Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Chapter 0Organization and IntroductionNonlinear Multiscale Methods for Image and Signal AnalysisSS 2015
Michael MoellerComputer Vision
Department of Computer ScienceTU Munchen
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What do you hear?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
Let’s look at the sound signal ...
0 0.05 0.1 0.15 0.2−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1Soundwave of a guitar string
Time (sec)
Am
plitu
de
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
Much better: DCT coefficients!
0 200 400 600 800 1000 12000
5
10
15DCT coefficients of sound signal
Frequency (Hz)
Mag
nitu
de
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What do you see?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What do you see?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
Shouldn’t there be a ’spectral’ representation with three peaks?
0 5 10 15 20 25 30 35 40 45 500
500
1000
1500
2000
2500Spectrum of a TV decomposition
Time (ISS flow)
Am
plitu
de
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
Shouldn’t there be a ’spectral’ representation with three peaks?
0 5 10 15 20 25 30 35 40 45 500
500
1000
1500
2000
2500Spectrum of a TV decomposition
Time (ISS flow)
Am
plitu
de
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
Shouldn’t there be a ’spectral’ representation with three peaks?
0 5 10 15 20 25 30 35 40 45 500
500
1000
1500
2000
2500Spectrum of a TV decomposition
Time (ISS flow)
Am
plitu
de
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What could spectral representations be good for?
0 20 40 60 80 100 120 140 160 180 200−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
x
ampl
itude
Input signal
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What could spectral representations be good for?
0 20 40 60 80 100 120 140 160 180 200
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
frequency
mag
nitu
de
DCT of input signal
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What could spectral representations be good for?
0 20 40 60 80 100 120 140 160 180 200
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
frequency
mag
nitu
de
DCT of input signal
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What could spectral representations be good for?
0 20 40 60 80 100 120 140 160 180 200
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
frequency
mag
nitu
de
DCT of input signal
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What could spectral representations be good for?
0 20 40 60 80 100 120 140 160 180 200−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
x
ampl
itude
Filtered signal
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
What could spectral representations be good for?
0 20 40 60 80 100 120 140 160 180 200−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
x
ampl
itude
Input signal
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
MotivationLooks boring for cosine signals?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
MotivationLooks boring for cosine signals?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
MotivationLooks boring for cosine signals?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
Looks boring for cosine signals?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
Looks boring for cosine signals?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
GUI
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Motivation
How does this work?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Variational Approach
u(t) = arg minu
12‖u − f‖2
2 + t J(u)
φ(t) = t ∂ttu(t), ψ(t) = φ(1/t)1t2
Forward Flow
∂tu(t) = −p(t), p(t) ∈ ∂J(u(t))
φ(t) = t ∂ttu(t), ψ(t) = φ(1/t)1t2
Inverse Scale Space Flow
∂tp(t) = f − v(t), p(t) ∈ ∂J(v(t))
ψ(t) = ∂tv(t), φ(t) = ψ(1/t)1t2
Filtering
ufiltered =
∫ ∞
0ω(t)ψ(t) dt
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Organizational Stuff
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Requirements, or “is this something for me?”
Necessary
• Interest in mathematical theory• Image processing and convex analysis• Numerics (Matlab)
Nice to know
• Optimization• Partial differential equations
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Requirements, or “is this something for me?”
For those who heard Variational Methods for Computer Vision
• It will become more theoretical and more mathematicallychallenging.
• We will prove theorems on the board.• You’ll have an advantage in terms of possible applications
and their numerical implementation.
For those who heard Ill-posed Problems
• We will need very little functional analysis – everything willbe in Rn.
• You’ll have an advantage in terms of the theoreticalmathematical concepts.
• I highly recommend learning how to implement thediscussed approaches. It is useful and fun!
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Requirements, or “is this something for me?”
For those who heard Variational Methods for Computer Vision
• It will become more theoretical and more mathematicallychallenging.
• We will prove theorems on the board.• You’ll have an advantage in terms of possible applications
and their numerical implementation.
For those who heard Ill-posed Problems
• We will need very little functional analysis – everything willbe in Rn.
• You’ll have an advantage in terms of the theoreticalmathematical concepts.
• I highly recommend learning how to implement thediscussed approaches. It is useful and fun!
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Formalities
Exercises
• There are no exercises for this lecture.• Occasionally: theory or programming problems.• Solution one week later in the lecture.• The more we discuss in the lecture, the more interesting
the course will be! Please don’t be shy to say something!• We’ll use OnlineTed to make the lecture more interactive.
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Formalities
Examination
• Depending on the number of attendees, the final exam willbe either oral or written.
• ECTS credits: 4
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Formalities
Miscellaneous
• My office: 02.09.061• Office hours: Tuesday 4–5pm• Lecture: Starts at 2:15pm. Short break in between.• Course website: https://vision.in.tum.de/teaching/ss2015/multiscale_methods
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Overview
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Introduction
Step 1: Make sure we are all on the same page!
Basics of convex analysis:• Convex extended real valued functions in Rn
• Minimization problems (existence, optimality condition)• Duality, Saddle point problems
Goal: Everyone knows all necessary tools to follow the lecture!
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Introduction
Step 2: Make sure everyone can try out what we are doing!
An optimization method for non-smooth convex minimization• Praxis oriented – focus on a Matlab implementation.• Idea rather than convergence analysis.
Goal: Everyone can try out him-/herself what we are doing!
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Multiscale decomposition
Step 3: Analyzing multiscale methods!
3.1 Classical theory: How does linear filtering work?
• Transform signal to different representation.• Filter coefficients.• Transform back.
→ Analyze behavior via eigendecomposition!
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Multiscale decomposition
Step 3: Analyzing multiscale methods!
3.2 A nonlinear singular vector analysis
• Is there any analogy to singular vectors for nonlinearregularization methods?
• What properties do nonlinear singular vectors have?
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Multiscale decomposition
Step 3: Analyzing multiscale methods!
3.3 Nonlinear variational methods
• Define a spectral decomposition for one-homogeneousregularizations.
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Multiscale decomposition
Step 3: Analyzing multiscale methods!
3.4 Nonlinear scale space flows
• Analyze behavior of scale space flows.• Define a spectral decomposition for one-homogeneous
regularizations.
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Multiscale decomposition
Step 3: Analyzing multiscale methods!
3.5 Nonlinear inverse scale space flows
• Analyze behavior of inverse scale space flows• Define a spectral decomposition for one-homogeneous
regularizations.
Organization andIntroduction
Michael Moeller
Motivation
Organizational Things
An Overview
updated 14.04.2015
Multiscale decomposition
Step 3: Analyzing multiscale methods!
3.6 Properties of spectral decomposition methods
• Under which conditions do the spectral decompositionapproaches yield a discrete spectrum?
• Under which conditions are the three spectraldecomposition approaches equivalent?
• Under which conditions do we obtain a nonlineareigenvalue decomposition?