-
Notifications
You must be signed in to change notification settings - Fork 6
/
poster_landscape.tex
239 lines (216 loc) · 9.64 KB
/
poster_landscape.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
\documentclass[landscape,a0paper,fontscale=0.292]{baposter}
\usepackage[vlined]{algorithm2e}
\usepackage{times}
\usepackage{calc}
\usepackage{url}
\usepackage{graphicx}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{relsize}
\usepackage{multirow}
\usepackage{booktabs}
\usepackage{graphicx}
\usepackage{multicol}
\usepackage[T1]{fontenc}
\usepackage{ae}
\usepackage{enumitem}
\usepackage[pagebackref=true,breaklinks=true,letterpaper=true,colorlinks,bookmarks=false,urlcolor = blue,]{hyperref}
\usepackage{colortbl}
\usepackage{xcolor}
\graphicspath{{images/}}
\definecolor{salmon}{rgb}{1.0, 0.90, 0.75}
\definecolor{LightCyan}{rgb}{0.88,1,1}
\definecolor{azure}{rgb}{0.94, 1.0, 1.0}
\definecolor{khaki}{rgb}{0.94, 0.9, 0.55}
\definecolor{lemonchiffon}{rgb}{1.0, 0.98, 0.8}
\definecolor{mistyrose}{rgb}{1.0, 0.89, 0.88}
\definecolor{palerobineggblue}{rgb}{0.59, 0.87, 0.82}
\definecolor{mossgreen}{rgb}{0.68, 0.87, 0.68}
\definecolor{magicmint}{rgb}{0.77, 1, 0.92}
\definecolor{grannysmithapple}{rgb}{0.66, 0.89, 0.63} \definecolor{lavenderblue}{rgb}{0.9, 0.9, 1.0}
\definecolor{peachorange}{rgb}{1.0, 0.8, 0.6}
\definecolor{pistachio}{rgb}{0.85, 1, 0.72}
\definecolor{brightmaroon}{rgb}{0.76, 0.13, 0.28}
\definecolor{bluencs}{rgb}{0.0, 0.53, 0.74}
\setlist[itemize]{leftmargin=*,nosep}
\setlength{\columnsep}{0.7em}
\setlength{\columnseprule}{0mm}
% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% % Save space in lists. Use this after the opening of the list
% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\newcommand{\compresslist}{%
\setlength{\itemsep}{0pt}%
\setlength{\parskip}{0pt}%
\setlength{\parsep}{0pt}%
}
\renewcommand{\rmdefault}{ptm} % Arial
\renewcommand{\sfdefault}{ptm} % Arial
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Begin of Document
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Here starts the poster
%%---------------------------------------------------------------------------
%% Format it to your taste with the options
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{poster}{
% Show grid to help with alignment
grid=false,
columns=4,
% Column spacing
colspacing=0.7em,
% Color style
headerColorOne=cyan!20!white!90!black,
borderColor=cyan!30!white!90!black,
% Format of textbox
textborder=faded,
% Format of text header
headerborder=open,
headershape=roundedright,
headershade=plain,
background=none,
bgColorOne=cyan!10!white,
headerheight=0.12\textheight}
% Eye Catcher
{
\includegraphics[width=0.07\linewidth]{VGG_logo.jpg}
\makebox[0.005\textwidth]{}
\includegraphics[width=0.08\linewidth]{Google_Research.png}
}
% Title
{\sc\huge\bf \textit{Speech2Action:} Cross-modal Supervision for Action Recognition}
% Authors
{Arsha Nagrani$^{1,2}$~~~~Chen Sun$^2$~~~~David Ross$^2$ ~~~~Rahul Sukthankar$^2$~~~~Cordelia Schmid$^{2}$ ~~~~Andrew Zisserman$^{1,3}$ \\[0.2em]
% \small{\url{https://www.robots.ox.ac.uk/~vgg/research/speech2action/}} \\[0.2em]
$^1$VGG, Oxford~~~~$^2$Google Research~~~~$^3$DeepMind~~~~ }
{
\begin{tabular}{r}
%\makebox[0.01\textwidth]{}
% \includegraphics[width=0.04\linewidth]{QR_code.png}
\includegraphics[width=0.14\linewidth]{CVPR_logo_2020.png}
\end{tabular}
}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Now define the boxes that make up the poster
%%%---------------------------------------------------------------------------
%%% Each box has a name and can be placed absolutely or relatively.
%%% The only inconvenience is that you can only specify a relative position
%%% towards an already declared box. So if you have a box attached to the
%%% bottom, one to the top and a third one which should be inbetween, you
%%% have to specify the top and bottom boxes before you specify the middle
%%% box.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\headerbox{\bf\color{blue} Problem Definition and Contribution}{name=contribution,column=0,row=0,span=1}{
\textbf{\color{blue}Goal:} Action Recognition in Movies and TV shows using only the speech as supervision
\begin{center}
\includegraphics[width=1\textwidth]{images/teaser_6.pdf}
\end{center}
\textbf{\color{blue}Motivation:}
\begin{itemize}
\item Manual annotation of human actions is expensive and not scalable
\item The audiotrack is usually freely available for large video corpuses.
\end{itemize}
\textbf{\color{blue}Key Contributions:}
\begin{itemize}
\item A \texttt{Speech2Action} model trained from literary screenplays that predicts actions from transcribed speech \textit{alone}
\item By applying this \texttt{Speech2Action} model to a large unlabelled corpus of videos, we obtain obtain weak action labels for over 800K video clips
\item An action classifier trained on these clips with \textit{no} fine-tuning beats fully supervised performance on the AVA dataset.
\item With finetuning, the classifier achieves state of the art results on HMDB51
\end{itemize}
}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\headerbox{\bf\color{blue} IMSDb Dataset}{name=imsdb,column=1,row=0,span=1}{
% \input{imsdb}
\begin{itemize}
\item Download 1,080 movie scripts from \url{www.IMSDb.com} spanning 22 genres
\item Separate out scene descriptions (which contain mention of \textbf{\textcolor{brightmaroon}{actions}}) and \textbf{\textcolor{bluencs}{speech segments}}.
\item Create a text dataset of \textbf{\textcolor{bluencs}{speech}} paired with \textbf{\textcolor{brightmaroon}{action}} labels, using proximity in the movie scripts as a cue.
\end{itemize}
\begin{center}
Examples of Movie Scripts
\includegraphics[width=\linewidth]{images/scripts_cropped.pdf}
\end{center}
}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\headerbox{\bf\color{blue} Results on Visual Action Recognition}{name=results,column=2,row=0,span=2}{
% \input{figures/syn_quant}
\vspace{0.5em}
\textbf{\color{blue}Examples of clips mined using Speech2Action:}
\begin{center}
\vspace{-0.8em}
\includegraphics[width=0.94\linewidth]{images/mining_cropped.pdf}
\vspace{-1em}
\end{center}
\textbf{\color{blue}Examples of abstract actions mined using Speech2Action:}
\begin{center}
\vspace{-0.8em}
\includegraphics[width=0.94\linewidth]{images/mining_unusual.pdf}
\vspace{-1em}
\end{center}
\vspace{0.5em}
{\bf\color{blue}Results on 14 AVA mid and tail classes}
\vspace{0.2em}
\input{avatable}
\begin{minipage}{0.5\linewidth}
{\bf\color{blue}Action Recognition Model}
\begin{itemize}
\item We train an S3D-G model for 18-way classification on video clips labelled with the Speech2Action model
\item We evaluate on AVA with NO finetuning, on mid and tail classes. These actions occur \textit{rarely}, and are hard to get manual supervision for. For 8 classes, we exceed fully supervised performance without a single manually labelled training example.
\item On HMDB51, we obtain a 17\% improvement over training from scratch and also outperform previous self-supervised and weakly supervised works.
\end{itemize}
\end{minipage}
\begin{minipage}{0.5\linewidth}
\vspace{1em}
{\bf\color{blue} \quad Results on HMDB51}
\vspace{-0.3em}
\begin{flushright}
\input{hmdb}
\end{flushright}
\end{minipage}
}
\headerbox{\bf\color{blue} Acknowledgments}{name=ack,column=2,below=results,span=2}{
Arsha is supported by a Google PhD Fellowship. We are grateful to Carl Vondrick for early discussions.
}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\headerbox{\bf\color{blue} Mining with the Speech2Action Model}{name=mining,column=0,below=contribution,span=2}{
\textbf{\color{blue}Main idea:} We train a text-based model to predict actions from transcribed speech alone. This model is trained on movie scripts downloaded from IMSDb. This can be applied to the transcribed speech from unlabelled videos to automatically get labels for video clips.
\begin{minipage}{0.4\linewidth}
{\bf\color{blue}Speech2Action Model}
\begin{itemize}
\item We obtain speech-action paired data for 18 action classes from the IMSDb data
\item We finetune a BERT model pretrained on English Wikipedia and the BooksCorpus
\end{itemize}
\end{minipage}
\begin{minipage}{0.6\linewidth}
\input{speech_preds}
\end{minipage}
\begin{minipage}{0.7\linewidth}
\textbf{\color{blue}Mining Clips Automatically:}
\begin{itemize}
\item We apply the Speech2Action model to the subtitles of unlabelled movies and TV shows.
\item We then assign the label for highly confident predictions of the model to the accompanying video clip.
\item In this manner we mined over 800K video clips and assign them with action labels based on the speech alone.
\end{itemize}
\end{minipage}
\begin{minipage}{0.3\linewidth}
\begin{center}
\begin{tiny}We mine two orders of magnitude more data than AVA automatically
\end{tiny}
\vspace{-2em}
\includegraphics[width=0.9\linewidth]{images/training_dist.png}
\vspace{-0.8em}
\end{center}
\end{minipage}
\vspace{1.8em}
\begin{minipage}{0.1\linewidth}
\includegraphics[width=0.9\linewidth]{images/QR_code.png}
\end{minipage}
\begin{minipage}{0.8\linewidth}
More details at: \small{\url{https://www.robots.ox.ac.uk/~vgg/research/speech2action/}} \\
Contact: \url{arsha@robots.ox.ac.uk}
\end{minipage}
}
\end{poster}
\end{document}