Textbook Energy Minimization Methods in Computer Vision and Pattern Recognition Marcello Pelillo Ebook All Chapter PDF

Energy Minimization Methods in
Computer Vision and Pattern

Recognition Marcello Pelillo
Visit to download the full and correct content document:
https://textbookfull.com/product/energy-minimization-methods-in-computer-vision-and
-pattern-recognition-marcello-pelillo/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
Energy Minimization Methods in Computer Vision and

Pattern Recognition 10th International Conference
EMMCVPR 2015 Hong Kong China January 13 16 2015
Proceedings 1st Edition Xue-Cheng Tai
https://textbookfull.com/product/energy-minimization-methods-in-
computer-vision-and-pattern-recognition-10th-international-
conference-emmcvpr-2015-hong-kong-china-
january-13-16-2015-proceedings-1st-edition-xue-cheng-tai/
Computer Vision Pattern Recognition Image Processing

and Graphics Renu Rameshan
https://textbookfull.com/product/computer-vision-pattern-
recognition-image-processing-and-graphics-renu-rameshan/
Image Processing Computer Vision and Pattern

Recognition Hamid R. Arabnia
https://textbookfull.com/product/image-processing-computer-
vision-and-pattern-recognition-hamid-r-arabnia/
Handbook Of Pattern Recognition And Computer Vision 6th

Edition C H Chen
https://textbookfull.com/product/handbook-of-pattern-recognition-
and-computer-vision-6th-edition-c-h-chen/
Pattern Recognition and Computer Vision Third Chinese
Conference PRCV 2020 Nanjing China October 16 18 2020
Proceedings Part III Yuxin Peng
https://textbookfull.com/product/pattern-recognition-and-
computer-vision-third-chinese-conference-prcv-2020-nanjing-china-
october-16-18-2020-proceedings-part-iii-yuxin-peng/
Artificial Life and Evolutionary Computation 12th

Italian Workshop WIVACE 2017 Marcello Pelillo
https://textbookfull.com/product/artificial-life-and-
evolutionary-computation-12th-italian-workshop-
wivace-2017-marcello-pelillo/
Pattern Recognition and Computer Vision First Chinese

Conference PRCV 2018 Guangzhou China November 23 26
2018 Proceedings Part IV Jian-Huang Lai
computer-vision-first-chinese-conference-prcv-2018-guangzhou-
china-november-23-26-2018-proceedings-part-iv-jian-huang-lai/
Pattern Recognition and Computer Vision First Chinese

Conference PRCV 2018 Guangzhou China November 23 26
2018 Proceedings Part I Jian-Huang Lai
computer-vision-first-chinese-conference-prcv-2018-guangzhou-
china-november-23-26-2018-proceedings-part-i-jian-huang-lai/
Pattern Recognition and Computer Vision Third Chinese

Conference PRCV 2020 Nanjing China October 16 18 2020
Proceedings Part I 1st Edition Yuxin Peng
computer-vision-third-chinese-conference-prcv-2020-nanjing-china-
october-16-18-2020-proceedings-part-i-1st-edition-yuxin-peng/
Marcello Pelillo
Edwin Hancock (Eds.)
LNCS 10746
Energy Minimization Methods

in Computer Vision
and Pattern Recognition
11th International Conference, EMMCVPR 2017
Venice, Italy, October 30 – November 1, 2017
Revised Selected Papers
123
Lecture Notes in Computer Science 10746
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zurich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology Madras, Chennai, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7412
Marcello Pelillo Edwin Hancock (Eds.)
•
Energy Minimization Methods

in Computer Vision
and Pattern Recognition
11th International Conference, EMMCVPR 2017
Venice, Italy, October 30 – November 1, 2017
Revised Selected Papers
123
Editors
Marcello Pelillo Edwin Hancock
Ca’ Foscari University of Venice University of York
Venice York
Italy UK
ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-319-78198-3 ISBN 978-3-319-78199-0 (eBook)
https://doi.org/10.1007/978-3-319-78199-0
Library of Congress Control Number: 2018937378
LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics
© Springer International Publishing AG, part of Springer Nature 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by the registered company Springer International Publishing AG
part of Springer Nature
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
This volume contains the papers presented at the 11th International Conference on
Energy Minimization Methods in Computer Vision and Pattern Recognition
(EMMVCPR). The year 2017 marked the 20th anniversary of the inaugural meeting
of the conference series, which took place in Venice in May 1997, and we wanted to
celebrate by bringing the event back to Venice. The conference was held from October
30 to November 1, 2017, in the unique setting of Palazzo Franchetti, a historical
Venetian palace facing the Grand Canal. The conference was co-located with ICCV
2017, the 16th International Conference on Computer Vision, which was held in
Venice Lido the week before EMMCVPR.
The call for papers produced 51 submissions, resulting in the 37 papers appearing in
this volume, nine of which were presented as long talks, and the remainder as short
talks. We make no distinction between these two types of paper in this book.
The accepted papers cover a wide range of problems and perspectives including
clustering and quantum computing, image processing and video analysis, shading and
light, as well as optimization methods, and address both theoretical issues as well as
real-world problems. In addition to the contributed papers, the conference featured
two invited talks by Joachim Buhmann, from ETH Zurich (Switzerland), and Paolo
Frasconi, from the University of Florence (Italy).
We would like to take this opportunity to express our gratitude to all those who
helped to organize the conference. First of all, thanks are due to the members of the
Steering and the Scientific Committees and to all the additional reviewers. Special
thanks are due to the members of the Organizing Committee, Sebastiano Vascon and
Ismail Elezi, together with Cristiana Fiandra and her team at “The Office” for managing
the conference organization, and also Roberta D’Argenio and Agnese Boscarol, from
the European Center for Living Technology (ECLT), for their administrative support.
We gratefully acknowledge generous financial support from the Bertinoro International
Center for Informatics (BICI) and the Department of Environmental Sciences, Infor-
matics and Statistics (DAIS) of Ca’ Foscari University of Venice.
We also offer our appreciation to the editorial staff at Springer in producing this
book, and for supporting the event through publication in the LNCS series. We thank
all the authors and the invited speakers for helping to make this event a success, and
producing a high-quality publication to document the event. We warmly acknowledge
the help of Sebastiano Vascon in assembling the final proceedings volume.
February 2018 Marcello Pelillo

Edwin Hancock
Organization
Program Committee
Ognjen Arandjelovic University of St. Andrews, UK
Xiang Bai Huazhong University of Science and Technology,
China
Michael Bronstein University of Lugano, Switzerland
Andres Bruhn University of Stuttgart, Germany
Sotirios Chatzis CUT, Cyprus
Xilin Chen Institute of Computing Technology, China
Jim Clark McGill University, Canada
Daniel Cremers Technical University of Munich, Germany
Gianfranco Doretto West Virginia University, USA
Francisco Escolano University of Alicante, Spain
Mario Figueiredo Instituto de Telecomunicações,
Instituto Superior Técnico, Portugal
Alexander Fix Oculus Research, USA
Boris Flach Czech Technical University in Prague, Czech Republic
Davi Geiger NYU, USA
Edwin Hancock University of York, UK
Hiroshi Ishikawa Waseda University, Japan
Ian Jermyn Durham University, UK
Fredrik Kahl Lund University, Sweden
Zoltan Kato University of Szeged, Hungary
Anton Konushin Lomonosov Moscow State University, Russia
Hamid Krim North Carolina State University, USA
Longin Jan Latecki Temple University, USA
Hongdong Li ANU, Australia
Stan Li CBSR, CASIA, China
Stephen Maybank Birkbeck College, UK
Marcello Pelillo University of Venice, Italy
Thomas Pock Graz University of Technology, Austria
Antonio Robles-Kelly CSIRO, Australia
Emanuele Rodola Technical University of Munich, Germany
Samuel Rota Bulò Fondazione Bruno Kessler, Italy
Christoph Schnoerr Heidelberg University, Germany
Fiorella Sgallari University of Bologna, Italy
Ali Shokoufandeh Drexel University, USA
Hugues Talbot Université Paris Est, France
Wenbing Tao Huazhong University of Science and Technology,
China
VIII Organization
Andrea Torsello Università Ca Foscari, Italy

Carole Twining ISBE, University of Manchester, UK
Olga Veksler University of Western Ontario, Canada
Richard Wilson University of York, UK
Josiane Zerubia Inria, France
Additional Reviewers
Chiotellis, John Piccirilli, Marco

Haefner, Bjoern Pidhorskyi, Stanislav
Jones, Quinn Rossi, Luca
Laude, Emanuel Stoll, Michael
Lei, Zhen Sun, Kun
Liao, Shengcai Wan, Jun
Loewenhauser, Benedikt Wang, Xiaobo
Ma, Zhanyu Welk, Martin
Marin, Dmitrii Xia, Guisong
Maurer, Daniel Xiao, Yang
Moellenhoff, Thomas Xu, Yongchao
Motiian, Saeid Yang, Yang
Möllenhoff, Thomas Zhu, Xiangyu
Contents
Clustering and Quantum Methods
Ising Models for Binary Clustering via Adiabatic Quantum Computing . . . . . 3

Christian Bauckhage, E. Brito, K. Cvejoski, C. Ojeda, Rafet Sifa,
and S. Wrobel
Quantum Interference and Shape Detection . . . . . . . . . . . . . . . . . . . . . . . . 18

Davi Geiger and Zvi M. Kedem
Structured Output Prediction and Learning for Deep Monocular

3D Human Pose Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Stefan Kinauer, Riza Alp Güler, Siddhartha Chandra,
and Iasonas Kokkinos
Dominant Set Biclustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

Matteo Denitto, Manuele Bicego, Alessandro Farinelli,
and Marcello Pelillo
Bragg Diffraction Patterns as Graph Characteristics . . . . . . . . . . . . . . . . . . . 62

Francisco Escolano and Edwin R. Hancock
Motion and Tracking
Variational Large Displacement Optical Flow Without Feature Matches . . . . . 79

Michael Stoll, Daniel Maurer, and Andrés Bruhn
Temporal Semantic Motion Segmentation Using Spatio

Temporal Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Nazrul Haque, N. Dinesh Reddy, and Madhava Krishna
Depth-Adaptive Computational Policies for Efficient Visual Tracking . . . . . . 109

Chris Ying and Katerina Fragkiadaki
Multiframe Motion Coupling for Video Super Resolution . . . . . . . . . . . . . . 123

Jonas Geiping, Hendrik Dirks, Daniel Cremers, and Michael Moeller
Illumination-Aware Large Displacement Optical Flow . . . . . . . . . . . . . . . . . 139

Michael Stoll, Daniel Maurer, Sebastian Volz, and Andrés Bruhn
Location Uncertainty Principle: Toward the Definition of Parameter-Free

Motion Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Shengze Cai, Etienne Mémin, Pierre Dérian, and Chao Xu
X Contents
Image Processing and Segmentation
Autonomous Multi-camera Tracking Using Distributed

Quadratic Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Yusuf Osmanlıoğlu, Bahareh Shakibajahromi, and Ali Shokoufandeh
Nonlinear Compressed Sensing for Multi-emitter X-Ray Imaging . . . . . . . . . 189

Maria Klodt and Raphael Hauser
Unified Functional Framework for Restoration of Image Sequences

Degraded by Atmospheric Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Naftali Zon and Nahum Kiryati
A Convex Approach to K-Means Clustering and Image Segmentation . . . . . . 220

Laurent Condat
Luminance-Guided Chrominance Denoising with Debiased

Coupled Total Variation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Fabien Pierre, Jean-François Aujol, Charles-Alban Deledalle,
and Nicolas Papadakis
Optimizing Wavelet Bases for Sparser Representations . . . . . . . . . . . . . . . . 249

Thomas Grandits and Thomas Pock
Bottom-Up Top-Down Cues for Weakly-Supervised Semantic

Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Qibin Hou, Daniela Massiceti, Puneet Kumar Dokania,
Yunchao Wei, Ming-Ming Cheng, and Philip H. S. Torr
Vehicle X-Ray Images Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

Abraham Marciano, Laurent D. Cohen, and Najib Gadi
Colour, Shading and Reflectance of Light
Superpixels Optimized by Color and Shape . . . . . . . . . . . . . . . . . . . . . . . . 297

Vitaliy Kurlin and Donald Harvey
Maximum Consensus Parameter Estimation by Reweighted ‘1 Methods . . . . . 312

Pulak Purkait, Christopher Zach, and Anders Eriksson
A Graph Theoretic Approach for Shape from Shading . . . . . . . . . . . . . . . . . 328

Robert Scheffler, Ashkan Mansouri Yarahmadi, Michael Breuß,
and Ekkehard Köhler
A Variational Approach to Shape-from-Shading Under

Natural Illumination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
Yvain Quéau, Jean Mélou, Fabien Castan, Daniel Cremers,
and Jean-Denis Durou
Contents XI
Sharpening Hyperspectral Images Using Spatial and Spectral Priors

in a Plug-and-Play Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Afonso M. Teodoro, José M. Bioucas-Dias, and Mário A. T. Figueiredo
Inverse Lightfield Rendering for Shape, Reflection

and Natural Illumination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Antonin Sulc, Ole Johannsen, and Bastian Goldluecke
Shadow and Specularity Priors for Intrinsic Light Field Decomposition . . . . . 389
Anna Alperovich, Ole Johannsen, Michael Strecke,
and Bastian Goldluecke
Propagation and Time-Evolution
Modelling Stable Backward Diffusion and Repulsive Swarms with Convex

Energies and Range Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
Leif Bergerhoff, Marcelo Cardénas, Joachim Weickert,
and Martin Welk
Euler-Lagrange Network Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424

Jianjia Wang, Richard C. Wilson, and Edwin R. Hancock
Slack and Margin Rescaling as Convex Extensions

of Supermodular Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Matthew B. Blaschko
Multi-object Convexity Shape Prior for Segmentation . . . . . . . . . . . . . . . . . 455

Lena Gorelick and Olga Veksler
Fast Asymmetric Fronts Propagation for Voronoi Region Partitioning

and Image Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
Da Chen and Laurent D. Cohen
PointFlow: A Model for Automatically Tracing Object Boundaries

and Inferring Illusory Contours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
Fang Yang, Alfred M. Bruckstein, and Laurent D. Cohen
An Isotropic Minimal Path Based Framework for Segmentation

and Quantification of Vascular Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 499
Emmanuel Cohen, Laurent D. Cohen, Thomas Deffieux,
and Mickael Tanter
XII Contents
Inference, Labeling and Relaxation
Limited-Memory Belief Propagation via Nested Optimization . . . . . . . . . . . . 517

Christopher Zach
Geometric Image Labeling with Global Convex Labeling Constraints . . . . . . 533

Artjom Zern, Karl Rohr, and Christoph Schnörr
Discretized Convex Relaxations for the Piecewise Smooth

Mumford-Shah Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
Christopher Zach and Christian Häne
A Projected Gradient Descent Method for CRF Inference

Allowing End-to-End Training of Arbitrary Pairwise Potentials. . . . . . . . . . . 564
Måns Larsson, Anurag Arnab, Fredrik Kahl, Shuai Zheng,
and Philip Torr
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581

Clustering and Quantum Methods
Ising Models for Binary Clustering
via Adiabatic Quantum Computing
Christian Bauckhage1,2(B) , E. Brito1,2 , K. Cvejoski1,2 , C. Ojeda1,2 ,

Rafet Sifa1,2 , and S. Wrobel1,2
1
Fraunhofer IAIS, Sankt Augustin, Germany
christian.bauckhage@iais.fraunhofer.de
2
B-IT, University of Bonn, Bonn, Germany
Abstract. Existing adiabatic quantum computers are tailored towards

minimizing the energies of Ising models. The quest for implementations
of pattern recognition or machine learning algorithms on such devices
can thus be seen as the quest for Ising model (re-)formulations of their
objective functions. In this paper, we present Ising models for the tasks
of binary clustering of numerical and relational data and discuss how to
set up corresponding quantum registers and Hamiltonian operators. In
simulation experiments, we numerically solve the respective Schrödinger
equations and observe our approaches to yield convincing results.
1 Introduction
After decades of being a mainly theoretical endeavor, quantum computing now
appears to become practical. First working quantum computers are being sold,
major industrial players invest into corresponding research and development,
and a growing number of voices predicts rapid technological progress [1–3]. All
these developments will likely impact machine learning and pattern recognition
because quantum computers promise fast solutions to the kind of optimization
problems dealt with in these fields [4–8].
In this paper, we are therefore concerned with the feasibility of quantum
computing for pattern recognition. In particular, we consider binary clustering
of numerical or relational data and show how to approach these problems using
adiabatic quantum computing [9].
Our interest in adiabatic quantum computing stems from the fact that recent
technological progress has happened mainly in this area [10,11]. Available devices
are geared towards finding low energy states of Ising models
n
n

H(s) = Qij si sj + qi si (1)
i,j=1 i=1
where H is a Hamiltonian or energy operator, s ∈ {−1, 1}n is a vector that

represents the state of a system of n bipolar entities si (say, magnetic spins),
c Springer International Publishing AG, part of Springer Nature 2018

M. Pelillo and E. Hancock (Eds.): EMMCVPR 2017, LNCS 10746, pp. 3–17, 2018.
https://doi.org/10.1007/978-3-319-78199-0_1
4 C. Bauckhage et al.
Q ∈ Rn×n is called the coupling matrix, and q ∈ Rn models external influences

to the system.
Adiabatic quantum computing relies on the adiabatic theorem [12] which
states that if a quantum mechanical system in a low energy configuration or
ground state of a Hamiltonian is subjected to gradual changes, it will end up
in the ground state of the resulting Hamiltonian. To harness this principle for
problem solving, one prepares a system in the ground state of a simple, problem
independent Hamiltonian and adiabatically evolves it to a Hamiltonian whose
ground state represents a solution to the problem at hand. One of the general
challenges of adiabatic quantum computing is thus to devise suitable problem
Hamiltonians. However, if the problem at hand can be expressed as an Ising
energy minimization problem, these Hamiltionians are readily constructed. In
this sense, the quest for adiabatic quantum implementations of pattern recog-
nition algorithms becomes the quest for Ising model (re-)formulations of their
objective functions.
Our presentation in this paper is therefore structured as follows: First of all,
we present Ising models for the problems of computing k = 2 means clustering
and balanced graph cuts.
Second of all, we discuss technical details of adiabatic quantum comput-
ing and show how to set up a system of qubits along with two appropriate
Hamiltonians to perform quantum adiabatic binary clustering.
Third of all, we present simple practical examples that demonstrate the feasi-
bility of adiabatic quantum clustering and illustrate how appropriately prepared
systems of qubits evolve towards a solution.
Finally, we conclude with a discussion of related work and a summary of our
contributions.
2 An Ising Model for Numerical Binary Clustering

In this section, we address the problem of binary clustering of numerical data
and derive a heuristic for k = 2 means clustering that leads to an Ising model.
Throughout, we let X ⊂ Rm denote a sample of n = |X| data points and assume
w.l.o.g. that these data are normalized such that the sample mean x̄ = 0.
Recall that hard k-means clustering is a prototype-based clustering technique
that partitions the data in X into k disjoint subsets
2 2

Xi = x ∈ X x − x̄i < x − x̄j ∀ i = j (2)
of size ni = |Xi | whose centroids are given by

1
x̄i = x. (3)
ni
x∈Xi
Ising Models for Binary Clustering via Adiabatic Quantum Computing 5
2.1 The “Standard” k-Means Objective Function

Given its definition of a cluster in (2), the basic problem of k-means clustering is
to find appropriate cluster prototypes x̄i . Well known algorithms such as those
of Lloyd [13], Hartigan [14], or MacQueen [15] accomplish this by minimizing
the within cluster scatter
k

SW (k) = x − x̄i 2 (4)
i=1 x∈Xi
with respect to the x̄1 , . . . , x̄k .

Since the minimization objective in (4) involves the distances x − x̄i that
occur in the definition of a cluster, it formalizes an intuitive idea. Yet, despite
its seeming simplicity, k-means clustering proves to be NP hard [16]. Algorithms
such as those in [13–15] are therefore but heuristics for which there is no guar-
antee that they will find the global minimum of (4).
2.2 An Alternative k-Means Objective Function

A rarely used fact in the context of k-means clustering is that the problem of
minimizing (4) is equivalent to the problem of maximizing the between cluster
scatter
k
2
SB (k) = ni nj x̄i − x̄j . (5)
i,j=1
This can be easily seen using Fisher’s analysis of variance [17]. It establishes
that the total scatter of a sample of data points can be written as

ST = x − x̄2 = SW (k) + 1 SB (k) (6)
2n
x∈X
which—since ST and n are constants—is to say that any decrease of SW in (4)

implies an increase of SB in (5).
2.3 An Ising Model for k = 2 Means Clustering

For the case where k = 2, the maximization objective in (5) simplifies to
2
SB (2) = 2 n1 n2 x̄1 − x̄2 . (7)
Interestingly, this expression provides an intuition for the fact that k-means
clustering is agnostic of cluster shapes and distances and thus tends to produce
clusters of about equal size even in situations where this seems unreasonable
[18]. In order for SB in (7) to be large, both the distance x̄1 − x̄2 between
the two cluster centers and the product n1 n2 of the two cluster sizes have to be
large. However, since the sum n1 + n2 = n is fixed, the product of the sizes will
be maximal if n1 = n2 = n2 .
This observation provides us with a heuristic handle to rewrite the objective

in (7) which will then allow for setting up an Ising model for k = 2 means
clustering.
Assuming that at a solution produced by conventional k-means clustering
we will likely have n1 ≈ n2 ≈ n2 allows for the following approximation or
reformulation of the objective function
2 2 2
2 n1 n2 x̄1 − x̄2 ≈ 2 n4 x̄1 − x̄2
2 2
= 2 n2 x̄1 − x̄2 = 2 n1 x̄1 − n2 x̄2 (8)
and, using this heuristic, the problem of k = 2 means clustering becomes to solve
2
argmax n1 x̄1 − n2 x̄2 . (9)
x̄ 1 , x̄ 2
Next, we express the norm in (9) in a form that does not explicitly depend
on the cluster means x̄i . To this end, we gather the given data in a data matrix
X = [x1 , . . . , xn ] ∈ Rm×n and introduce two binary vectors z 1 , z 2 ∈ {0, 1}n
which indicate cluster memberships in the sense that entry l of z i is 1 if xl ∈ Xi
and 0 otherwise. This way, we can write n1 x̄1 = Xz 1 as well as n2 x̄2 = Xz 2
and thus
n1 x̄1 − n2 x̄2 2 = X z 1 − z 2 2 = Xs2 . (10)
Here, s is guaranteed to be a bipolar vector because, in hard k-means clustering,
every given data point is assigned to one and only one cluster so that
z 1 − z 2 = z 1 − (1 − z 1 ) = 2 z 1 − 1 = s ∈ {−1, 1}n . (11)
This, however, establishes that there is an Ising model formulation for the
problem of k = 2 means clustering of zero mean data. On the one hand, zero
mean data implies
x̄ = 1
n (n1 x̄1 + n2 x̄2 ) = 0 ⇔ n1 x̄1 = −n2 x̄2 (12)
so that the two means will always be of opposite sign. On the other hand, since
Xs2 = sT X T Xs, the problem in (9) is equivalent to solving
argmax sT X T X s ⇔ argmin −sT X T X s. (13)

s∈{−1,1}n s∈{−1,1}n
Because of (12) this will necessarily produce a vector s whose entries are not all
equal and thus induce a clustering of the data in X. If we furthermore substitute
Q = X T X, the problem in (23) can also be written as
n

argmin − Qij si sj (14)
s∈{−1,1}n i,j=1
and is now recognizable as an instance of an Ising energy minimization problem.

Looking at (14), two final remarks appear to be in order. First of all, the
coupling matrix Q = X T X is a Gram matrix. The clustering heuristic derived
in this section therefore allows for invoking the kernel trick and is thus applicable
to a wide variety of practical problems.
Second of all, (14) exposes the “hardness” of k = 2 means clustering for it
reveals it to be an integer programming problem. That is, it shows that binary
clustering is to find a suitable label vector s ∈ {−1, 1}n whose entries assign data
points to clusters. A naı̈ve approach to k = 2 means clustering would therefore
be to evaluate (14) for each of the 2n possible assignments of n data points to 2
clusters. On a conventional digital computer this is of course impractical if n is
large. On a quantum computer, on the other hand, we could prepare a system of
n qubits that is in a superposition of 2n states each of which reflects a possible
solution. The trick is then to evolve the system such that, when its wave function
is collapsed, it is more likely to be found in a state that corresponds to a good
solution. Below, we discuss how to accomplish this.
3 An Ising Model for Relational Binary Clustering

Next, we address the problem of binary clustering of relational data. In particu-
lar, we revisit the idea of graph clustering via minimum cuts, focus on normalized
cuts as popularized in [19], and analyze their properties in order to derive a novel
cut criterion that leads to an Ising model.
Throughout, we restrict our discussion to undirected and unweighted graphs
G = (V, E) consisting of vertices V = {v1 , . . . , vn } and edges E ⊆ V × V . For
these, we recall several basic properties:
The adjacency matrix A of an unweighted graph is a binary n × n matrix
where Aij = 1 if (vi , vj ) ∈ E and Aij = 0 otherwise. Moreover, if the graph is
undirected, A will be symmetric.
The degree vector d of G is an n-dimensional vector whose entries
n

di = Aij . (15)
j=1
count the degrees of the vertices vi . The degree matrix of G is a diagonal matrix
D = diag(d ) and it is easily verified that both the degree matrix as well as the
outer product matrix ddT are symmetric and positive semidefinite.
The Laplacian matrix of G is given by L = D − A. Again, if G is undirected,
its Laplacian is a symmetric and positive semidefinite matrix.
Finally, the volume vol(G) of G amounts to the sum of all its vertex degrees
n

vol(G) = di = 1T d (16)
i=1
and, for brevity, we will henceforth write it as η ≡ vol(G).

3.1 Normalized Cuts for Balanced Graph Clustering

The problem of clustering a given graph G into two disjoint subgraphs G1 and
G2 such that V1 ∪ V2 = V , V1 ∩ V2 = ∅, and E1 ∩ E2 = ∅ is often formalized as
the problem of finding an appropriate graph cut of minimum cost

cut(V1 , V2 ) = cut(V2 , V1 ) = Aij . (17)
vi ∈V1
vj ∈V2
Interestingly, the costs of cutting G can be expressed in terms of its graph

Laplacian. Assuming a bipolar indicator vector s ∈ {−1, 1}n whose entries
+1, if vi ∈ V1
si = (18)
−1, if vi ∈ V2
indicate which of the clusters a vertex will belong to after the cut, straightforward
algebra [20] reveals that
cut(V1 , V2 ) = 1
4 sT L s. (19)
However, since naı̈vely minimizing (19) might lead to unbalanced partitions,
various modifications have been proposed. A particularly popular heuristic is to
consider normalized cuts
sT L s sT L s η1 + η2
Ncut(V1 , V2 ) = + = sT L s · (20)
η1 η2 η1 η2
which trade off the size of cuts and the volumes of the resulting subgraphs [19].
Finally, it is easy to see that the problem of minimizing (20) with respect to
s ∈ {−1, 1}n is an integer programming problem and thus NP hard. The Ncut
problem is therefore usually relaxed to spectral clustering [19,21,22]. However,
on an adiabatic quantum computer, we could again prepare a system of n qubits
which are in a superposition of the 2n possible solutions and evolve the system
such that the optimal solution is more likely to be measured. This, however,
requires an optimization objective that can be expressed as an Ising model.
Since the Ncut objective in (20) does not meet this requirement, we next derive
a balanced cut criterion that corresponds to an Ising model.
3.2 An Ising Model for Balanced Graph Clustering

Looking at (20), we realize that, in order for Ncut to be small, the quadratic
form sT L s has to be small and the product η1 η2 has to be large.
However, since η1 + η2 = η is a constant, the product η1 η2 will be maximal if
η1 = η2 = η2 . In other words, the Ncut criterion implicitly enforces η1 ≈ η2 . We
may therefore just as well require the squared difference (η1 − η2 )2 to be small.
Regarding this idea, we note that

η1 − η2 = di − d j = sT d (21)
vi ∈V1 vj ∈V2
so that the squared difference of the volumes of the subgraphs that result from
a graph cut can be written as
(η1 − η2 )2 = sT ddT s. (22)
This observation immediately leads to yet another heuristic for balanced
minimum cuts: combining (19) and (22), we can formalize graph clustering as the
task of solving the following equality constrained quadratic binary optimization
problem
s∗ = argmin sT L s s. t. sT ddT s = 0 (23)
s∈{−1,1}n
whose Lagrangian is given by

L(s, λ) = sT L s + λ sT ddT s. (24)
Contrary to the Ncut formulation, our formulation is now an instance of an
Ising energy minimization problem. However, in order to minimize (24) on an
adiabatic quantum computer, we need to specify the Lagrange multiplier λ so
that the problem indeed only depends on the bipolar indicator vector s. To this
end, we propose to simply choose λ = η2 . We then have 0 ≤ η2 sT ddT s ≤ 2 η
and, since we also always have 0 ≤ sT L s < 2 η, neither term on the right hand
side of (24) will dominate the minimization objective.
In summary, our derivation leads to the following graph cut criterion, which,
for want of a better name, we will call an Ising cut
2 T T
Icut(V1 , V2 ) = sT L s + s dd s
η
n

2
= sT L+ ddT s = sT Q s = si Qij sj . (25)
η i,j=1
4 Adiabatic Quantum Binary Clustering

To perform adiabatic quantum binary clustering of n entities, we consider a
system of n qubits that is in a superposition of 2n basis states
n
−1
2

ψ(t) = ai (t) ψi (26)
i=0
whose time dependent amplitudes ai ∈ C obey i |ai |2 = 1. We understand each

of the different basis states

ψ0 = 000 . . . 000 (27)

ψ1 = 000 . . . 001 (28)

ψ2 = 000 . . . 010 (29)

ψ3 = 000 . . . 011 (30)
..
.
as an indicator vector that represents one of the 2n possible assignment of n

entities to 2 clusters and use the common shorthand to express tensor products,
for instance
ψ1 = 000 . . . 001 = 0 ⊗ 0 ⊗ . . . ⊗ 1 . (31)
If a system such as in (26) evolves under the influence of a time-dependent
Hamiltonian H(t), its behavior is governed by the Schrödinger equation
∂
ψ(t) = −i H(t) ψ(t) (32)
∂t
where we have set = 1. In adiabatic quantum computing, we consider periods
ranging from t = 0 to t = τ and assume the Hamiltonian at time t to be given
as a convex combination of two static Hamiltonians, namely

H(t) = 1 − τt HB + τt HP . (33)
HB is called the beginning Hamiltonian whose ground state is easy to construct

and HP is the problem Hamiltonian whose ground state encodes the solution to
the problem at hand.
For simple Ising models such as the ones in (14) and (25), there are by now
standard suggestions for how to set up a suitable problem Hamiltonian [9]. In
particular, we may define
n

HP = Qij σzi σzj (34)
i,j=1
where σzi denotes the Pauli spin matrix σz acting on the ith qubit, that is
σzi = I ⊗ I ⊗ . . . ⊗ I ⊗ σz ⊗ I ⊗ I . . . ⊗ I . (35)

i−1 terms n−i terms
The beginning Hamiltonian is then typically chosen to be orthogonal to the

problem Hamiltonian, for instance
n

HB = σxi (36)
i=1
where σxi is defined as above, this time with respect to the Pauli spin matrix
σx .
To compute a clustering, we then let |ψ(t) evolve from |ψ(0) to ψ(τ )
where |ψ(0) is chosen to be the ground state of HB . That is, if λ denotes the
smallest eigenvalue of HB , the initial state |ψ(0) of the system corresponds to
the solution of
HB ψ(0) = λ ψ(0) . (37)
Finally, upon termination of its evolution, a measurement is performed on
the n qubit system. This will cause the wave function ψ(τ) to collapse to a
particular basis state and the probability for this state to be ψi is given by the
amplitude |ai (τ )|2 . However, since the adiabatic evolution was steered towards
the problem Hamiltonian HP , basis states that correspond to ground states of
HP are more likely to be found.
On an adiabatic quantum computer, this process be carried out physically.
On a digital computer, we may simulate it by numerically solving
τ

ψ(τ ) = −i H(t) ψ(t) dt. (38)
0
5 Practical Examples
In this section, we present several examples which demonstrate the practical
feasibility of the above ideas. Our examples are of didactic nature and mainly
intended to illustrate the adiabatic evolution of n qubit systems. We thus restrict
ourselves to rather small n = n1 + n2 in order to be able to comprehensibly
visualize how the amplitudes of 2n basis states evolve over time.
In each experiment, we simulate quantum adiabatic evolutions on a digital
computer. To this end, we set up the corresponding problem
Hamiltonian HP ,
the beginning Hamiltonian HB , and its ground state ψ(0) as discussed above
and use the Python quantum computing toolbox QuTiP [23] to numerically
solve (38) for t ∈ [0, τ = 50].
5.1 Adiabatic Quantum k = 2 Means Clustering

Next, we consider two examples of adiabatic quantum clustering of numerical
data. In both cases, we consider a data matrix X = [X 1 , X 2 ] ∈ R2×n whose
column vectors form two clusters of size n1 and n2 , respectively, and let an n
qubit system evolve in order to uncover these clusters.
Figure 1(a) shows a zero mean sample of n = 8 data points. In order to
quantum adiabatically cluster these data into two clusters, we consider the Ising
model in (14) using Q = X T X. Figure 1(c) illustrates the temporal evolution
of the amplitudes |ai (t)|2
of the 28
= 256 basis states ψi the quantum system

ψ(t) can be in. At t = 0, all states are equally likely but over time their
amplitudes begin to increase or decrease. At t = τ , two of the basis states have
considerably higher amplitudes than the others so that a measurement will likely
cause the system to collapse to either of these more probable states.
The top 8 most likely basis states for the system to be found in at the end
of this experiment are listed in Table 1. Looking at the table, we recognize that
the two most likely states are indeed 00111111 and 11000000 which, when
understood as cluster indicator vectors, both produce the result in Fig. 1(b).
For the six next most likely states, we observe that they will produce clusters
of sizes 3 and 5 and 4 and 4, respectively. In other words, although our heuristic
approximation of the k-means objective in (9) and the resulting Ising model in
(14) were based on the assumption of clusters of about equal size, this practical
result shows that the model can nevertheless produce clusters of imbalanced
sizes if the data are distributed in a manner that would require it.
|ai |2
0.025
0.020
(a) n = 8 data points in R2
0.015
0.010
0.005
0.000
t→
|ψ i

(b) binary clustering result (c) evolution of amplitudes of basis states ψi
Fig. 1. Example of adiabatic quantum k = 2 means clustering using the Ising model
in (14). (a) zero mean sample of 8 data points; (b) corresponding clustering result; (c)
adiabatic evolution of a system of 8 qubits. During its evolution over time t, the system
is in a superposition of 28 = 256 basis states ψi each representing a possible binary
clustering. Initially, it is equally likely to find the system in any of these states. At the
end, two basis states have noticeably higher amplitudes
|ai |2 than
the others
and are
therefore more likely to be measured; these are 00111111 and 11000000 and they
both induce the result in (b).
Table 1. The 8 most likely states for the system in Fig. 1 to be in at t = τ .

ψi |ai |2
|00111111 0.02547562
|11000000 0.02547562
|11010000 0.02104650
|00101111 0.02104650
|00110111 0.01929660
|11001000 0.01929660
|11010001 0.01812229
|00101110 0.01812229
Figure 2(a) shows n = 16 data points arranged in a pattern that suggests the
use of kernel k-means clustering. Accordingly, we choose the coupling matrix for
the Ising model in (14) to be a centered kernel matrix
1 1 1
Qij = k(xi , xj ) − k(xi , xl ) − k(xk , xj ) + 2 k(xk , xl ) (39)
n n n
l k k,l
2
where k xi , xj = exp −xi − xj /2 σ 2 and σ = 0.25. During its quantum
adiabatic evolution, the corresponding 16 qubit system ψ(t) is in a super-
position of 216 = 65536 basis states but evolves to a configuration in which
two of these basis states have much higher amplitudes than the others (see
Fig. 2(c)). Here, the two most likely basis states are 1111111100000000 and

0000000011111111 and they cluster the data in a manner a human observer
would expect (see Fig. 2(b)).
(a) n = 16 data points in R2

(b) binary clustering result (c) evolution of amplitudes of basis states ψi
Fig. 2. Example of adiabatic quantum kernel k = 2-means clustering. (a) sample of 16

data points; (b) corresponding clustering result; (c) adiabatic evolution of a system of
16 qubits. Throughout, the system is in a superposition of 216 = 65536 basis states.
Upon termination of the evolution,the two
most likely basis states for the system to
be found in are 1111111100000000 and 0000000011111111 .
5.2 Adiabatic Quantum Graph Clustering

Next, we present an example of adiabatic quantum graph clustering using the
Ising model in (25).
Figure 3(a) shows a graph composed of two recognizable communities. To
produce it, we created two Erdős-Rényi random graphs of 6 and 10 vertices,
respectively, where the edge probability was set to p = 0.5 in both cases. To
connect these subgraphs, we then randomly drew two edges between them.
Figure 3(c) shows the temporal evolution of the amplitudes |ai (t)|2 of the
2 = 65336 basis states |ψ i the quantum system |ψ(t) used to cluster this
16
(a) graph of n = 16 vertices

(b) binary clustering result (c) evolution of amplitudes of basis states ψ i
Fig. 3. Example of adiabatic quantum clustering of a graph composed of two Erdős-

16
the corresponding qubit system is in a superposition of 2 =
Rényi graphs. Over time,
65536 basis states ψ i representing possible clusterings. Initially, it is equally likely to
find the system in any of these states. At the end, two states have noticeably higher
amplitudes
|ai |2 than the
others
and are therefore more likely to be measured; these
are 0000000000111111 and 1111111111000000 and they both induce the clusters
in (b).
graph can be in. At t = 0, all basis states are equally likely but over time their
amplitudes begin to increase or decrease. At t = τ , two of the basis states have
considerably higher amplitudes than the others so that a measurement of the
quantum system at this time will likely cause it to collapse to either one of these
more probable states.
Table 2 shows the top 8 most likely basis states for the qubit system of this
experiment to be found in at the end of its adiabatic evolution. Looking at this
table, we recognize that the two most likely basis states are |0000000000111111
and |1111111111000000 which, due to obvious symmetries, both produce the
result shown in Fig. 3(b).
Table 2. The 8 most likely states for the system in Fig. 3 to be in at t = τ .

ψi |ai |2
|0000000000111111 0.00094453
|1111111111000000 0.00094453
|1111111111010000 0.00058813
|0000000000101111 0.00058813
|1111111111100000 0.00058813
|0000000000011111 0.00058813
|0000000000110111 0.00056642
|1111111111001000 0.00056642
6 Related Work
Interest in quantum computing for machine learning and pattern recognition

is steadily increasing [5,7,8,24–28]. Indeed, many algorithms in these fields
deal with demanding optimization problems for which quantum computing is
expected to yield considerable speed-up.
Unsurprisingly, quantum computing approaches to problems in the context
of binary clustering have been considered before. However, prior work reported
in [5,6,8] follows the paradigm of quantum gate computing. Though equivalent
to adiabatic quantum computing, quantum gate computing algorithms do not
attempt to optimize Ising models and are, as of this writing, not yet practically
feasible for lack of corresponding hardware and, crucially, the availability of
quantum oracles.
Lloyd et al. [7] present an adiabatic quantum implementation of Lloyd’s
heuristic
algorithm
[13] for k-means
clustering
which has a run time complexity
of O k log(mn) rather than O poly(mn) where m denotes data dimensionality.
However, their approach hinges on the availability of a quantum random access
memory (qRAM), a theoretical concept which, as of this writing, is not yet
technically feasible. While the approach we discussed in this paper is admittedly
more restricted in that it only allows for k = 2 means clustering, its Hamiltonians
are simple and easy to work work with, and it does not require a qRAM.
Graph clustering, too, has been previously studied from the point of view of
quantum computing. In a very recent contribution, Ushijima-Mwesigwa et al. [29]
actually perform graph clustering experiments on D-Wave computers [10,11].
Rather than an Ising model derived from the idea of normalized cuts as proposed
above, they consider an Ising model based on Newman’s modularity measure
[30]. Their experiments show this approach to equal or even outperform current
state-of-the art methods.
7 Summary
After decades of being a mainly theoretical concept, quantum computing is now
becoming practical. Companies like IBM, Google, or Microsoft invest increasing
resources into corresponding research and development [1] and further rapid
progress is expected [2,3]. All of this will likely impact statistical learning and
pattern recognition, because quantum computers have the potential to accelerate
the kind of optimization procedures that are at the heart of many algorithms in
these fields.
Since at the time of this writing, adiabatic quantum computers appear to
be most advanced, we were concerned with adiabatic quantum computing for
data analysis, especially for data clustering. From an abstract point of view,
the problem of setting up pattern recognition algorithms for adiabatic quantum
computing can be seen as the problem of expressing their objective functions in
terms of Ising models because adiabatic quantum computers are tailored towards
minimizing Ising energies. We thus derived Ising models for (kernel) k = 2-means
clustering and balanced graph cuts. Simulated examples demonstrated that the
Schrödinger equations governing the dynamics of corresponding qubit systems
can indeed be used to cluster data.
References
1. Technology Quarterly: Quantum Devices. The Economist, March 2017
2. Castelvecchi, D.: Quantum computers ready to leap out of the lab in 2017. Nature
541(7635), 9–10 (2017). https://doi.org/10.1038/541009a
3. Gibney, E.: D-Wave upgrade: how scientists are using the world’s most controversial
quantum computer. Nature 541(7638), 447–448 (2017). https://doi.org/10.1038/
541447b
4. Grover, L.: From Schrödinger’s equation to the quantum search algorithm. J. of
Phys. 56(2), 333–348 (2001). https://doi.org/10.1119/1.1359518
5. Aı̈meur, E., Brassard, G., Gambs, S.: Quantum clustering algorithms. In: Proceed-
ings ICML (2007)
6. Aı̈meur, E., Brassard, G., Gambs, S.: Quantum speed-up for unsupervised learning.
Mach. Learn. 90(2), 261–287 (2013). https://doi.org/10.1007/s10994-012-531305
7. Lloyd, S., Mohseni, M., Rebentrost, P.: Quantum algorithms for supervised and
unsupervised machine learning. arXiv:1307.0411 [quant-ph] (2013)
8. Wiebe, N., Kapoor, A., Svore, K.: Quantum algorithms for nearest-neighbor meth-
ods for supervised and unsupervised learning. Quantum Inf. Comput. 15(3–4),
316–356 (2015)
9. Albash, T., Lidar, D.: Adiabatic quantum computing. arXiv:1611.04471 [quant-ph]
(2016)
10. Bian, Z., Chudak, F., Macready, W., Rose, G.: The Ising model: teaching an old
problem new tricks. Technical report, D-Wave Systems (2010)
11. Johnson, M., Amin, M., Gildert, S., Lanting, T., Hamze, F., Dickson, N., Harris,
R., Berkley, A., Johansson, J., Bunyk, P., Chapple, E., Enderud, C., Hilton, J.,
Karimi, K., Ladizinsky, E., Ladizinsky, N., Oh, T., Perminov, I., Rich, C., Thom,
M., Tolkacheva, E., Truncik, C., Uchaikin, S., Wang, J., Wilson, B., Rose, G.:
Quantum annealing with manufactured spins. Nature 473(7346), 194–198 (2011).
https://doi.org/10.1038/nature10012
12. Born, M., Fock, V.: Beweis des Adiabatensatzes. Zeitschrift für Physik 51(3–4),
165–180 (1928). https://doi.org/10.1007/BF01343193
13. Lloyd, S.: Least squares quantization in PCM. IEEE Trans. Inf. Theory 28(2),
129–137 (1982). https://doi.org/10.1109/TIT.1982.1056489
14. Hartigan, J., Wong, M.: Algorithm AS 136: a k-means clustering algorithm. J. Roy.
Stat. Soc. C 28(1), 100–108 (1979). https://doi.org/10.2307/2346830
15. MacQueen, J.: Some methods for classification and analysis of multivariate observa-
tions. In: Proceedings Berkeley Symposium on Mathematical Statistics and Prob-
ability (1967)
16. Aloise, D., Deshapande, A., Hansen, P., Popat, P.: NP-hardness of euclidean
sum-of-squares clustering. Mach. Learn. 75(2), 245–248 (2009). https://doi.org/
10.1007/s10994-009-5103-0
17. Fisher, R.: On the probable error of a coefficient correlation deduced from a small
sample. Metron 1, 3–32 (1921)
18. MacKay, D.: Information Theory, Inference, and Learning Algorithms. Cambridge
University Press, Cambridge (2003)
19. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Trans. PAMI
22(8), 888–905 (2000). https://doi.org/10.1109/34.868688
20. Dhillon, I.: Co-clustering documents and words using bipartite spectral graph par-
titioning. In: Proceedings KDD (2001)
21. von Luxburg, U.: A tutorial on spectral clustering. Stat. Comput. 17(4), 395–416
(2007). https://doi.org/10.1007/s11222-007-9033-z
22. Fiedler, M.: A property of eigenvectors of nonnegative symmetric matrices and its
application to graph theory. Czech. Math. J. 25(4), 619–633 (1975)
23. Johansson, J., Nation, P., Nori, F.: QuTiP 2: a python framework for the dynamics
of open quantum systems. Comput. Phys. Commun. 184(4), 1234–1240 (2013).
https://doi.org/10.1016/j.cpc.2012.11.019
24. Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., Lloyd, S.: Quan-
tum machine learning. arXiv:1611.09347 [quant-ph] (2016)
25. Dunjiko, V., Taylor, J., Briegel, H.: Quantum-enhanced machine learning. Phys.
Rev. Lett. 117(13), 130501 (2016). https://doi.org/10.1103/PhysRevLett.117.
130501
26. Schuld, M., Sinayskiy, I., Petruccione, F.: An introduction to quantum
machine learning. Contemp. Phys. 56(2), 172–185 (2014). https://doi.org/10.1080/
00107514.2014.964942
27. Wiebe, N., Kapoor, A., Svore, K.: Quantum perceptron models. In: Proceedings
NIPS (2016)
28. Wittek, P.: Quantum Machine Learning. Academic Press, London (2014)
29. Ushijima-Mwesigwa, H., Negre, C., Mniszewski, S.: Graph partitioning using quan-
tum annealing on the D-Wave system. arXiv:1705.03082 [quant-ph] (2017)
30. Newman, M.: Modularity and community structure in networks. PNAS 103(23),
8577–8582 (2006). https://doi.org/10.1073/pnas.0601602103
Quantum Interference and Shape
Detection
Davi Geiger and Zvi M. Kedem(B)
Department of Computer Science, Courant Institute of Mathematical Sciences,

New York University, New York, USA
{dg1,kedem}@nyu.edu
Abstract. We address the problem of shape detection in settings where

large shape deformations and occlusions occur with clutter noise present.
We propose a quantum model for shapes by applying the quantum path
integral formulation to an existing energy model for shapes (a Bayesian-
derived cost function). We show that the classical statistical method
derived from the quantum method, via the Wick rotation technique,
is a voting scheme similar to the Hough transform. The quantum phe-
nomenon of interference drives the quantum method for shape detection
to excel, compared to the corresponding classical statistical method or
the statistical Bayesian (energy optimization) method. To empirically
demonstrate our approach, we focus on simple shapes: circles and ellipses.
Keywords: Shape · Hough transform · Interference

Statistical methods · Energy minimization · Wick rotation
1 Introduction
We are given an image I containing multiple shapes, some of which are possibly
occluded and/or deformed, and there is possible clutter noise data. To detect
the shapes in I, we propose a new method, which is rooted in quantum theory.
The main advantage of the quantum method over existing methods is its high
accuracy and resistance to deformations thanks to the quantum interference
phenomenon as described, e.g., by Feynman and Hibbs [1] and Feynman [2].
The input data is specified by a set of N feature points X = {x1 , x2 , ..., xN }
in RD . Here we focus on 2-dimensional data, so D = 2. An image I may be
specified by the set of points, as shown for example in Fig. 1. Otherwise, feature
points can be extracted from I, using, for example SIFT features, HOG features,
or maximum of wavelet responses, etc.
1.1 Our Approach

We start with an energy-minimization view of shape detection. We consider a
standard error criterion of the energy model for fitting a shape to the data using
c Springer International Publishing AG, part of Springer Nature 2018

M. Pelillo and E. Hancock (Eds.): EMMCVPR 2017, LNCS 10746, pp. 18–33, 2018.
https://doi.org/10.1007/978-3-319-78199-0_2
Another random document with
no related content on Scribd:
his master, and had got about half way to the shore when the girl
gave a terrible shriek, and in sight of her parents disappeared
beneath the water.
Hearing the shriek the dog turned at once and swam out again
towards the place where he had left his young mistress. The poor
father sprang into the boat and pushed it off into the stream, while
her mother, poor soul, stood wringing her hands, and moaning
piteously on the shore. With every nerve strained her husband pulled
after the dog, which seemed to be swimming towards the entrance of
the lake barking all the time. The poor man knew that his child had
been seized by a shark, but still he followed the faithful animal, the
mother’s frantic cries ringing in his ear, and urging him on. When he
got to about one hundred yards from the dog, he could see that
between its barks it was trying to snap at something beneath the
water. Suddenly it gave a terrified howl, and seemed to be in
difficulty. When he reached it, he found one of the poor dog’s hind
legs had been bitten clean off, and the faithful beast was drowning.
There was no sign or trace of his daughter, and after rowing about
for some time he was forced to give up the search and return to his
grief-stricken wife. It was an awful fate for one so young and full of
life to meet her death in sight of the door of her home and loved
ones, and they powerless to help or save.
But the lake and those who lived near it had not seen the last of
the shark, as I found to my cost a few weeks later. As I have already
said I was passionately fond of swimming, and often, when my day’s
work was done, I would swim across the lake to Southland, rest
awhile and swim back. After the tragedy at Boyd’s I was very careful
to keep a good look out, but on this particular day it had been
fearfully hot, the temperature being I should think 90 degrees in the
shade. I had been out with a riding party looking after the horses and
by the time we returned and I had settled the horses for the night, I
felt pretty fagged, the water looked very tempting, so in I went, and
struck out for Southland, landed, had a run along the beach, and
started to swim back to Belmont.
I had got about half way over when I noticed quite a commotion
among the people at Mr. Williams’s, who usually at this time of the
day were sitting out on the lawn, and in the grounds after dinner, but
now I noticed that they were running about and pointing to me, while
some were looking through spy-glasses. I thought they were
admiring my good swimming for I was a fairly good swimmer in those
days, and a bit proud of it, too, so, youth-like, I put on a spurt, just to
show off a bit. In doing so I kicked something with my right foot,
which hurt my toes considerably, but thinking it was some floating
rubbish I swam on, never giving a thought to it, or troubling to
ascertain what it was. The number of people out on the hill was
increasing, until I should think everybody belonging to the
establishment was watching me, and waving their handkerchiefs. I
began to feel quite flattered by the attention they were giving to me,
and wondered what there was in my swimming to cause it. Suddenly
I saw several men break away from the rest, and rush down to the
boatshed on the beach. This rather surprised me, they surely did not
think I was drowning. To relieve their fears, as I thought, and give
them a change of programme I turned over and commenced to swim
along leisurely on my back. In doing so I happened to cast a glance
behind me, and “Oh God in heaven help me!” I cried, for there, not
twenty feet from me was the ugly dorsal fin of a monstrous shark.
This was the cause of the commotion on the hillside, not admiration
of my swimming, but to warn me of the fate that was closing around
me like a net.
I nearly sank with fright, there seemed to be no escape, I trembled
from head to foot, I knew now only too well what I had struck with my
foot. Quickly I turned over on my breast and struck out with every
ounce of strength that was in me. In fancy I could already feel the
loathsome brute’s teeth tearing my flesh. “Dear God,” I prayed, “not
that, not a terrible death like that.” Thoughts of home and those left
behind crowded into my mind and like a flash my whole past life
passed before me. Just then to my joy I saw two boats coming to my
rescue. Oh, how slowly they seemed to be coming towards me. My
heart was thumping like a sledge hammer in my breast. Then, oh
horror, I felt something touch my right leg, and I lacked, splashed,
shouted and almost fainted with fright, but still I swam on, and the
shark dropped behind for a moment or two. Then, with another
prayer to God for help, my nerve seemed to come back to me, and I
swam on, but this time it was with the madness of despair. A few
seconds afterwards I saw the dark shadow of the shark pass just
beneath me and dart ahead and then go rushing past on my right
side. I knew then that my time was come and death was hovering
over me, the boats were drawing nearer, in despair I shouted “Help,
help, for God’s sake help me. Shark, a shark!” and I saw them bend
to their oars and pull for all they were worth. The man who was
steering the first boat sang out “swim to the left, to the left,” and I did
so with all my might. The first boat dashed past me on my right side,
and the second ran up alongside of me, and in an instant I was
grasped by the arms and legs, and almost jerked into the boat. And
not a moment too soon, for someone in the forward boat called out
—“Look out! Look out!” and the next minute the shark darted past
our boat, his dorsal fin sending up quite a spray as he tore through
the water, swishing his tail in anger and madness at the loss of his
prey.
“By Jove, you’ve had a narrow squeak, young fellow,” said one of
the gentlemen as he mopped his perspiring face. “I never thought to
see you reach land again.” I am glad I was not the one in the water, I
should have sunk in sheer fright. It was simply marvellous that the
brute did not attack you; we saw him go up to you more than once,
and I for one gave you up for lost.
It was not until some hours later that I was able to thank my
preservers, for when we reached Belmont I collapsed, and was put
to bed by some of our men who were kindness itself. Mr. Williams
came and gave me a dose of medicine that sent me off into a deep
sleep which lasted some hours. When I awoke it was morning, and I
felt quite myself, but the nearness to which I had approached death
left its mark upon me. I never attempted to swim across to Southland
again, but was content to bathe near the beach and then kept a
sharp lookout for my enemy the shark.
I stayed at Belmont twelve months, and they were some of the
happiest in my life; it was a happy, healthy, outdoor life, and suited
me in every way. Every day brought its work and its pleasure, now it
would be yachting, then it would be hunting wild horses, kangaroos
and wallabies during the day, and at night we would hunt the
opossum in the gum trees. The time seemed to fly.
One day while gathering apricots in the orchard, I noticed a black
cloud rising across the lake. I looked and looked at it, but could make
nothing of it. I had seen many squalls in my time, and had learnt a
good bit about them, but this puzzled me. It kept altering its shape in
a strange unaccountable manner, and yet one point seemed to be in
advance of all the rest. I was watching it closely and turning it over in
my mind as to what it could mean, when I heard the whistle that
called us back to the house. When I got there I was told that the
black cloud, as I thought it, was in reality thousands and thousands
of flying foxes, who were migrating south and if they were not
frightened away they would steal every particle of fruit in the orchard.
Now these destructive creatures are one of the scourges of the
fruit orchards, and are called by some the fruit bat and by others the
flying fox, and are, I think, only found in the warmer regions of the old
world. The bats are the only members of the Mammalia which
possess the power of true flight, and the way this is accomplished
was pointed out to me afterwards by one of the visitors staying at the
house. The wings of the bat have been formed by the modification of
the fore limbs, the finger bones having become lengthened to serve
as a support to a thin web of very sensitive skin extending outwards
from the body, not unlike the covering of an umbrella. The hand of
the bat is quite a unique organ. When resting they attach themselves
to the boughs of the trees by these fingers and in the dusk might
easily be taken for fruit.
I was handed a cumbersome wooden rattle, and told to give it full
force all the time they were passing over head, and a terrible noise it
made. The master and several men were discharging guns at the
same time. So great were the numbers of the flock that they were
quite ten minutes in passing over.
Now ten miles from Belmont there was a large fruit farm owned by
a Mr. Warner. One large orchard was entirely of quinces for the
Sydney market. He estimated that the crop would yield quite eleven
tons of fruit. Unfortunately for him he and his family were away at the
time, and those left in charge were either careless of their
responsibilities, or had not noticed the cloud coming their way, but
this was a grand opportunity for the foxes, who live on fruit. Down
they swooped and stripped both that orchard and the garden around
the house of every particle of fruit that was growing. The whole
year’s crop vanished in a few minutes, the cloud swept on, and no
cheque was left to pay for the fruit.
Another very interesting sight to be seen at that time of the year
was the migrating south of immense flocks of black swans. There is
no more beautiful sight than to see them flying overhead for then the
pure white of their under pinions is exposed and glistens like snow in
the sun, spreading themselves out in the shape of a triangle, with
one bird leading in front. At no time do the other birds approach the
leader, but each keeps its own place like a regiment of soldiers with
an officer leading them on. They fly very high, and seldom, if ever,
come within shooting range when migrating.
CHAPTER XX
Snake Stories—Two Brave Girls
In the previous chapter I have spoken of Mr. Warner and his fruit
orchard. The old saying that misfortunes, like blessings, never come
singly was verified in their case. One day, not long after the previous
incident, Miss Warner was in the orchard pruning some young trees.
As she moved away from a tree she had finished with, she felt a
sharp slap on her right thigh and knew at once she had trodden on a
snake, which animals are very numerous in that part of the country.
Her bush training had taught her that it is safest in such a case to
stand quite still until you know which end of the snake you are
treading on. Perchance it may be on its head, and if so you can
easily dispatch it; and if, unfortunately, you are on its tail, well, no
earthly power can save you from being bitten before you can jump
clear. As ill-luck would have it, Miss Warner had trod upon the
snake’s tail and it had retaliated by digging its poisoned fang into her
thigh. It was just about to make another stab, when she struck at it
with her pruning knife, cutting it in two and killing it instantly. Then
she coolly wiped the point of the knife on her dress, and deliberately
made a cross cut into her thigh where the snake had bitten her.
Then, while the blood was spurting from the wound, she called out to
her father who came running to her, knowing by the sound of her
voice that something was the matter. In a few seconds he had torn
his handkerchief into strips and tightly bound the leg above and
below the wound. Then, saddling his horse and one for the wounded
girl, they set off on their twelve mile ride to Newcastle.
You can imagine what that ride was like to Miss Warner, up hill and
down dale as fast as the horses could go, the great gash in her leg
was very agony of pain. It required endurance, nerve, and pluck—
qualities our colonial bush-reared maidens are in no way deficient in.
Her father, too, had a very good reason for letting her ride on
horseback. Snake poison, as is well known, causes sleepiness,
which, if succumbed to, knows no waking. Had Mr. Warner taken her
in a trap, he would not have been able to prevent her from falling
asleep, so had put her upon her horse, hoping to reach Newcastle
before the poison took effect. They had ridden about nine miles
when Miss Warner became very faint, and could scarcely keep her
seat on her horse. Just then they met a young man riding out to the
lake district, and as soon as he heard the state of affairs, he at once
turned his horse round and went back to Newcastle to obtain a
doctor.
Fortunately the doctor was in, he immediately ordered his carriage,
and taking his instrument case and some antidote to counteract the
snake poison, he set out and met Mr. and Miss Warner just outside
the town. He at once helped her off her horse, then she was taken
more dead than alive to the nearest house, and all that medical skill
could do and suggest was done for her, but it was fully three months
before she was able to return home, and then looked a perfect wreck
in comparison to her former robust self. But the brave spirit was in no
way quenched by the suffering she had gone through.
The other brave girl was the daughter of an old boat-builder name
Parrell, who lived on the banks of Lake McQuarrie, with his wife and
his one daughter, Jennie, who was seventeen years of age, and had
been born in the bush of Australia. Like most of the girls reared in the
bush, she was a fearless horse-woman, a strong swimmer, a first-
class shot with a revolver, as cool as a cucumber at all times, and, to
crown all, one of the prettiest girls in those parts, at least I thought
so, and many a young fellow beside, but Jennie would have none of
us, but would laugh and shake her head at our attempts to oust each
other in our efforts to win her favour, but it was all to no purpose,
Jennie remained heart-whole, and we sighed in vain.
The house they lived in was built of weatherboard, and stood at
the mouth of a small creek, where it emptied itself into the lake. All
the rooms were on the ground floor, and were divided by a thin
partition about six feet high, thus making two bedrooms and a good-
sized living room, all furnished very comfortably, the beds used being
the ordinary trestle camp beds. One night I had gone over to try and
get a chat with Jennie—the night was hot and sultry, there was not a
breath of air moving, the day had been one of the hottest we had
had, the mosquitoes were terribly vicious. When I got there I found
that Mrs. Parrell and Jennie, unable to bear the heat and mosquitoes
any longer for that day, had gone to bed under their mosquito
curtains, so Mr. Parrell and I sat on a log outside the cottage door.
There was some satisfaction in being near the object of my
admiration at any rate, and her father always gave me a very hearty
welcome, which was something in my favour, at least I thought so.
They had been in bed about two hours—while we had been yarning
—when Mrs. Parrell heard her daughter quietly calling her.
“What do you want, lassie?” she replied. “Can’t you sleep?”
“No, mother, there is a snake in my bed,” the girl answered. “It is
lying on my naked legs. I dare not move or it will bite me. Tell my
father quick.”
The terrified mother needed no second bidding, springing out of
bed she rushed to the door and told her husband.
“Hush, mother,” he said quietly, laying his hand on her arm, “don’t
make a sound, or you will do more harm than good.” Then quietly he
crept into his own room and speaking softly, said:
“Where is it now, Jennie? Don’t be afraid, lassie.”
“On my stomach, father, and it is working up to my breast,” she
replied in a low tone.
I thought, of course, the father would have rushed into the room,
and at one blow would have killed the reptile. Not so, the old man
had learnt by experience a better way than that.
The door of the bedroom was just at the foot of the girl’s bed, and
any noise, either by opening or shutting it, would have startled the
snake, and, perhaps, made it plunge its poisoned fang into Jennie’s
body, and thus have ended her young happy life, but the father knew
just what to do under the circumstances. Going to a cupboard that
stood in the corner of his room, he took down an old violin, then
crossing, without a sound, over to the end of the room farthest away
from the bed, he drew the bow lightly over the strings making a soft
plaintive sound. The moment the snake heard the noise, it poised its
head up on the girl’s breast, and as the soft plaintive notes floated
about the room it began to wriggle along towards the foot of the bed
in the direction of the sound. The brave girl kept her nerve and
presence of mind marvellously. She knew full well her very life
depended on it, for the slightest movement on her part, while the
snake was on her body, would have been fatal. But as soon as she
felt the snake slip off the bed she sprang up and out of the way. The
snake, when it heard her move, made a dart for a small knot hole in
the planking of the floor, through which it had entered the room, but it
had Jennie to reckon with, and before it could reach it she had
thrown a pillow upon it, then her father rushed in and dispatched it
with a stick, he brought it outside, and we measured it, and found it
was three feet nine inches in length, and was about as thick as a
broom handle. It was a carpet snake, one of the most deadly
enemies to be encountered in bush life. A few minutes afterwards,
Jennie and her mother having dressed, came outside, and I could
not help telling her how much I admired her nerve and courage, and
asked how she knew it was a snake?
“Why because it was so cold and clammy,” she replied, making a
wry face. “A snake is always cold, no matter how hot the place may
be where it gets into.”
The next day I had occasion to go to Newcastle on some business
for Mr. Williams and looked forward with some pleasure to the twelve
mile ride through the bush on the back of old Blunderbuss, a horse
that had once belonged to that king of bushrangers, Captain Morgan,
and after having passed through several people’s hands, had
become the property of Mr. Williams. We had only gone a few miles
on our way, when I began to notice that what little air there was had
an acrid smell about it, and every step it seemed to become more
close and stifling. Blunderbuss, too, began to twitch his ears and
sniff. Then a little farther on the cause of it burst upon me—the bush
was on fire—for a second or two I pulled up, and looking round, to
my amazement the trees near by began to blaze and crackle, and
there I was with fire in the front of me and sweeping along the path
over which I had come. There was nothing to be done but try and get
through to where it was already burnt, so putting spurs to the horse
we began our race for life. Hotter and hotter grew the blast, as I
urged the panting beast forward. Kangaroos leaped out from among
the burning scrub, and fled onward, a greater enemy than man was
upon them. Then a drove of wild horses went madly past, while the
birds shrieked and fluttered above our heads, powerless to help
themselves. Onward, still onward we went, the flames hissing and
crackling over our heads, and still we seemed to be no nearer to the
outlet, no track was visible, and all around was the thick hot air.
Blunderbuss, too, seemed to be pulling in quite an opposite direction,
then it flashed across my mind that perhaps the instinct of the poor
beast would lead us into safety, so giving him his head and hoping
for the best I let him go his own way. In an instant he wheeled almost
round, and sped onward towards what seemed a veritable inferno of
flaming trees, with head outstretched and feet scarce touching the
hot earth he dashed through the blazing mass, and presently to my
joy I saw that he was making for an old bridle path that led around
the cliffs to Newcastle, and thankful I was, for had it not been for the
instinct of Blunderbuss we might both have perished miserably. As
we slackened our speed now that the danger was past, I turned to
look at the sight we were leaving behind. Overhead was a dusky
canopy of thick black smoke, there stood the black bare trees
between us and the still raging fire, farther away the sparks were
dropping from the thick smoke like a hailstorm, then again it looked
like a moving curtain of crimson smoke, with the falling and blazing
twigs and small branches, like coloured fireworks—all blue, red and
yellow, while tongues of flame licked up the dry scrub and grass.
It was a grand sight to look at when you were safely out of it, but
we were both in a sorry state, for our hair was singed and our skin
blistered where the flames had touched us. The next day
Blunderbuss and I returned to Belmont, needless to say we gave the
still burning bush a wide berth, and reached home before nightfall.
The fire was burning for the best part of a week, and when some
weeks afterwards I passed that way again, I marvelled how we had
ever passed through that furnace of flame.
CHAPTER XXI
Widow Smith’s Pig or “Barkis is Willin’”
At Belmont, our next door neighbour, Mrs. Rebecca Smith, lived a

quarter of a mile away, but this was thought little of in the Australian
bush, where, frequently, your nearest neighbour lives five miles off.
Mrs. Smith was a buxom, good-looking widow, and she knew it. She
had a good weatherboard house and a large patch of freehold land,
with a nice well-kept vegetable patch, and she kept a good number
of pigs, which were noted for their size and quality.
Now there was one pig in particular that she took a great pride in,
though, at the same time, it was the plague of her life. It was very
fond of roaming, and would not confine itself to the large paddock
where the stye was placed, but preferred other people’s patches,
especially if they happened to be vegetable patches. It would travel
miles away from home at times; a gate was no obstacle to it—it
would burrow under it in no time—and it was really surprising what a
small hole it would squeeze itself through.
Mrs. Smith’s garden was alongside the main road leading from
Newcastle, New South Wales, to the Wallsend coal mines. The fence
of the garden was a low one, and everyone passing could see how
clean and tidy she kept it. A good number of people passed by on
horseback and in vehicles, and whenever Mrs. Smith heard anybody
coming along, she would stand at the door facing the road and
smilingly bid them good-day. The widow was about forty, strong,
healthy, and fair to look on—her fine, fat, round arms were a sight to
see—and many a man passed the cottage just to catch a glimpse of
the bonny widow standing at the door.
Now, one Saturday evening this torment of a pig got out of the
stockyard, burrowed under the fence, and got into the widow’s
garden patch, where it made terrible havoc of the beds. It turned up
a lot of vegetables, took a bite of a dozen water melons, tore the
bark off several fig-trees, and bit the stem of a lovely passion-fruit
vine right through, killing the plant.
The following morning—Sunday—when Mrs. Smith saw the
damage that had been done, she vowed there and then that the pig
should die. But it could not be found anywhere, until about ten
o’clock she heard it in the bush at the back of the house. She
immediately sent her two young sons and a workman to drive it into
the stockyard, and when they had done so, Mrs. Smith told her son
to get his rifle and shoot the pig at once. The boy slipped into the
house, and returned with the gun in his hands, and made over
towards the pig, who was crunching some turnip.
“Shoot him in the head, Willie,” cried the widow, as she stood in
the gateway to prevent the pig from escaping.
As the boy approached, the pig looked up, and, seeing the shining
barrel pointing at his head, scented danger, and turned right-about
ready to bolt. Just then the boy fired, and the shot entered the pig at
the wrong end. With a piercing squeal and a grunt, it made a dart for
the gate, turning the boy head over heels in its mad rush. Mrs. Smith,
seeing the pig coming straight for the gate, and having no time to
close it, spread out her legs to block the passage with her skirts.
(Mrs. Smith made her own clothes and put good material into them).
The pig, nothing daunted by Mrs. Smith’s presence, made a dash
between her legs, his snout ripped a hole in the skirt, his head went
through, and Mrs. Smith was jerked off her feet, the skirt being
bridled over the pig’s head. The widow fell flat on her face on the
pig’s back, with her head to its tail. Finding herself being dragged
along, she threw her arms around the pig’s body and clipped its neck
with her legs, and held on like grim death, a second Mazeppa.
On dashed the pig squealing, Mrs. Smith screeching, the boys and
the workman yelling as they raced after them. The pig, thinking to
escape his pursuers, left the road and struck into the bush, and soon
the covering was torn off Mrs. Smith’s back by the bushes. But still
the pig rushed on, and still Mrs. Smith held on to the pig, and by this
time quite a number of people had started to follow.
Now, there was no church or chapel at Belmont, but a sprinkling of
people used to gather together and hold a Methodist meeting in a
little school-room where a one-legged schoolmaster taught the
children of the district reading, writing and arithmetic. On Sunday one
of the men read Spurgeon’s sermons out of a book, and each in turn,
as the spirit moved them, prayed loud and long for forgiveness of
their sins which were few, and of their neighbour’s, which were many.
On this particular Sunday morning a local preacher had come from
Minmi to hold service in the little schoolroom. He had opened the
Good Book, wiped his face, coughed, and given out his text: “Repent
ye, for the kingdom of heaven is at hand.” He paused for a moment
to let the words take effect. Then, in a sterner and more determined
tone, he repeated the text: “Repent ye, for the kingdom of heaven is
at hand.” There was silence for a moment.
Then an unearthly noise was heard outside. Pig squealing, women
screeching, men shouting. Everyone in the meeting was on his feet
in an instant. If this was the kingdom of heaven—well, they all
seemed mighty scared to meet it. The preacher especially—for he
dropped the Book and made for the door first.
As they all got outside, the pig dashed past with Mrs. Smith on its
back, holding on for dear life, and I do not know which was making
the most noise—Mrs. Smith, or the pig. Few of the neighbours could
recognise the rider as she dashed past, for they were not acquainted
with the features presented to view, but they all took up the chase,
the preacher taking the lead.
The pig was making a bee-line for the tea-tree scrub about a
quarter of a mile from the school-house; but to get to the scrub it had
to cross a small creek about ten yards wide and three feet deep. As
soon as it took the water the weight of its rider turned it completely
over, and so released the skirt from its head. But before it could get
away Mrs. Smith had seized its leg and held its head under water
until help arrived, when one of the men despatched it, and a trap was
sent for to take it home.
And so ended Mrs. Smith’s Sunday morning ride on the pig. But
there was a sequel to it. The local preacher, who witnessed the
famous ride, was apparently struck with the widow’s proportions
where visible, and became very attentive and solicitous for her
welfare. He walked home with her and left his sermon for a future
day, sympathising with her in her trouble, and pouring soothing
unction on her outraged feelings in a true Methodistic manner. The
widow was so pleased with his sympathy and attention that she
invited him to stay to dinner, which invitation the preacher was very
pleased to accept.
During dinner the preacher made himself so very agreeable that
the widow Smith opened her heart to him and told him how very
lonely it was for her at times, since her poor husband died.
“But the dear man left me a good home and fifty acres of freehold
land, so I must not complain,” she added; “only sometimes I do feel it
lonesome.”
And the preacher took it all in with an eye to future events. He had
a nice little store at Minmi, but no land, and he thought how nice it
would be if the two were joined together as one concern.
He did not stay late that day, as he had twenty-seven miles to ride
through the bush to his home at Minmi, but he asked permission to
come again, and this was freely granted.
Very frequently afterwards the preacher was seen riding down to
Widow Smith’s homestead. It was a case of “Barkis is willing” on
both sides.
Although his first view of the widow had been a hasty view, still it
had impressed the preacher so much that he soon persuaded her to
change her name, and the little store at Minmi knew him no more.
I had now been with Mr. Williams at Belmont about twelve months
when the old restless spirit began to take possession of me again.
The sea was calling me, with its wild days of stress and storm, or
hardships and peril, and would not be denied.
When I told Mr. Williams he did his best to dissuade me, but the
call had come, so packing up my belonging’s, which included a good
testimonial from Mr. Williams and a beautiful Bible from Miss
Williams, I said good-bye, and set out for Newcastle, hoping to get a
ship from that port.
CHAPTER XXII
A Dangerous Enterprise
I arrived at Newcastle and was fortunate enough on the first day to

ship as able seaman on the barque “P.C.E.”—Captain Law—bound
for Noumea, New Caledonia, with a cargo of coal. My wages were to
be seven pounds a month: I felt that I had “struck oil” at once for I
had never heard of seamen getting such fabulous wages. I found
out, however, that colonial owners paid good wages, and required
good workers for their money, no limejuice methods would suit the
coast trade. But I was young, strong and healthy, and I got along
famously with both officers and men. Captain Law was as rough and
unpolished a specimen of the British seaman as it was possible to
meet, but a sailor from the crown of his head to the sole of his feet.
The mate was a most illiterate man, coarse in speech and manner,
and proud of his manners, or lack of them, but a good seaman, and
utterly fearless. It was generally believed on board that he would
have gone to sea on a plank if the pay had been good enough.
We had a fine passage of eight days to Noumea, New Caledonia,
which is a penal settlement under the French flag. At Noumea there
are a large number of prisoners both political and criminal out on
parole—many of them are in business and doing well for themselves,
but they cannot, of course, leave the island, and naturally there are
always a number of daring and eager spirits ready and willing to run
any risk to obtain their freedom, if they can get anyone to assist
them. The political prisoners, many of whom are well off, or have
friends who are well supplied with money for such a purpose, are
always on the qui vive for such a chance.
Whilst unloading our cargo, the captain spent most of his time on
shore at one of the saloons. Captain Law was very fond of a glass of
grog when on shore, but, to his credit be it said, he never touched
liquor at sea. One day while at the saloon Captain Law was
approached by a Frenchman named Balliere as to his willingness to
assist some prisoners to escape, and a good round sum was offered
on account of Henri Rochefort, French Communist, who was then on
Ducos Peninsular. Captain Law had heard of the famous journalist
and was rather in sympathy with him, but the sum of ten thousand
francs won his sympathy entirely. This was the sum Henri Rochefort
had offered through Balliere. Captain Law accepted the offer,
stipulating that no other member of the crew should be told for fear of
them informing the officials, by whom a large reward would be paid
for such information. Altogether it was a dangerous undertaking, for
had he failed, or been caught in the act, the vessel would have been
confiscated and all the crew imprisoned, perhaps shot. It was,
therefore, arranged that the exiles should board the vessel
themselves at their own risk, so that in the event of failure the
authorities would not confiscate the vessel. But there was another
side to the question which Captain Law, apparently, had everlooked
—that he was risking the lives and liberty of his officers and crew
without their consent or knowledge, which, in law is a criminal
offence.
There was no difficulty for the three exiles at Noumea—Jourde,
Balliere and Granthille—to get on board the “P.C.E.” They were out
on parole and were in business in Noumea, but for Henri Rochefort,
Oliver Pain and Pascal Grousett, on the Ducos Peninsular, it was a
difficult and dangerous undertaking. But these men were quite
accustomed to dangerous undertakings, their very lives had been
spent among risks and perils, and they were prepared to do or die.
Better far better to die in a struggle for freedom than to live a
hopeless, lingering life, desolate and alone in their sun-scorched
island prison. All of them had wives and families in their beloved
France, and to see their faces again they were prepared to risk all,
feeling that they were bound to be the gainers, whichever way things
went. If they were caught, they would be shot—well, their misery
would be over—if they succeeded they would be free. Free men
again—Oh! how their hearts throbbed at the very thought of it! The
joy of liberty is seldom appreciated until it is lost.
But to return to the exiles on Ducos. It was not possible to
approach the bay in a boat without being seen by the sentry on duty.
Now, just off the peninsula there was a small volcanic rock about one
mile from the beach and in a line between Noumea and the Bay of
Ducos. On the side facing Ducos was a mass of coral reefs, the
other side had a bit of sandy beach, on which the surf breaks heavily
at times. The waters around the island were infested with sharks,
this, to those in authority, greatly lessening the possibility of
prisoners escaping.
Now on the day the barque “P.C.E.” finished unloading at Noumea,
Balliere had sent word to Henri Rochefort that a boat would be off
the rock at 8 p.m., and he and his two friends Oliver Pain and Pascal
Grousett were to make their way to the rock. All that day they had
been talking about it, and as they talked the great ugly sharks were
lazily swimming about in front of them, a sight that was enough to
make the stoutest heart quail and quake. Yet, in spite of this, they
resolved to swim out to the rock when the time came.
As night drew near the exiles grew anxious lest anything should
happen that would defeat their plans, and towards 7 o’clock their
hearts sank, for the sky grew dark and threatening, and a fierce
tropical storm broke over their heads, the thunder, lightning and rain
being truly terrific, while it lasted. But strange though it may seem,
although the exiles at the time did not realize it, the storm was a
Godsend to them. First it drove the lookout sentries in, and secondly
no fish will remain near the surface during thunder and lightning, they
always sink to the still calm water below. Thus were the exiles
spared a battle with these monsters of the deep. Just before 8
o’clock they stripped and walked down to the beach. But here a fresh
difficulty presented itself. The night was so dark they could not see
the rock, but there was no time to hesitate, somewhere out there lay
the means that was to help them to regain their freedom, so into the
water they went. Rochefort, ever a leader, first, and they struck out in
the direction they knew the rock to be. After swimming for some time,
expecting every moment to find themselves in the jaws of a shark,
and seeing no signs of the rock, their hearts began to sink for fear
they had missed it, then suddenly they found themselves among the
kelp on the reef, so with thankful hearts they landed and crawled
around to the other side.
They had only just got there, when the sound of muffled rowing fell
upon their ears. Holding their breath they waited, for perhaps they
had already been missed, and the sound they heard might be the
guard boat. They did not know if it were friend or foe. Presently a
voice called out cautiously.
“Are you there?”
They were afraid to answer for fear they had been followed, and
the voice seemed strange to them.
Again out of the darkness came a voice:
“Are you there? It is I, Granthille!”
Thus assured that it was the voice of a friend the three fugitives
replied, and were told to swim out to the boat as there was too much
swell on the reef to risk taking the boat inside. This they did, and
were soon assisted into the boat, and supplied with dry clothes. Few
words were spoken, their hearts were too full for words, and there
was still dangerous work before them, this was but the first step
towards freedom. But a silent clasp of the hand was given, and
understood by all.
The boat was at once headed for Noumea. The darkness was
intense, but a faint glimmer of the harbour lights could be seen. Do
what they would, they could not prevent a long phosphorescent light
streaming behind the boat, and as the night was so black, this
brilliant light was very conspicuous and might possibly betray them to
any observant watcher. Their whole souls were full of doubt and fear
and racked with anxiety as to the issue of the attempt, for more than
one of them had resolved never to return alive to their late prison.
As the boat drew near to the vessels they had a truly marvellous
escape from detection, but thanks to a heavy downpour of rain,
which drove all watchers into shelter. Before leaving Noumea with
the boat, Balliere had strolled down to the harbour and taken note of
the position of the vessels, and the “P.C.E.” in particular to enable
them to find her in the dark, but during the short storm the wind had
changed, and had altered the position of the vessels at anchor,
bringing a small French despatch-boat into the position previously
occupied by the “P.C.E.” The Frenchmen, not being seamen, did not
notice the change, and pulled over alongside the French boat. One
of the exiles actually had his foot on the ladder to climb on board,
when they heard voices above them speaking in French. At once
they realized the extreme danger that threatened them: with faces
sternly set, but with their hearts in their mouths as it were, for they
scarcely dared to breathe, they let the boat drop quietly astern, not
knowing whether they had been seen or heard. Lightly they dipped
their muffled oars and rowed over to the next vessel, which proved to
be the one they were seeking. The side ladder was hanging over the
side, quickly and quietly the exiles climbed on board. The steward
was on deck aft and stared with astonishment at the six big fellows
as they came over the rail. Before any explanation could be asked or
given, Captain Law made his appearance, the steward was ordered
to his room, and the captain pointed to the after hold, where the
fugitives all defended and stowed themselves away among a lot of
old ropes, tarpaulins and cargo gear. The hatch was closed, but
afterwards re-opened to disarm suspicion, should an official by
chance come off. The only persons who knew the escaped prisoners
were on board were Captain Law, the steward and myself, as I was
keeping anchor watch. The plug of the boat was taken out, and a
quantity of stone ballast that had been hoisted on deck during the
afternoon, was put into her and she was sunk alongside the ship. It
was not safe to let the boat go adrift, suspicion would have been
aroused, and a muster of prisoners ordered, and then all hope of
getting away would vanish.
At daylight the pilot came off to take the vessel out, but
unfortunately the wind had died away and he refused to get the ship
under weigh. The delay in sailing was very serious. Both Captain
Law and the fugitives were in a terrible state of anxiety and
suspense, expecting every moment a search party would come off to
overhaul the vessel. Up to the present the alarm gun on the hill had
not been fired, so far we knew that the escape of the prisoners had
not been discovered, but at the same time we also knew that the
moment the escape was discovered the signal gun would be fired
and that no vessel would be able to leave port without being
searched.
At noon a breeze sprang up, but it was in the wrong quarter to
enable us to get through the usual passage in the reef, however,
Captain Law insisted on the pilot taking the vessel out of the harbour,
so the anchor was lifted and we stood out towards the entrance.
Once outside, the pilot was dismissed and we made all sail, and with
a fair wind steered for the Balearic Channel, coasting along the
Ducos Peninsular, where the poor exiles had spent so many weary
months of imprisonment. As the evening drew on the wind increased,
and before dark we had cleared the coast and waters of New
Caledonia.
The captain then called the Frenchmen on deck and told them that
thus far they were free. Henri Rochefort was the first to come on
deck, and a more pitiable looking object it would have been hard to
find. He stood quite six feet in height and was as thin as a lath, with a
long head and a very prominent forehead. The only clothes he wore
were an old pair of pants, a shirt, and a seaman’s sou’wester on his
conspicuous head, and he was covered from head to foot with coal
dust from the vessel’s hold, where they had been lying concealed. I
doubt if his best friends would have recognised, in the miserable-
looking object before them, the most famous French journalist of the
day, whose writings and doings had stirred France to the core. They
were all more or less suffering from sea-sickness, and were taken
into the cabin to have a clean up and make themselves presentable.
There was only one who could speak a few words of English, but the
warm clasp of the hand and the happy look in their eyes spoke
volumes of gratitude for their deliverance.
All through the following day the strong north-east wind kept up,
and running before the breeze with every stitch of canvas set, we
soon put a good distance between us and the Island. The crew did
nothing but work the ship and keep a lookout for overtaking vessels.
The only one we were afraid of was the despatch-boat that we left at
anchor at Noumea when we sailed.
The second and third day remained hazy with a strong wind. The
exiles rarely left the deck, but would pace to and fro for hours, so
anxious and fearful were they of recapture. All day long they would
scan the horizon, their lips muttering the thoughts that would not be
kept back, thoughts of home and loved ones. Surely fate would not

Textbook Energy Minimization Methods in Computer Vision and Pattern Recognition Marcello Pelillo Ebook All Chapter PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Textbook Energy Minimization Methods in Computer Vision and Pattern Recognition Marcello Pelillo Ebook All Chapter PDF

Uploaded by

Copyright:

Available Formats

Energy Minimization Methods in

Computer Vision and Pattern

Energy Minimization Methods in Computer Vision and

Computer Vision Pattern Recognition Image Processing

Image Processing Computer Vision and Pattern

Handbook Of Pattern Recognition And Computer Vision 6th

Artificial Life and Evolutionary Computation 12th

Pattern Recognition and Computer Vision First Chinese

Pattern Recognition and Computer Vision First Chinese

Pattern Recognition and Computer Vision Third Chinese

Energy Minimization Methods

Energy Minimization Methods

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Library of Congress Control Number: 2018937378

© Springer International Publishing AG, part of Springer Nature 2018

Printed on acid-free paper

February 2018 Marcello Pelillo

Andrea Torsello Università Ca Foscari, Italy

Chiotellis, John Piccirilli, Marco

Clustering and Quantum Methods

Ising Models for Binary Clustering via Adiabatic Quantum Computing . . . . . 3

Quantum Interference and Shape Detection . . . . . . . . . . . . . . . . . . . . . . . . 18

Structured Output Prediction and Learning for Deep Monocular

Dominant Set Biclustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

Bragg Diffraction Patterns as Graph Characteristics . . . . . . . . . . . . . . . . . . . 62

Motion and Tracking

Variational Large Displacement Optical Flow Without Feature Matches . . . . . 79

Temporal Semantic Motion Segmentation Using Spatio

Depth-Adaptive Computational Policies for Efficient Visual Tracking . . . . . . 109

Multiframe Motion Coupling for Video Super Resolution . . . . . . . . . . . . . . 123

Illumination-Aware Large Displacement Optical Flow . . . . . . . . . . . . . . . . . 139

Location Uncertainty Principle: Toward the Definition of Parameter-Free

Image Processing and Segmentation

Autonomous Multi-camera Tracking Using Distributed

Nonlinear Compressed Sensing for Multi-emitter X-Ray Imaging . . . . . . . . . 189

Unified Functional Framework for Restoration of Image Sequences

A Convex Approach to K-Means Clustering and Image Segmentation . . . . . . 220

Luminance-Guided Chrominance Denoising with Debiased

Optimizing Wavelet Bases for Sparser Representations . . . . . . . . . . . . . . . . 249

Bottom-Up Top-Down Cues for Weakly-Supervised Semantic

Vehicle X-Ray Images Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

Colour, Shading and Reflectance of Light

Superpixels Optimized by Color and Shape . . . . . . . . . . . . . . . . . . . . . . . . 297

Maximum Consensus Parameter Estimation by Reweighted ‘1 Methods . . . . . 312

A Graph Theoretic Approach for Shape from Shading . . . . . . . . . . . . . . . . . 328

A Variational Approach to Shape-from-Shading Under

Sharpening Hyperspectral Images Using Spatial and Spectral Priors

Inverse Lightfield Rendering for Shape, Reflection

Propagation and Time-Evolution

Modelling Stable Backward Diffusion and Repulsive Swarms with Convex

Euler-Lagrange Network Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424

Slack and Margin Rescaling as Convex Extensions

Multi-object Convexity Shape Prior for Segmentation . . . . . . . . . . . . . . . . . 455

Fast Asymmetric Fronts Propagation for Voronoi Region Partitioning

PointFlow: A Model for Automatically Tracing Object Boundaries

An Isotropic Minimal Path Based Framework for Segmentation

Inference, Labeling and Relaxation

Limited-Memory Belief Propagation via Nested Optimization . . . . . . . . . . . . 517

Geometric Image Labeling with Global Convex Labeling Constraints . . . . . . 533

Discretized Convex Relaxations for the Piecewise Smooth

A Projected Gradient Descent Method for CRF Inference

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581

Christian Bauckhage1,2(B) , E. Brito1,2 , K. Cvejoski1,2 , C. Ojeda1,2 ,