
18. Δορυφορικά
τηλεπιςκοπικά
δεδομζνα MODIS από το
Dataset του Γνϊραςη
μετά την πυρκαγιά του
2007. Φυςικό ζγχρωμο
ςφνθετο 3-2-1
(αριςτερά) και
ψευδζγχρωμο 5-4-3
ςφνθετο (δεξιά).
Υπολογιςμόσ δεικτϊν
NDVI ςτα δορυφορικά
δεδομζνα πριν
(αριςτερά) και μετά
(δεξιά) την πυρκαγιά του
2007
19. Αποτελζςματα από την
εφαρμογή αλγορίθμων
κατάτμηςησ (αριςτερά) με
χρήςη προςαρμοςμζνου
κατωφλίου και αποτζλεςμα
τησ μορφολογικήσ
επεξεργαςίασ
(μορφολογικό opening με
δομικό ςτοιχείο μεγζθουσ
ενόσ εικονοςτοιχείου,
δεξιά).
20. Κατάτμηςη τηλεπιςκοπικϊν
δεδομζνων TERASAR-X με τη μζθοδο
των επιπεδοςυνόλων για την
ανίχνευςη του υδρογραφικοφ
δικτφου
(Συνεργαςία με ΕΜΠ, η
ανάπτυξη αλγορίθμων κατάτμηςησ
ζγινε από το EΜΠ, εργαςτήριο
τηλεπιςκόπιςησ, ςτα πλαίςια του
ζργου ΓΝΩΡΑΗ, www.gnorasi.gr)
33. Forensic Image Retrieval for Nudity Detection
• Police needs to search in unknown hard-drives to identify suspicious digital content
• Workflow to satisfy the requirements
Folder scanning and Image/video identification
Automatic indexing and feature/metadata extraction
Search Engine, Results presentation and Report generation
34. Indexing and Search
Image
Dataset
Query by visual example
•MPEG-7 descriptors
•R-tree structure for indexing
Visual Feature extraction
•MPEG-7 color & texture descriptors
•Dense SIFT GRAY & Dense SIFT RGB
Nudity Concept detection based on SVM classification
New
Image
Nudity
detection model
Probability of
Nudity
Search Engine with results and reports
Visual Features
(MPEG-7, SIFT)
Support Vector
Machines
Training Image
Dataset
37. Diadohokinetic (DDK) Test
DDK tests are used by speech-language pathologists for assessment of
motor speech impairments, e.g. dysarthria
Quick and accurate production of rapid, alternating sound tokens
involving different parts of mouse, e.g. “puh-tuk-kuh—puh-tuh-kuh—
…”
Usually measurements are done manually and include the DDK rate statistics
Our hypothesis and research direction:
The DDK test may challenge both motor and cognitive control over speech
production
Motor and cognitive faults deteriorate the temporal regularity of the audio
signal normally expected in this type of utterances
We can develop a regularity measure of the DDK performance
Is the DDK regularity useful in distinguishing between Control/MCI/AD
groups?
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
44
41. DDK Regularity Mapping for all the
Participants
DDK Regularity of Group Participants
0.045
0.04
Statistically significant
inter-group differences
with T-test pvalues:
DDK Regularity
0.035
Control vs. MCI
p<0.05%
MCI
0.025
p<1.45%
Control vs. AD
0.03
p<1.65%
vs. AD
0.02
0.015
Control
MCI
AD
0.01
Participants
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
49
44. Vertex & edges indicate degrees
• Degree of a graph vertex v: the
number of graph edges which
touch v.
• Indegree of a graph vertex v: the
number of inward directed graph
edges from a given graph vertex
in a directed graph
• Outdegree of a graph vertex v:
The number of outward directed
graph edges from a given graph
vertex in a directed graph.
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
undirected
directed
52
45. Degrees & adjancencies
Adjacency matrix on an undirected graph : A(i,j), i,j <= n
v1
v2
v3
v5
v4
degree of a vertex v
(number of edges incident upon it):
kv
A(v, w)
w
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
53
46. Μεξηθά παξαδείγκαηα
Webpage www.x.com
href=“www.y.com”
href = “www.z.com”
x
y
a
z
b
Webpage www.y.com
href=“www.x.com”
href = “www.a.com”
href = “www.b.com”
Webpage www.z.com
href=“www.a.com”
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
54
49. blogosphere as a graph
technical - gadgets
nodes = blogs
edges = hyperlinks
society - poli cs
h p://datamining.typepad.com/gallery/blog-map-gallery.html
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
9
57
50. social web as a graph
announcement of Mubarak’s resigna on
nodes = twi er users
edges = retweets on #jan25 hashtag
h p://gephi.org/2011/the-egyp an-revolu on-on-twi er/
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
10
58
52. Αλάιπζε Τπν – Γξάθσλ / Τπν - Γνκψλ (subgraphs)
k=3 (triangle)
k=4
k=5
• k-clique
Each node is
connected to all k-1
nodes
• N-clique
N=2 (star)
N is the length of
the path allowed to
all other members
2-core
• k-core
all vertices have
degree at least k
4-core
3-core
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
1-core
0-core
31
62
57. Challenges in Social Media network mining
No prior assumptions about structure:
Complex & evolving structure
No possibility for knowing structural features (e.g. number of clusters
on a graph) in advance
Unsupervised
Scale
Tens of millions of active users frequently contributing loads of
content links + metadata (tags, comments, ratings)
Efficient - scalable
Quality
Spam is very common. Only a portion of user contributions is worth
further analysis.
Noise resilient
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
67
60. Photo clustering results
Geographic localization of results was also found to be very high.
Most clusters correspond to landmarks or events.
baptism
EVENTS
LANDMARKS
conference
castels
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
70
70
61. Sample results:
[Visual] vs. [Tag] vs. [Visual + Tag]
VISUAL
HYBRID
TAG
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
71
71
62. clusttour.gr application
PHOTOS & METADATA
tags: sagrada familia,
cathedral, barcelona
SPATIAL CLUSTERING + TEMPORAL ANALYSIS
taken: 12 May 2009
lat: 41.4036, lon: 2.1743
CLASSIFICATION TO LANDMARKS/EVENTS
COMMUNITY DETECTION
VISUAL
TAG
HYBRID
S. Papadopoulos, C. Zigkolis, Y. Kompatsiaris, A. Vakali. “Cluster-based Landmark and Event Detection on Tagged Photo
Collections”. In IEEE Multimedia Magazine 18(1), pp. 52-63, 2011
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
72
65. Η Θεςςαλονίκη μέςα από το ClustTour
“Everything is automatic and when we say everything we mean everything”
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
75
72. πκπεξάζκαηα – Πεξηνρέο Πξνβιήκαηα
• Αλαγλψξηζε πξνηχπσλ
• Δπξεία ρξήζε ηερληθψλ αλαγλψξηζεο πξνηχπσλ ζε
εθαξκνγέο αλάιπζεο πνιπκέζσλ θαη θνηλσληθψλ
δηθηχσλ
• Απαηηνχληαη εμεηδηθεπκέλεο ιχζεηο θαη ζπλδπαζκφο
ηερληθψλ
• Η απηφκαηε επεμεξγαζία είλαη απαηηεηηθφ πξφβιεκα
• Πξνβιήκαηα - Δθαξκνγέο
• Μεγάινο φγθνο δεδνκέλσλ – Big Data - Social Media Data Mining – real–time – scalability
• Semantic Gap – Visual Similarity
• eHealth – Personalized Health
• Security – Forensics
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
83
73. Improve My City
App for Citizens Reporting Issues in Municipalities – Regions
Currently used by Thermi Municipality, Thessaloniki, Greece
http://mklab.iti.gr/imc/
74. Why do we need an app for that?
• Municipalities and Regions cover large areas
– Issues collection, prioritization and addressing is time
consuming and costly
– Regional authorities want to listen (or at least show they are
listening) what their citizens are saying
• Citizens want to participate
– Social Networks create new culture and technical possibilities
– They are interested in their everyday problems (especially if
they receive timely feedback)
• Citizens can become a cost and time efficient real-time
sensor of issues and the best source for solution ideas
and prioritization (Collective Intelligence)
75. Features at a Glance
• Login – authentication
– Synchronization with web
version
• Map and list-based view of
issues
– Close to current location
• Customized filters per
category and distance
• New issue
– Location + image
• Issue comment and voting
• Greek and English
77. Features
List View of
issues with
current progress
Local cache
(SQLite)
Filter by distance
from current
position (km)
Efficient
bandwidth usage
Filter by Category
78. Features
Full integration
with web-based
app, backend and
database
Easily customizable
for other
municipalities –
regions – cases
Authorities must
provide feedback
and adapt their
processes
Web based app and backend developed by URENIO Research Unit, Aristotle University of
Thessaloniki
79. Future Plans
• Open source distribution (under discussion)
• Augmented Reality Visualization for future and on-going
projects
– Library for Android, 3D objects support, OpenGL
• Automatic image (and issue – topic - category) annotation
using visual-based processing
80. Multimedia Group
http://mklab.iti.gr
Information Technologies Institute
http://www.iti.gr
Urban and Regional Innovation Research
http://www.urenio.org
Aristotle University of Thessaloniki
Smart City Services
Municipality of Thermi - Greece
https://smartcity.thermi.gov.gr/
“Google play„ search: Multimedia Group - CERTH-ITI,
http://mklab.iti.gr/imc
86. CERTH-ITI Multimedia Group
• Personnel
•
25 people (researchers, developers, administration)
• Participation in European and national research projects:
•
•
FP7: SocialSensor (coordination), Dem@Care
(coordination), WeKnowIt
(coordination), Pescado, JUMAS, CHORUS+, etc
FP6: AceMedia, X-Media, MESH, BOEMIE, VIDI-Video,
K-Space, PATExpert, ELU, etc.
• Contracts with Industry (Motorola US)
• Local collaborations (Thessaloniki Film Festival, Thermi
Municipality, Alzheimer Society, Police, TEDx)
• 55 Journal publications, 170+ conference publications, 30 book
chapters, 7 patents (2012)
• Numerous events: SSMS212, EVENT2010, ACM
CIVR09, WWW09 tutorial, WIAMIS 2007, etc.
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
97
87. How Tim Berners Lee told me in front of
thousand people: “Go geek and do it”
“You know it! Think of a world that you want. Just imagine it!
Data - Users
• What would be the distribution?
Demos - Apps
• what would be the user interface?
• What would be the processes?
• What third parties would be involved.
Programming
Implementation
Forums – Social Networks - Teamwork
Go out and build it! Talk to the people here.
Install an apache server and just go geek
and make it happen!”
Libraries –
Frameworks
http://www.rene-pickhardt.de/how-tim-berners-lee-told-me-in-front-of-thousand-people-%E2%80%9Cgo-geekand-do-it%E2%80%9D/
Oκάδα Πνιπκέζσλ
ΔΚΔΣΑ - Ιλζηηηνχην Σερλνινγηψλ Πιεξνθνξηθήο θαη Δπηθνηλσληψλ
98