Professional Documents
Culture Documents
P t
ro in VeDigital Signal Processing
n a l rt
i g a
i g i tal S i and M
D d on 013 Module 10: The End!
Pra n
©2
Paolo
Module Overview:
g
ssinfrom hererli
roce in Vette
◮ Module 10.1: Signal processing courses you can take
a l P
◮ Module 10.2: Some research projectsigthat
l S n useM art
techniques you learned here
i t a n d
Module 10.3: ig
Examples ofDstart-ups
d ni ause signal
othat 13 processing as a core technology
n
◮
r a 2 0
lo P ©
Module 10.4:
Pao
Acknowledgements
◮
10 1
c e s sing terli
P t
ro in VeDigital Signal Processing
n a l rt
Si g a
i g i tal 10.1:i a n d Mprocessing courses you can take from here
D
Module
don 2013
Signal
r a n
Paolo P ©
Overview:
lo P
Pao
Advanced Topics in Signal Processing
◮
10.1 2
Overview:
lo P
Pao
Advanced Topics in Signal Processing
◮
10.1 2
Overview:
lo P
Pao
Advanced Topics in Signal Processing
◮
10.1 2
Overview:
lo P
Pao
Advanced Topics in Signal Processing
◮
10.1 2
Statistical Signal and Processing through Applications 1/3
D i g i
don 2013
models and statistical techniques
P r a n
©
aolo to address?
What are the problems wePwant
◮ Spectral estimation for Wireless Transmissions, Biomedical Signals, Audio Signals
◮ Classification of Neurobiological Spikes
◮ Adaptive filtering for echo cancellation in Voice-Over-IP transmissions
10.1 3
Statistical Signal Processing 2/3
An exemplary application:
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.1 4
Statistical Signal Processing 3/3
c e s sing terli
Statistical Signal Processing Tools for
ro in Vet
◮
a l P
• Wireless Transmissions
l S i gn M art
ta d
• Echo Cancellation Digi doni an 13
Neurobiological Spikes lo Pra
n 20
©
Pao
•
10.1 5
Signal Processing for Audio and Acoustics 1/3
10.1 6
Signal Processing for Audio and Acoustics 1/3
10.1 6
Signal Processing for Audio and Acoustics 1/3
10.1 6
Signal Processing for Audio and Acoustics 1/3
10.1 6
Signal Processing for Audio and Acoustics 1/3
10.1 6
Signal Processing for Audio and Acoustics 2/3
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
10.1 7
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Signal Processing for Audio and Acoustics 3/3
• Audio Recording
c e s sing terli
a l P ro in Vet
• Multichannel Audio Reproduction
l S i gn M art
ta d
• Spatial Filtering Digi doni an 13
Coding of Audio Signalslo Pr
an © 20
Pao
•
• Auralization
10.1 8
Mathematical Foundations of Signal Processing
◮ The world is analog but computation is digital
◮ How to go between these representations?
g
Processingessin li
Analog World
r o c et t erAnalog World
l P V
l S i gna Martin
ta and simulations etc
D i donigraphics,
Ex: Audio, sensing, imaging,igcomputer
Key mathematical concepts: Pran 2 013
◮
lo ©
Pao
• Sampling and interpolation
10.1 9
From Analog to Digital
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
10.1 10
Example: sensor networks for environmental monitoring
◮ How many sensors?
◮ How to reconstruct?
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
10.1 11
Course Outline
◮ Brief outline:
1. From Euclid to Hilbert
2. Sequences and Discrete-Time Systems
3. Functions and Continuous-Time Systems
c e s sing terli
a l P ro in Vet
i gn
4. Sampling, Interpolation and Approximation
l S M art
ta d
5. Applications Digi doni an 13
lo P ran © 20
Pao
◮ Details on course website
http://lcav.epfl.ch/SP_Foundations
http://moodle.epfl.ch/course/view.php?id=13431
◮ Textbook: M. Vetterli, J. Kovacevic and V. Goyal, “Foundations of Signal Processing”,
Cambridge U. Press, 2013. Available in open access at
http://www.fourierandwavelets.org
10.1 12
Advanced Topics in Signal Processing 1/3
◮
c e s
Models and methods used for describing and processingsingsignals
t erli
arising from physical
P
phenomena constantly increase in complexityr o V e t
l
igna artin
◮
i tal S anymore,
Classical approaches are not sufficient
i g i a nd Msince they either cannot be applied, or
D do n 13
ran © 20
they lead to wrong results or to labyrinthian computations
lo P
◮
Pao to optimally model and analyse the data
Advanced tools are required
10.1 13
Advanced Topics in Signal Processing 2/3
Compressed Sensing Example: Expansion with respect to an overcomplete set of vectors is not
unique. One can optimize various norms (l1 , l2 , l∞ )
a1
3 dia
c e s sing terli
a l P rosqu in Vet
cir
l S
2
i gn M art
ta d
Digi d1 oni an 13
an © 20
apr
P r
P aolo a0
-3 -1 1 3 5
-1
-2
-3
10.1 14
Advanced Topics in Signal Processing 3/3
c e s sing terli
ro in Vet
◮ Mathematical principles of signal processing
a l P
l S i gn
◮ Selected current topics in signal processing
M art
ta d
d o ni anfor1sensor
Digi processing
(sparse signal processing, signal
3 networks, etc.)
Pra n 0
o lo ©2
a
P class
Website and textbook of the
◮ http://lcav.epfl.ch/teaching/advanced sp
◮ Volume 2 of http://www.fourierandwavelets.org
10.1 15
c e s sing terli
END OF MODULE 10.1 a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
c e s sing terli
P t
ro in VeDigital Signal Processing
n a l rt
i g Ma
i g i tal Sresearch
i a n dprojects
D
Module 10.2: Some
d on 013 that use techniques you learned here
a n
o lo Pr ©2
P a
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
Overview:
◮ eFacsimile or art work acquisition, representation and rendering
◮ Signal processing for sensors networks
sing terli
◮ Source localization in graphs
c e s
Finite rate of innovation sampling
a l P ro in Vet
art
◮
l S i gn M
ta d
Digi doni an 13
◮ Sampling of fields
◮
P ran binary
Image acquisition using oversampled
lo ©
0
2pixels
◮ Pao
Predicting the stock market
◮ Inversion of the diffusion equation
◮ The Fukushima inverse problem
◮ Can you hear the shape of a room
10.2 16
eFacsimile 1/3
People involved:
◮ Joint project between LCAV and Google Inc.
◮ Dr. Loı̈c Baboulaz
c e s sing terli
a l P ro in Vet
◮ Ms. Mitra Fatemi
l S i gn M art
ta d
◮ Digi doni an 13
Mr. Niranjan Thanikachalam
lo P ran © 20
Pao
◮ Mr. Zhou Xue
What are the problems addressed?
◮ How to capture, represent and render an Artwork faithfully?
◮ How to achieve relighting, 3D manipulation and high-resolution on mobile devices?
10.2 17
eFacsimile 2/3
Demo: interactive relighting and manipulation of Oil Paintings and Stained Glasses
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
10.2 18
eFacsimile 3/3
sing terli
◮ Traditional and light-field cameras
c e s
a l P ro in Vet
art
◮ Light transport theory
l S i gn M
ta d
◮ Sparse recovery methods
Digi doni an 13
Websites
lo P ran © 20
Pao
◮ http://lcav.epfl.ch/cms/site/lcav/lang/en/efacsimile
◮ http://www.youtube.com/user/lcavepfl
10.2 19
Visual Monitoring using Wireless Sensor Networks
People involved:
◮ Zichong Chen, Guillermo Barrenetxea, Martin Vetterli
Video monitoring with autonomous wireless camera nodes. Challenges include:
◮ Each node has very limited energy (e.g., 100 mW from solar panel)
ces(e.g.,V900 r sing li
◮
r o
Need to capture, process and transmit 3-D signals
P ettekilobytes per image), which
a l t i n
ign
requires far more energy than the budget ar
i g i tal S i and M
D don 2013
NO2 sensor CMOS image sensor Light-field camera
r a n
Sensors
lo P ©
Trends in wireless sensor
networko(WSN):
Pa
physical quantity measured by sensors
1-D signal
I (t )
3-D signal
I ( x, y , t )
4-D signal
I ( x, y , z , t )
WSN nodes
videos and even beyond
Tools for better representing and understanding image signals: image compression (low
level),
10.2 computer vision (high level) 20
Visual Monitoring using Wireless Sensor Networks
◮ Many real-world applications by using wireless cameras
Monitor wild animals
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
Weather
P r an © 20 Avalanche
inference lo
detection
P a o
Wireless monitoring
applications
10.2 21
Visual Monitoring using Wireless Sensor Networks
Results so far
◮ N cameras can collaboratively share the monitoring task and reduce energy per node by a
g
ssin rli
roce in Vette
factor of N
a l P
i
Understanding video content via computer
◮
S gn visionM rt
atechniques improves video compression
i t al n d
Dig doni a 13
efficiency by 70% over conventional H.264 codec
Further references
lo P ran © 20
◮ Thesis on Cooperative a o Monitoring in Energy-Constrained Wireless Sensor
P Visual
Networks: http://infoscience.epfl.ch/record/183072
◮ Supported by the grant from NCCR-MICS, http://www.mics.org
10.2 22
Source localization
People involved: Dr Pedro Pinto, Prof Patrick Thiran
◮ Goal: Locate the source of diffusion in a network with thousands of nodes
c e s sing tterli
ro in Vet
3
a l P
gn art
observer
l S i d M
ta
Digi doni an 13
ran © 20
t4
o lo P
source
Pa
t1
observer
t2
10.2 23
Source localization
Our contributions:
Future directions:
◮ Developed optimal and efficient
algorithms that exploit ◮
s s i ng placement
Optimal
r l i
of observers.
time-of-arrivals. roc◮e Disease te and public health.
Vetcontrol
n a l P t i n
Achieve 90% accuracy with al Sig
i t n d M◮arSocial media analytics.
oni a 013
◮
ig
typically less than 20%Dobservers.
a n d
o lo Pr ©2
P a
◮ ”Locating the Source of Diffusion in Large-Scale Networks”, P. Pinto, P. Thiran, M.
Vetterli, Physical Review Letters, Aug. 2012.
◮ Project funded by the Bill and Melinda Gates Foundation.
10.2 24
Applications
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.2 25
Finite Rate of Innovation (FRI)
People involved:
◮ Dr. Nikolaos Freris, Senior Scientist ◮ Ehsaneddin Asgari, M.Sc. Assistant
r o c V
What is FRI?
S i g nal P Martin
l
◮ Sampling and reconstruction of
D i g itasignals anda
with
onofi time)
finite number of parameters (pernunit
P r a d 2 013
olo of Dirac pulses,
◮ Typical signals: periodicastreams
©
P
periodic non-uniform splines
DSP Tools
◮ Fast Fourier Transform ◮ Annihilating filter (root finding for complex
exponentials)
◮ Singular Value Decomposition
10.2 26
ECG Modeling
amplitude
sing terli
1
waveforms
c e s
a l P ro in Vet 0
◮ Model signals as a sum of asymmetric
l S i gn M art
ta d
Digi doni an 13
þ 1
pulses 0 1
time (sec)
0 1
time (sec)
0 1
time (sec)
lo P ran © 20
Industrial Collaborator Pao Advantages
◮ Compression
◮ Noise removal
10.2 27
Publications
◮ M. Vetterli, P. Marziliano, and T. Blu. Sampling Signals With Finite Rate of Innovation.
IEEE Transactions on Signal Processing, vol. 50, no. 6, June 2002.
◮ G. Baechler, N. Freris, R.F.Quick, and R.E. Crochiere. Finite rate of innovation based
ing
modeling and compression of ECG signals. ICASSP s2013.
r li
oces ette
l Pr ofrtMultiple
◮
i g a
Hormati, A., Vetterli, M., Compressive nSampling
a in V Sparse Signals Having
S nd M Principles, Signal Processing Letters,
al of iInnovation
Common Support Using Finite tRate
Digi May
do n a
3
ran © 201
IEEE, vol.18, no.5, pp.331,334, 2011.
lo P
Pao Fast and Robust Parametric Estimation of Jointly Sparse
Y. Barbotin and M. Vetterli.
◮
Channels, in Journal on Emerging and Selected Topics in Circuits and Systems, IEEE, vol.
PP, num. 99, p. 1-11, 2012.
◮ A. Hormati and M. Vetterli. Distributed Compressed Sensing: Sparsity Models and
Reconstruction Algorithms Using Annihilating Filter. ICASSP, Las Vegas, Nevada, USA,
March 30-April 4, 2008.
10.2 28
Sampling spatial fields using mobile sensors 1/3
People involved:
◮ Jayakrishnan Unnikrishnan
◮ Martin Vetterli
c e s sing terli
Research challenges a l P ro in Vet
S i gnfield using
M rt
adiscrete
l
igita oni and 3
◮ Algorithms for reconstructing a spatial measurements taken by sensors
D
◮ Sampling spatial fields with mobile
P r andsensors
© 201
lo
Pao trajectories that minimize the total distance traveled by the sensors
• Design of efficient sensor
10.2 29
Sampling spatial fields using mobile sensors 2/3
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.2 30
Sampling spatial fields using mobile sensors 3/3
Insights from research
◮ Sampling along equispaced parallel lines (see figure) is
optimal over a wide range of configurations of sensor
trajectories for 2D low-pass fields
n g
r o c essi etterli y
mobile V
nal P Martin
Spatial anti-aliasing can be induced by using
◮
l S i g x
sensors with time-domain filtering
i t a n d
Dig doni a 13
lo P ran © 20
Key papers available at http://lcav.epfl.ch/people/jayakrishnan.unnikrishnan
◮ J. Unnikrishnan and M.
PaoVetterli, “Sampling High-Dimensional Bandlimited Fields on
Low-Dimensional Manifolds” accepted for publication in IEEE Transactions on
Information Theory, Oct 2012.
◮ J. Unnikrishnan and M. Vetterli, “Sampling and Reconstruction of Spatial Fields using
Mobile Sensors” accepted for publication in IEEE Transactions on Signal Processing, Feb
2013.
10.2 31
Image acquisition using oversampled binary pixels 1/3
People involved:
◮ Feng Yang, Luciano Sbaiz (Google Zürich), Yue M. Lu (Harvard University), Sabine
Süsstrunk, Hyungjune Yoon, Edoardo Charbon, Martin Vetterli
sing terli
What are the problems addressed?
c e s t
◮ Propose an image sensor with binary pixels,Pused
a l ro forinhigh
Vedynamic range imaging
i g n a rt
tal S i and M
sm
D i g i s2
don 2013
λ (x)
0 s0 s1
n
Light intensity
r a
lo P ©
Lens m
Pao
Photon counting y0 y1 y2 ym
Binary
λ(x) quantization
(q = 2)
Sensor b0 b1 b2 bm
Binary
measurements
c e s sing Pixel
t i
erlvalue: binary
P r o Vet
S i g nal Martin Resolution: 32 × 32photon
Sensitivity: single
tal d
Digi doni an 13 Temporal oversampling to
P r a n 2 0
©
aolo
emulate spatial oversampling
P
Reconstruction from 4096 consecutive frames
Spatial: 64 × 64
Temporal: 64
10.2 33
Image acquisition using oversampled binary pixels 3/3
Main publications
◮ Feng Yang, Yue M. Lu, Luciano Sbaiz, and Martin Vetterli. Bits from photons:
Oversampled image acquisition using binary Poisson statistics. IEEE Transactions on
ng
essi etterli
Image Processing, vol. 21, issue 4, pp. 1421-1436, 2012.
r o c V sensor. Proc. of IEEE Int.
n P Processing,
Feng Yang, Martin Vetterli. Oversampledalnoisy
tin image
binary
r
◮
Conference on Acoustics, Speechaland S igSignal dM
a Vancouver, May 26-31, 2013.
i g i t i a n
D n
doCharbon, 3 Süsstrunk, and Martin Vetterli. The
◮
P r
Luciano Sbaiz, Feng Yang, Edoardo a n 2 01Sabine
©
aolo of IEEE Int. Conference on Acoustics, Speech and Signal
gigavision camera. In Proc.
P
Processing, pp.1093-1096, Taipei, Apr. 2009.
Website and other pointers
◮ Feng Yang’s PhD thesis http://infoscience.epfl.ch/record/174695?ln=en
◮ http://lcav.epfl.ch/research/gigavision
10.2 34
Predicting the stock market
People involved:
◮ Lionel Coulot
Peter Bossaerts
◮
g
sinand erli
What are the problems addressed?
o c e s t t
l P r
◮ prediction means identifying features in observed data
n V e extrapolating them in the
future in the hope they persist ign a ar t i
i g i t al S i and M
◮ D
features expressed using statistical on that
dmodels 13account for the inherent randomness in
P r a n 2 0
data ©
P aolo
◮ how do we decide between a simple and a more complex model?
on which portion of the observed data do we learn the model?
◮
What tools used?
◮ coding theory
◮ dynamic programming
10.2 35
Predicting the stock market
An exemplary application: dynamic trend line
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.2 36
Predicting the stock market
Results so far
sing terli
◮ project completed
c e s
Website and other pointers
a l P ro in Vet
◮ PhD thesis
l S i gn M art
ta d
Digi doni an 13
ran © 20
◮ http://infoscience.epfl.ch/record/182936
lo P
◮ Sponsor Pao
◮ http://www.lgt.com
10.2 37
Inversion of the diffusion equation 1/3
People involved:
◮ Juri Ranieri, Ivan Dokmanić, Amina Chebira, Yue Lu, Martin Vetterli
ta l S d M
◮
Digi locations
We measure the field at certain
d o ni ainn space
1 3 and time
a n 2 0
◮
o lo Pofr the field©
Target: estimate the sources
a
P
◮ Variations: advective diffusive phenomena, time varying sources, . . .
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.2 39
Inversion of the diffusion equation 3/3
What we have shown so far
◮ Recovery spatially sparse sources appearing at unknown time
◮ Recovery of time-varying emissions with FRI
sing li
Recovery of the field using traditional sampling es
roctechnique
V tter
eexploiting the low-pass kernel
nal P Martin
◮
S i g
tal i anSampling
Main Publications
d
DigiM. Vetterli.
◮ J. Ranieri, Y. Lu, A. Chebira and
d o n 1 3 and Reconstructing Diffusion Fields with
a n 0
o Pr ©2
Localized Sources
a o l
P A. Chebira and M. Vetterli. Sensor Networks for Diffusion Fields:
◮ I. Dokmanic, J. Ranieri,
Detection of Sources in Space and Time
◮ J. Ranieri, I. Dokmanic, A. Chebira and M. Vetterli. Sampling and Reconstruction of
Time-Varying Atmospheric Emissions
◮ J. Ranieri and M. Vetterli. Sampling and reconstructing diffusion fields in presence of aliasing
10.2 40
The Fukushima Inverse Problem 1/3
People involved:
◮ Marta Martinez-Camara, Ivan Dokmanic, Juri Ranieri, Robin Scheibler, Martin Vetterli.
sing terli
◮ Andreas Stohl, from NILU, Norway.
c e s
a l P ro in Vet
What are the problems addressed?
l S i gn M art
ta radionucleidesd
Digi dof
◮ We want to estimate the emissions
o ni an 13 during a nuclear accident.
P n
raaround 0
2world.
lo ©
Pao
◮ Using a few sensors located all the
10.2 41
The Fukushima Inverse Problem 2/3
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
10.2 42
The Fukushima Inverse Problem 3/3
g
was released in the first 5 days afterSithe a
n earthquake.
a rt
i g i tal i and M
D don 2013
r a n
lo P ©
Pao
Website and other pointers
◮ http://infoscience.epfl.ch/record/182697
◮ M.Martinez-Camara, I.Dokmanic, J.Ranieri, R.Scheibler, M. Vetterli, A.Stohl, ”The
Fukushima Inverse Problem,” Proc IEEE ICASSP, May 2013.
10.2 43
Can One Hear the Shape of a Room?
◮ People: Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Martin Vetterli and Yue M. Lu
(Harvard School of Engineering) g n
r o c essi etterli
◮ Remember 5.12: Dereverberation and Echo
i g a rtin V
nal PCancelation?
S
tal i rid d M echoes
◮
Digi candoget
If you know “room shape”, you
n anof1the
P r anwe know 2 0 3
lo
If you know the shape... but can © it?
Pao
◮
◮ Imagine you’re blindfolded inside an unknown room and you snap your fingers. Can you
hear the shape of the room?
10.2 44
Can One Hear the Shape of a Room?
◮ People: Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Martin Vetterli and Yue M. Lu
(Harvard School of Engineering) g n
r o c essi etterli
◮ Remember 5.12: Dereverberation and Echo
i g a rtin V
nal PCancelation?
S
tal i rid d M echoes
◮
Digi candoget
If you know “room shape”, you
n anof1the
P r anwe know 2 0 3
lo
If you know the shape... but can © it?
Pao
◮
◮ Imagine you’re blindfolded inside an unknown room and you snap your fingers. Can you
hear the shape of the room?
10.2 44
Can One Hear the Shape of a Room?
◮ People: Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Martin Vetterli and Yue M. Lu
(Harvard School of Engineering) g n
r o c essi etterli
◮ Remember 5.12: Dereverberation and Echo
i g a rtin V
nal PCancelation?
S
tal i rid d M echoes
◮
Digi candoget
If you know “room shape”, you
n anof1the
P r anwe know 2 0 3
lo
If you know the shape... but can © it?
Pao
◮
◮ Imagine you’re blindfolded inside an unknown room and you snap your fingers. Can you
hear the shape of the room?
10.2 44
Can One Hear the Shape of a Room?
◮ People: Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Martin Vetterli and Yue M. Lu
(Harvard School of Engineering) g n
r o c essi etterli
◮ Remember 5.12: Dereverberation and Echo
i g a rtin V
nal PCancelation?
S
tal i rid d M echoes
◮
Digi candoget
If you know “room shape”, you
n anof1the
P r anwe know 2 0 3
lo
If you know the shape... but can © it?
Pao
◮
◮ Imagine you’re blindfolded inside an unknown room and you snap your fingers. Can you
hear the shape of the room?
10.2 44
Can One Hear the Shape of a Room?
◮ People: Ivan Dokmanić, Reza Parhizkar, Andreas Walther, Martin Vetterli and Yue M. Lu
(Harvard School of Engineering) g n
r o c essi etterli
◮ Remember 5.12: Dereverberation and Echo
i g a rtin V
nal PCancelation?
S
tal i rid d M echoes
◮
Digi candoget
If you know “room shape”, you
n anof1the
P r anwe know 2 0 3
lo
If you know the shape... but can © it?
Pao
◮
◮ Imagine you’re blindfolded inside an unknown room and you snap your fingers. Can you
hear the shape of the room?
10.2 44
How Does One Hear the Room?
Wall i
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Wall j
Pao
◮ Aim: Reconstruct the shape of the room from acoustic reponses (as few as possible!)
◮ How many suffice? We can get a robust algorithm with 4!
10.2 45
How Does One Hear the Room?
Wall i
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Wall j
Pao
◮ Aim: Reconstruct the shape of the room from acoustic reponses (as few as possible!)
◮ How many suffice? We can get a robust algorithm with 4!
10.2 45
How Does One Hear the Room?
Wall i
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Wall j
Pao
◮ Aim: Reconstruct the shape of the room from acoustic reponses (as few as possible!)
◮ How many suffice? We can get a robust algorithm with 4!
10.2 45
Yes We Can!
D
sing terli
3 2
c e s t
A
n a l Pro rtin Ve E DD
i g a
90.1° 116.8° Height 89.7° 0.4
tal S i and M
79.1°
90.0° 118.0°
2.08 Exp. B 2.70m 90.0° 2.52 79.5°
Exp. C 2.71m
1 F
i
Real 2.69m
g
0.99
D i
1
don 2013
6.48m
7.01m 5 4 6.53m
7.08m 5
3
n
4 1 3
2
2
6.71m
3
a
2
6.76m
P r
4.55m
©
aolo
90.3°
4.59m
100.9°
L
89.9° 63.2° 90.0° 100.5°
P
90.0° 62.0°
B C E
10.2 46
Yes We Can!
D
sing terli
3 2
c e s t
A
n a l Pro rtin Ve E DD
i g a
90.1° 116.8° Height 89.7° 0.4
tal S i and M
79.1°
90.0° 118.0°
2.08 Exp. B 2.70m 90.0° 2.52 79.5°
Exp. C 2.71m
1 F
i
Real 2.69m
g
0.99
D i
1
don 2013
6.48m
7.01m 5 4 6.53m
7.08m 5
3
n
4 1 3
2
2
6.71m
3
a
2
6.76m
P r
4.55m
©
aolo
90.3°
4.59m
100.9°
L
89.9° 63.2° 90.0° 100.5°
P
90.0° 62.0°
B C E
10.2 46
Yes We Can!
D
sing terli
3 2
c e s t
A
n a l Pro rtin Ve E DD
i g a
90.1° 116.8° Height 89.7° 0.4
tal S i and M
79.1°
90.0° 118.0°
2.08 Exp. B 2.70m 90.0° 2.52 79.5°
Exp. C 2.71m
1 F
i
Real 2.69m
g
0.99
D i
1
don 2013
6.48m
7.01m 5 4 6.53m
7.08m 5
3
n
4 1 3
2
2
6.71m
3
a
2
6.76m
P r
4.55m
©
aolo
90.3°
4.59m
100.9°
L
89.9° 63.2° 90.0° 100.5°
P
90.0° 62.0°
B C E
10.2 46
c e s sing terli
END OF MODULE 10.2 a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
c e s sing terli
P t
ro in VeDigital Signal Processing
n a l rt
Si g a
i g i talstart-ups
i a n d Muse signal processing as a core technology
Module 10.3: D
Examples of
don 2013
that
r a n
lo P ©
Pao
Overview:
Illusonic
c e s sing terli
ro in Vet
◮
a l P
◮ Quividi
l S i gn M art
ta d
Sensorscope Digi doni an 13
ran © 20
◮
lo P
◮ Vidinoti
Pao
10.3 47
Overview:
Illusonic
c e s sing terli
ro in Vet
◮
a l P
◮ Quividi
l S i gn M art
ta d
Sensorscope Digi doni an 13
ran © 20
◮
lo P
◮ Vidinoti
Pao
10.3 47
Overview:
Illusonic
c e s sing terli
ro in Vet
◮
a l P
◮ Quividi
l S i gn M art
ta d
Sensorscope Digi doni an 13
ran © 20
◮
lo P
◮ Vidinoti
Pao
10.3 47
Overview:
Illusonic
c e s sing terli
ro in Vet
◮
a l P
◮ Quividi
l S i gn M art
ta d
Sensorscope Digi doni an 13
ran © 20
◮
lo P
◮ Vidinoti
Pao
10.3 47
Illusonic 1/3
10.3 48
Illusonic 2/3
An exemplary application: Sound in 3D
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.3 49
Illusonic 3/3
sing terli
◮ 5 employees
c e s
P ro in Vet
◮ Licensable technologies for pro, consumer,l voice
a
l S i gn M art
ta d
Digi doni an 13
◮ Custom technology adaptation/development
10.3 50
Quividi: automated audience measurement
Quividi provides qualified audience measurement using
video analytics and off-the-shelf hardware:
◮ real-time processing of video streams
◮ data relayed to main data warehouse
c e s sing terli
a l P ro in Vet
◮
l S i gn
customers access formatted data remotely
M art
ta d
Digi doni an 13
P an metrics
rMost © 20based on face detection/analysis:
lo
Pao ◮ number of viewers, attention time, dwell time...
10.3 51
Quividi: audience data dashboards
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
10.3 52
Quividi
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
http://www.quividi.com
10.3 53
Sensorscope
S i g a rtinanywhere/anytime
ital ni and
◮ Sensor data management: access/control/share
D i g
What tools used?
P r a ndo 2013
lo ©
Pao
◮ Wireless sensor networks (WSN)
10.3 54
Sensorscope 2/3
An exemplary application: environmental monitoring with sensor networks
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.3 55
Sensorscope 3/3
sing terli
◮ 5 employees
c e s
a P
◮ Main markets: environmental monitoring, lprecision
t i n Vet
ro agriculture
a ign d Mar
l Smodules
t
Dig doni an to13easily compose a WSN
i
◮ Main product: DS3 system, flexible
◮ http://www.climaps.com
10.3 56
Vidinoti 1/3
Vidinoti is active in the field of Augmented Reality on mobile devices. More specifically, the
main challenges are:
◮ Image recognition robust to environment changes (lighting, distortion, ...)
ng
i recognition
erli
◮
r o c
Time and memory constraints to perform real-timeessimageet t and tracking on
P V
mobile devices nal
Sig nd Ma rtin
i t al
◮ Load balancing between remote
Dig server, a
onicontaining a large amount of images to recognize,
n d 0 1 3
and mobile devices
o lo Pra ©2
a
P Vidinoti uses:
To tackle these challenges,
◮ Advanced image recognition algorithms
◮ Objects tracking taking advantage of sensors and image data
◮ Algorithm optimization tailored to mobile devices
10.3 57
Vidinoti 2/3
Vidinoti technology is cloud-based, providing a seamless integration of all Augmented Reality
components to customers.
c e s sing terli
a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
P ao
10.3 58
Vidinoti 3/3
Vidinoti started in 2010 as a spin-off of the LCAV, now in Fribourg (CH).
The company currently employs about 10 people having the missions of:
◮ Business development
◮ Research and development
c e s sing terli
a l P ro in Vet
◮ IP watch, investigation and creation
l S i gn M art
ta i andadditional
Digi dtoondeliver
Vidinoti objective is to allow customers multimedia content to their
end-users using Augmented Reality,ra innmarkets 2 0
such13as:
lo P ©
Pao Museums
◮ Marketing, Printed media,
For more information, please visit Vidinoti website and try the iPhone app, PixLive, using
Vidinoti technology:
◮ http://www.vidinoti.com
◮ http://www.pixlive.info
10.3 59
c e s sing terli
END OF MODULE 10.3 a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
c e s sing terli
P t
ro in VeDigital Signal Processing
n a l rt
i g a
i g i tal S i and M
D d on 013 Module 10.4: Acknowledgements
Pra n
©2
Paolo
Acknowledgements
Content reviewers:
Operations: ◮ Jay Unnikrishnan
◮in g
s Youssef rli Baba
◮ Pedro Pinto
o c e s t t eEl
r V e
nal P Mar◮tinTaher Nguira
◮ Juri Ranieri
S i g
tal d
Digi doni an 13
◮ Solene Buet
10.4 60
Acknowledgements
Homework Sets:
Juri Ranieri g Fatemi
sinMitra erli
◮
o c e s◮
t t
r V e
nal P Mar◮tinDirk Schroder
◮ Yann Barbotin
S i g
ital ni and
◮ Ivan Dokmanic
D i g o
◮ Andreas Walther
d 3
ran © 201 ◮ Alireza Ghasemi
◮ Zhou Xue
lo P
Pao
◮ Reza Parhizkar
◮ Marta Martinez-Camara
◮ Feng Yang
◮ Zichong Chen
◮ Mihailo Kolundzija
10.4 61
Acknowledgements
c e s sing terli
Administration: a l P ro Funding:
t i n Vet
◮ Jacqueline Aeberhardal Si
gn Mar◮ EPFL
i t n d
Dig doni a 13
lo P ran © 20
Pao
10.4 62
c e s sing terli
END OF MODULE 10.4 a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao
c e s sing terli
END OF MODULE 10 a l P ro in Vet
l S i gn M art
ta d
Digi doni an 13
lo P ran © 20
Pao