full_report.tex

\documentclass[12pt, a4paper]{memoir} % for a short document
\usepackage[french,english]{babel}

\usepackage [vscale=0.76,includehead]{geometry}                % See geometry.pdf to learn the layout options. There are lots.
%\geometry{a4paper}                   % ... or a4paper or a5paper or ...
%\geometry{landscape}                % Activate for for rotated page geometry
%\OnehalfSpacing
% \setSingleSpace{1.05}
%\usepackage[parfill]{parskip}    % Activate to begin paragraphs with an empty line rather than an indent


%===================================== packages
\usepackage{lipsum}
\usepackage{graphicx}
\usepackage{subcaption}
\usepackage{fullpage}
\usepackage{mathptmx} % font = times
\usepackage{helvet} % font sf = helvetica
\usepackage[utf8]{inputenc}
\usepackage{relsize}
\usepackage[T1]{fontenc}
\usepackage{tikz}
\usepackage{booktabs}
\usepackage{textcomp}%textquotesingle
\usepackage{multirow}
\usepackage{pgfplots}
\usepackage{url}
\usepackage{footnote}
\usepackage{amsmath,amssymb}
\usepackage{hyperref}
\usepackage{colortbl}
%============================================
\usetikzlibrary{arrows,shapes,positioning,shadows,trees}
\makesavenoteenv{tabular}
\makesavenoteenv{table}
%==============================================
\def\checkmark{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;}
%Style des têtes de section, headings, chapitre
\headstyles{komalike}
\nouppercaseheads
\chapterstyle{dash}
\makeevenhead{headings}{\sffamily\thepage}{}{\sffamily\leftmark}
\makeoddhead{headings}{\sffamily\rightmark}{}{\sffamily\thepage}
\makeoddfoot{plain}{}{}{} % Pages chapitre.
\makeheadrule{headings}{\textwidth}{\normalrulethickness}
%\renewcommand{\leftmark}{\thechapter ---}
\renewcommand{\chaptername}{\relax}
\renewcommand{\chaptitlefont}{ \sffamily\bfseries \LARGE}
\renewcommand{\chapnumfont}{ \sffamily\bfseries \LARGE}
\setsecnumdepth{subsection}


% Title page formatting -- do not change!
\pretitle{\HUGE\sffamily \bfseries\begin{center}}
\posttitle{\end{center}}
\preauthor{\LARGE  \sffamily \bfseries\begin{center}}
\postauthor{\par\end{center}}
\newcommand{\jury}[1]{%
\gdef\juryB{#1}}
\newcommand{\juryB}{}
\newcommand{\session}[1]{%
\gdef\sessionB{#1}}
\newcommand{\sessionB}{}
\newcommand{\option}[1]{%
\gdef\optionB{#1}}
\newcommand{\optionB} {}

\renewcommand{\maketitlehookd}{%
\vfill{}  \large\par\noindent
\begin{center}\juryB \bigskip\sessionB\end{center}
\vspace{-1.5cm}}
\renewcommand{\maketitlehooka}{%
\vspace{-1.5cm}\noindent\includegraphics[height=12ex]{./img/logo-uga.png}\hfill\raisebox{2ex}{\includegraphics[height=14ex]{./img/logoINP.png}}\\
\bigskip
\begin{center} \large
Master of Science in Informatics at Grenoble \\
Master Informatique \\
Specialization \optionB  \end{center}\vfill}
% =======================End of title page formatting

\option{Data Science}
\title{Deep learning with Siamese networks for instance search or identification} %\\\vspace{-1ex}\rule{10ex}{0.5pt} \\sub-title}
\author{Matthias Kohl}
\date{June 20, 2017} % Delete this line to display the current date
\jury{
Research project performed at LIG - MRIM\\\medskip
Under the supervision of:\\
Georges Quénot, Jean-Pierre Chevallet\\\medskip
% Defended before a jury composed of:\\
% Head of the jury (TODO)\\
% Jury member 1\\
% Jury member 2\\
}
\session{June \hfill 2017}
\setcounter{tocdepth}{4}
\setcounter{secnumdepth}{4}

%%% BEGIN DOCUMENT
\begin{document}
\selectlanguage{English} % french si rapport en français
\frontmatter
\begin{titlingpage}
\maketitle
\end{titlingpage}

%\small
\setlength{\parskip}{-1pt plus 1pt}

\renewcommand{\abstracttextfont}{\normalfont}
\abstractintoc
\begin{abstract}
Our research is part of the GUIMUTEIC project, which aims at building an augmented
audio-guide for museums or tourist sites. In particular, our goal is to
develop a system capable of identifying a
piece of art using solely
the visual clues of an image provided by a camera, in order to provide
further information about the piece of art to the user of the audio-guide.
We focus on instance retrieval approaches based on deep learning.
We make three main contributions:
we analyze advantages and limitations of the current state of the art
in instance retrieval,
we propose a novel approach based on a fully
convolutional network (FCN),
and we evaluate the performance of these different approaches in depth.
We show that the proposed architecture produces state-of-the-art results,
is fast to run and reasonably fast to setup for a new museum or
tourist site. The evaluation is carried out on the CLICIDE and GaRoFou
datasets, both created in the context of the project, containing images
of pieces of art from museums in Grenoble and Lyon.
\end{abstract}
\abstractintoc

\renewcommand\abstractname{Acknowledgement}
\begin{abstract}
This work has been carried out in the context of the Guimuteic project
funded by Fonds Européen de Développement Régional (FEDER) of région
Auvergne Rhône-Alpes.

I would like to express my sincere gratitude to Georges Quénot,
Jean-Pierre Chevallet and Maxime Portaz for their invaluable assistance
and comments in reviewing this report.
\end{abstract}

\renewcommand\abstractname{Publication}
\begin{abstract}
The work developed as part of this research project will be submitted as a
paper to the second international workshop on egocentric perception,
interaction and computing at ICCV 2017.
\end{abstract}

\renewcommand\abstractname{R\'esum\'e}
\begin{abstract} \selectlanguage{French}
Notre travail de recherche fait partie du projet GUIMUTEIC qui doit produire un audio-guide
augment\'e pour des mus\'ees ou des sites touristiques. En particulier, notre objectif est
de d\'evelopper un système d'identification automatique de l'\oe{}uvre d'art
contempl\'e par un utilisateur de l'audio-guide, en utilisant uniquement
une image issue d'une caméra.
Notre recherche se focalise sur les approches d'apprentissage profond
pour l'identification d'instances.
Nous apportons trois contributions majeures:
une analyse des avantages et inconv\'enients
des solutions d'identification d'instances de l'\'etat de l'art, la proposition d'une nouvelle approche bas\'ee sur
un r\'eseau enti\`erement convolutionnel (FCN),
et l'\'evaluation de la performance de ces diff\'erentes
approches. Nous montrons ainsi que l'architecture propos\'ee donne des
r\'esultats performants tout en \'etant rapide \`a l'interrogation et raisonnablement rapide \`a mettre en place pour une nouvelle collection de mus\'ee ou site touristique. L'\'evaluation se fait sur les deux corpus d'images CLICIDE et GaRoFou, tous deux produits dans le cadre du projet, et contenant des images d'\oe{}uvres d'art d'un mus\'ee de Grenoble et de Lyon.
\end{abstract}
\selectlanguage{English}

\cleardoublepage

\tableofcontents* % the asterisk means that the table of contents itself isn't put into the ToC
\normalsize

\mainmatter
\SingleSpace
%==============================CHAPTERS==================
\include{./chapters/intro}
\include{./chapters/state_of_the_art}
\include{./chapters/contributions}
\include{./chapters/evaluation}
\include{./chapters/results}
\include{./chapters/conclusion}
\include{./chapters/appendix}
%=========================================================


%=========================================================
\backmatter

\bibliographystyle{plain} % plain-fr si rapport en français
\bibliography{siamese}

%\cleardoublepage % Goes to an odd page
%\pagestyle{empty} % no page number
%~\newpage % goes to a new even page

\end{document}