Assessing Team Strategy using Spatiotemporal Data

Patrick Lucey Dean Oliver

Disney Research ESPN
Pittsburgh, PA, USA Bristol, CT, USA
[email protected] [email protected]

Peter Carr Joe Roth Iain Matthews

Disney Research Disney Research Disney Research
Pittsburgh, PA, USA Pittsburgh, PA, USA Pittsburgh, PA, USA
[email protected] [email protected] [email protected]

ABSTRACT Canada
9(3) Shots (on Goal)
USA
12(4)
In CVPR, 2009.
[16] W. Lu, J. Ting, K. Murphy, and
The Moneyball revolution coincided with a shift in the way 13 Fouls 11 Players in Broadcast Sports Vide
professional sporting organizations handle and utilize data in 3 Corner Kicks 8 Random Fields. In CVPR, 2011.
1 O↵sides 4
terms of decision making processes. Due to the demand for 38% Time of Possession 62%
[17] P. Lucey, A. Bialkowski, P. Carr
I. Matthews. Characterizing Mul
better sports analytics and the improvement in sensor tech- 3 Yellow Cards 1
Behavior from Partial Team Tra
nology, there has been a plethora of ball and player tracking 0 Red Cards 0
the English Premier League. In A
4 Saves 3
information generated within professional sports for analyt- [18] L. Madden. NFL to Follow Army
ical purposes. However, due to the continuous nature of the (a) (b)Sensors in Attempt to Prevent H
data and the lack of associated high-level labels to describe8. REFERENCES
Figure 1: (a) An example of standard soccer
www.forbes.com/sites/lancema
statis-
nfl-to-follow-armys-lead-on-
it - this rich set of information has had very limited use espe-[1] tics
S. Ali andbased
M. Shah.on hand-labeled
Floor event
Fields for Tracking data which
in High describe
-attempt-to-prevent-head-inj
cially in the analysis of a team’s tactics and strategy. In this Density
what Crowd Scenes. In ECCV,
happened. 2008.
(b) Spatiotemporal data[19]has the po- Y. Chang, A. H
R. Masheswaran,
paper, we give an overview of the types of analysis currently[2] N. Allen, J. Templon, P. McNally, L. Birnbaum, and
tential to describe the where and how, but S. Danesis.
as it Destructing
is the Rebo
K. Hammond. StatsMonkey: A Data-Driven Sports Tracking Data. In MIT Sloan Sp
performed mostly with hand-labeled event data and high- a continuous
Narrative Writer. In AAAI signal which is
Fall Symposium not associated
Series, with
Conference, 2012.
a
light the problems associated with the influx of spatiotem- 2010.fixed event, using this data for analysis is V.difficult.
[20] Morariu and L. Davis. Multi-
poral data. By way of example, we present an approach[3] BBC-Sports. Footballers may trial wearing microchips Recognition in Structured Scena
which uses an entire season of ball tracking data from the to monitor health. [21] T. Moskowitz and L. Wertheim.
English Premier League (2010-2011 season) to reinforce the
1. INTRODUCTION
www.bbc.co.uk/sport/0/football/21460038, 14 Feb Hidden Influences Behind How S
2013. Games Are Won. Crown Publish
common held belief that teams should aim to “win home[4] In his 2003 book Moneyball
M. Beetz, N. von Hoyningen-Huene, B. Kirchlechner,
[14], Michael Lewis docu-
[22] NBA Shot Charts. www.nba.com/
games and draw away ones”. We do this by: i) forming a mented
S. Gedikli, how M.
F. Siles, Oakland
Durus, andA’sM.General
Lames. Manager Billy Beane was
[23] D. Oliver. Basketball on Paper: R
representation of team behavior by chunking the incoming able to Automated
ASPOGAMO: effectivelySports
use Game
metrics derived from hand-crafted
Analysis Performance Analysis. Brassey’s
Models. International
statistics Journal of
to exploit theComputer Science inin the value of indi-
inefficiencies
spatiotemporal signal into a series of quantized bins, and ii) [24] D. Oliver. Guide to the Total Qu
Sport, 8(1), 2009.
generate an expectation model of team behavior based on[5] vidual baseball players. Around
P. Carr, Y. Sheikh, and I. Matthews. Monocular
the same time, Basketball
espn.go.com/nfl/story/_/id/6
explaining-statistics-total-
a code-book of past performances. We show that home ad- on Paper [24] was published
Object Detection using 3D Geometric Primitives. which outlined methods
4 August 2011.
for
vantage in soccer is partly due to the conservative strategy 2012.valuing player performance in basketball which is a far more
[25] Opta Sports. www.optasports.co
of the away team. We also show that our approach can flag[6] K. Goldsberry.
challenging CourtVision:
problemNew Visual and
because it isSpatial
Analytics for the NBA. In MIT Sloan Sports Analytics
a continuous[26] team sport.A. Ess, K. Schindle
S. Pellegrini,
anomalous team behavior which has many potential appli- Due to the popularity and effectiveness of the tools
Conference, 2012.
that
You’ll em-
Never Walk Alone: Model
anated from these works, there for Multi-Target Tracking. In CV
cations. [7] A. Gupta, P. Srinivasan, J. Shi, and L. Davis.has been enormous interest
[27] M. Perse, M. Kristan, S. Kovacic
in the field
Understanding of sports
Videos, analytics
Constructing over the
Plots: Learning a last 10 years with
Trajectory-Based Analysis of Co
Visually Grounded Storyline Model from Annotated
many organizations (e.g. professional teams, media groups)
Categories and Subject Descriptors Videos. In CVPR, 2009.
housing their own analytics department. However,
Activity
nearly
Image
in Basketball Game. Co
all
Understanding, 2008.
[8] Hawk-Eye. www.hawkeyeinnovations.co.uk. [28]hand-labeled
Prozone. www.prozonesports.co
H.4 [Information Systems Applications]: Miscellaneous;[9] of the analytical works have
D. Henschen. IBM Serves New Tennis Analytics At dealt solely with
[29] B. Siddiquie, Y. Yacoob, and L.
I.2.6 [Learning]: General event data
Wimbledon. which describes what happened (e.g.
www.informationweek.com/software/ basketball
Plays in American Football Vide
- rebounds, points scored, assists, football - yards
business-intelligence/ per carry,
University of Maryland, 2009.
Keywords ibm-serves-new-tennis-analytics-at-wimbl/
tackles, sacks,
240002528, 23 June 2012.
soccer - passes, shots, tackles (see
[30] Figure 1(a))).
SportsVision. www.sportsvision
Sports Analytics, Spatiotemporal Data, Representation [10] Once and
A. Hervieu the P.data is in Understanding
Bouthemy. this form, most sportsapproaches just relate
[31] STATS SportsVU. www.sportvu.
videotousing
parsing
playersin the relevant
trajectories. data L.
In J. Zhang, from [32] Statsheet.
Shao,a database, then ap- www.statsheet.com.
L. Zhang, andsport-based
G. Jones, editors, Intelligent Video [33] D. Stracuzzi, A. Fern, K. Ali, R.
plying rules and standard statistical methods,
T. Konik, and D. Shapiro. An A
Event Analysis and Understanding. Springer Berlin /
including regression and optimization.
Heidelberg, 2010. to American Football: From Obs
Permission to make digital or hard copies of all or part of this work for [11] Asand
S. Intille most sporting
A. Bobick. environments
A Framework tend to be dynamic
for Recognizing Video to withControl in a Simulated
personal or classroom use is granted without fee provided that copies are multiple
Multi-Agent players
Action fromcontinuously
Visual Evidence.moving
In AAAI, and competing Magazine,
against32(2), 2011.
not made or distributed for profit or commercial advantage and that copies 1999.each other, simple event statistics do not capture [34] X. Wei, P. Lucey, S. Morgan, and
the com-
Sweet-Spot: Using Spatiotempor
bear this notice and the full citation on the first page. To copy otherwise, to [12] K. Kim,
plexM.aspects
Grundmann, of theA. Shamir,
game.I. Matthews,
To gain an advantage and over
Predict the
Shots in Tennis. In M
republish, to post on servers or to redistribute to lists, requires prior specific J. Hodgins, and I. Essa. Motion Fields to Predict Play
rest of the field, sporting
permission and/or a fee. Evolution in Dynamic Sports Scenes.organizations
In CVPR, 2010. have recently Analyticslooked
Conference, 2013.
[13] to employ
M. Lewis. Moneyball:commercial trackingan technologies
The Art of Winning Unfair [35] C. Xu,can
which Y. Zhang,
lo- G. Zhu, Y. Ru
KDD’13, August 11–14, 2013, Chicago, Illinois, USA. Q. Huang. Using Webcast Text f
Copyright 2013 ACM 978-1-4503-2174-7/13/08 ...$15.00. Game.cate the position
Norton, 2003. of the ball and players at each time instant
Detection in Broadcast. T. Multi
[14] R. Li and R. Chellappa. Group Motion Segmentation
Using a Spatio-Temporal Driving Force Model. In [36] Zonalmarking. www.zonalmarkin
CVPR, 2010.
[15] R. Li, R. Chellappa, and S. Zhou. Learning
Multi-Modal Densities on Discriminative Temporal
1366
Interaction Manifold for Group Activity Recognition.
in professional leagues [29, 32, 9, 26]1 - to determine where alizations for the television broadcasters [9]. Partial data
and how events happen. Even though there is potentially sources normally generated by human annotators such as
an enormous amount of hidden team behavioral information shot-charts in basketball and ice-hockey are often used for
to mine from such sources, due to the sheer volume as well analysis [23], as well as passing and shot charts in soccer
as the noisy and variable length of the data, methods which [26]. Recently, ESPN developed a new quarterback rating
can adequately represent team behaviors are yet to be devel- in American Football called “TotalQBR” [25] which attempts
oped. The value of this data is limited as little analysis can to assign credit or blame to the quarterback depending on
be conducted. What compounds the difficulty of this task a host of factors such as pass or catch quality, importance
is the low-scoring and/or continuous nature of many team in the match, pass thrown under pressure or not. As these
sports (e.g. soccer, hockey, basketball) which makes it very factors are quite subjective, annotators who are reliable in
hard to associate segments of play with high-level behaviors labeling such variables are used. In terms of strategic anal-
(e.g. tactics, strategy, style, system, formations). Without ysis, zonalmarking.net [37] attempts to describe a soccer
these labels, which essentially give the game context, infer- match from a tactical and formation point of view. Whilst
ring team strategy or behavior is impossible as there are no interesting, this approach is still qualitative and is based
factors to condition against (see Figure 1(b)). solely on the opinion of the analyst.
Due to these complexities, there has been no effective As the problem of fully automatic multi-agent tracking
method of utilizing spatiotemporal data in continuous sports. from vision-based systems is still an open one, most aca-
Having suitable methods which can first develop suitable demic research has centered on the tracking problem [1, 27,
representations from imperfect (e.g. noisy or impartial) data, 17, 5]. The lack of fully automated tracking approaches has
and then learn team behaviors in an unsupervised or semi- limited team behavioral research to works on limited size
supervised manner, as well as recognize and predict future datasets. The first work which looked at using spatiotem-
behaviors would greatly enhance decision making in all ar- poral data for team behavior analysis was conducted over 10
eas of the sporting landscape (e.g. coaching, broadcasting, years ago by Intille and Bobick [12]. In this seminal work,
fantasy-games, video games, betting etc.). We call this the the authors used a probabilistic model to recognize a single
emerging field of Sports Spatiotemporal Analytics - and we football play from hand annotated player trajectories. Since
show an example of this new area of analytics by compar- then, multiple approaches have centered on recognizing foot-
ing the strategies of home and away teams in the English ball plays [16, 30, 15, 34], but only on a very small number of
Premier League by using ball tracking data. plays (i.e. 50-100). For soccer, Kim et al. [13] used the global
motion of all players in a soccer match to predict where the
play will evolve in the short-term. Beetz et al. [4] proposed
2. RELATED WORK a system which aims to track player and ball positions via
The use of automatic sports analysis systems have re- a vision system for the use of automatic analysis of soccer
cently graduated from the virtual to the real-world. This matches. In basketball, Perse et al. [28] used trajectories of
is due in part to the popularity of live-sport, the amount player movement to recognize three type of team offensive
of live-sport being broadcasted, the proliferation of mobile patterns. Morariu and Davis [21] integrated interval-based
devices, the rise of second-screen viewing, the amount of temporal reasoning with probabilistic logical inference to
data/statistics being generated for sports, and demand for recognize events in one-on-one basketball. Hervieu et al. [11]
in-depth reporting and analysis of sport. Systems which use also used player trajectories to recognize low-level team ac-
match statistics to automatically generate narratives have tivities using a hierarchical parallel semi-Markov model. In
already been deployed [33, 2]. Although impressive, these addition to these works, plenty of work has centered on an-
solutions just give a low-level description of match statistics alyzing broadcast footage of sports for action, activity and
and notable individual performances without giving any tac- highlight detection [36, 8]2 . Even though notable, the lack
tical analysis about factors which contributed to the result. of tracking data to adequately train models has limited the
In tennis, IBM has created Slamtracker [10] which can pro- usefulness of the above research.
vide player analysis by finding patterns that characterize It is clear from the overview given above, that there ex-
the best chance a player has to beat another player from an ists a major disparity in resources between industry and
enormous amount of event labeled data - although no spa- academia to deal with this problem domain. Sporting orga-
tiotemporal data (i.e. player or ball tracking information) nizations that receive large volumes of spatiotemporal data
has been used in their analysis yet. from third-party vendors but often the people within these
Spatiotemporal data has been used extensively in the vi- organizations lack the computational skills or resources to
sualization of sports action. Examples include vision-based make use of it. Contrastingly, due to the proprietary nature
systems which track baseball pitches for Major League Base- of commercial tracking systems, and the cost and method of
ball [31], and ball and players in basketball and soccer [32, generating the tracking data, research groups who have the
29]. Hawk-Eye deploy vision-based systems which track the necessary skills can not access these large data repositories.
ball in tennis and cricket, and is often used to aid in the Recently however, due to the potential payoff, some industry
officiating of these matches in addition to providing visu- groups are investing in analytical people with these skill sets,
1 or have teamed up with research groups to help facilitate a
As nearly all professional leagues currently forbid the use of
wearable sensors on players, unobtrusive data capture meth- solution. The release of STATS Sports VU data [32] to some
ods such as vision-based systems or armies of human annota- research groups has enabled interesting analysis of shots and
tors are used to provide player and ball tracking information. rebounding in the NBA[7, 20]. In tennis, Wei et al. [35]
However, this restriction may change soon as monitoring the
2
health and well-being of players has attracted significant in- These works only capture a portion of the field, making
terest lately, especially for concussions in American Foot- group analysis very difficult as all active players are rarely
ball [19], as well as heart issues in soccer [3]. present in the all frames.

1367
 Home Away
N Team
W D L P GF SF GA SA W D L P GF SF GA SA
1 Man Utd 18 1 0 55 49 347 12 191 5 10 4 25 29 272 25 271
2 Chelsea 14 3 2 45 39 379 13 233 7 5 7 26 30 367 20 208
3 Man City 13 4 2 43 34 306 12 216 8 4 7 28 26 240 21 298
4 Arsenal 11 4 4 37 33 350 15 154 8 7 4 31 39 305 28 251
5 Tottenham 9 9 1 36 30 383 19 228 7 5 7 26 25 274 27 339
6 Liverpool 12 4 3 40 37 319 14 220 5 3 11 18 22 266 30 270
7 Everton 9 7 3 34 31 321 23 227 4 8 7 20 20 259 22 279
8 Fulham 8 7 4 31 30 307 23 262 3 9 7 18 19 245 20 271
9 Aston Villa 8 7 4 31 26 273 19 263 4 5 10 17 22 233 40 340
10 Sunderland 7 5 7 26 25 287 27 243 5 6 8 21 20 246 29 311
11 West Brom 8 6 5 30 30 329 30 237 4 5 10 17 26 273 41 297
12 Newcastle 6 8 5 26 41 300 27 250 5 5 9 20 15 209 30 256
13 Stoke City 10 4 5 34 31 298 18 256 3 3 13 12 15 186 30 294
14 Bolton 10 5 4 35 34 311 24 256 2 5 12 11 18 261 32 346
15 Blackburn 7 7 5 28 22 254 16 259 4 3 12 15 24 200 43 360
16 Wigan 5 8 6 23 22 290 34 227 4 7 8 19 18 221 27 284
17 Wolves 8 4 7 28 30 256 30 266 3 3 13 12 16 205 36 306
18 Birmingham 6 8 5 26 19 231 22 324 2 7 10 13 18 174 36 362
19 Blackpool 5 5 9 20 30 296 37 297 5 4 10 19 25 240 41 441
20 West Ham 5 5 9 20 24 325 31 317 2 7 10 13 19 250 39 378
SUM 179 111 90 648 617 6162 446 4926 90 111 179 381 446 4926 617 6162
AVG(per game) 0.47 0.29 0.24 1.71 1.62 16.2 1.17 13.0 0.24 0.29 0.47 1.00 1.17 13.0 1.62 16.2

Table 1: Table showing the statistics for the home and away performances for each team in the 2010 EPL
season: (left columns) home matches (right columns) away columns (Key: W = wins, D = draws, L = losses,
P = points (3 for a win, 1 for a draw, 0 for a loss), GF = goals for, SF = shots for, GA = goals against,
SA = shots against).

used ball and player tracking information to predict shots cant difference between teams at home and away (10.01%
using data from the 2012 Australian Open. For soccer, re- vs 9.05% for shooting and 73.46% vs 72.99% for passing -
searchers have characterized team behaviors in the English see the bottom row in Table 1).
Premier League using ball-motion information across an en- However, there is a large difference between the amount of
tire season using OPTA data [18]. In this paper, we extend shots (16.2 vs 13.0) and goals scored (1.62 vs 1.17) at home
this method to explain that the home advantage in soccer and away. An illuminating example is the league champions
is due to the conservative strategy that away teams use (or for that season, Manchester United (see top row in Table 1).
more aggressive approach of the home team) which rein- At home, they were unbeaten (winning 18 and drawing 1),
forces the commonly held belief that teams aim to win their but away from home they only won 5 games, drew 10 and
home games and draw their away ones. lost 4. The telling statistic is that at home they scored 49
goals from 347 shots, compared to only 29 goals from 272
3. CASE STUDY: shots away from home. Comparatively, the opposition at
home games only scored a paltry 12 goals from 191 shots
HOME ADVANTAGE IN SOCCER while at away games they scored 25 goals from 271 shots.
In soccer, there is the commonly held belief that team should
3.1 “Win at Home and Draw Away” aim to win their home games and draw their away ones. If
In a recent book by Moskowitz and Wertheim [22], they you skim down Table 1, you will find that: i) all teams won
highlight that the home advantage exists in all professional more home games (except for Blackpool who won the same
sports (i.e. teams win more at home than away). The au- amount), ii) all teams score more goals at home (except Ar-
thors hypothesized that referees play a significant role by senal and Blackburn), iii) all teams had more shots at home
giving home teams favorable calls at critical moments. They compared to away, and iv) all teams gained more points at
then quantitatively showed this in baseball through the use home. These event statistics tell us what has occurred, in
of pitch tracking data. For soccer, hand-labeled event statis- the rest of the paper we use spatiotemporal data to help
tics such as the amount of injury time, number of yellow explain where and why this occurred. Before we detail the
cards and number of penalties awarded to reinforce their method, we first describe the data.
hypothesis. As soccer is a very tactical game, we hypothe-
size that the strategy of the home and away teams also plays
a role in explaining the home advantage. 3.2 Ball Tracking and Event Data
A great case study of home advantage is the 2010-2011 Due to the difficulty associated with accurately tracking
English Premier League soccer season. In that season, the players and the ball, data containing this information is still
home team earned an average 1.71 points out of a total 3 scarce. Most of the data collected is via an army of human
points per match. This is in stark contrast with the away annotators who label all actions that occur around the ball
team, which earned only 1.00 points per game: a rather - which they call ball actions. The F24 soccer data feed col-
large discrepancy, especially considering that teams play ev- lected for the English Premier League (EPL) by Opta [26]
ery other team both at home and away so that any talent is a good example of this. The F24 data is a time coded
disparities apply to both home and away averages. In terms feed that lists all player action events within the game with
of shooting and passing proficiency, there was no signifi- a player, team, event type, minute and second for each ac-

1368
 ement
acking
as over fixedwindows
information windows (?; ically hardness
solve
proach,
changedand
coded Sycara
MAPR ofthe provided single proposed
dynamic
structures problems andthat team
of another
multi-agent
in
ateam fully
groups the
membership recognizing
model
observed
of plan
agents, using
to team
prune algorithm
recognition,abeing
Sukthankar first-cut
trace
thethe sizeand andthat
of ap-
aplaybook
the li-en- with ing hardness
solve ically partial
aconnection new MAPR changed of team model, single 4problems
tracings. structures
revealingand inWethe
multi-agent of the show groups
model distinction this plan 5using ofbythe agents, framing
recognition, between
a first-cut Sukthan the the and ap
a aspatio-temporal
spatio-temporal oc- hardness of single
new statistics
sue by including
representing spatial location
behavior asof events
aaovercome
spatio-temporal gen- oc- cated the between 3and multi-agent
agents, to trackplan recognition, and
and Sycara proposed aadynam-
ment over fixed 4another recognizing algorithm that en
cation
nts. despite
er
ment of
tracking
By over events
the
analyzing use being
information
fixed of oc-
windows
aamatch gen- (?; plan
and solveproach,
proach, sue
In
often In
verifying
library.
cupancysuethis this
by
MAPR
new •by provided
unreliable paper,
representing
Characterize
paper,
provided map
•humans
statistics such
Banerjee
Characterize
representing problems
we
basedwe
as that
analysis
that propose
propose
there
including team
et aateam
on fully
team
al. in
ball
are
fully behavior
(?; ano
proposed
team the
aspatial method
observed
?).
tactics
method
movement
behavior model asas
This
tactics
quantitative
observed locationas
to to
ato is team
as overcome
spatio-temporal
function
using
over
ainput despite
formalize
of ameanstrace
function
fixed
events
spatio-temporal
team a this
trace this
and
of MAPR
first-cut
windows
of
being and the
use
is- ofis-
aoc-
oc- ali-
gen- of
the ap-
li- playbook
coded solve
hardness the size.
MAPR dynamic of single problems 5team and membership in multi-agent the model to prune using
plan the
recognition, first-cut
sizeand ofand ap-
the I(X
an
ents. By analyzing match and solve
Sycaraproach,
brary
Automatically coded
plan
erated •MAPR Show
of
proposed
cupancy the
provided
library.
by full that
dynamic
problems
team
map Banerjee
analyze teams
another that plans
(?)
based team
a
anin
and have
onfully
were
et the
recognizingmembership
unseen al.
player
ball unique
model
observed
given
proposed
movement match
tracking asplaying
using
algorithm to
teamto
in prune
a styles,
(?).first-cut
trace
formalize
terms
information
over that
fixed theofand
en-
windows despite
size
tac-
(?; ap-
MAPRa of
li- ically the hav- problem
hardness solve proach,
solve
changed and coded MAPR
Sycara
MAPR as
of theprovided a5single
structures team
dynamic problems
proposed
problems identification
and
that 4of
team another
5multi-agent
a groups inin fully membership theobserved
the task.
model
recognizing
model
of agents, plan using
to using teamrecognition,
prune Sukthankar algorithm
a trace first-cut
first-cut
the size that
ap-
of aap th l
fent over fixed windows • 5 6
nt ts.
r,nts. we By
tracking
over
eventsBy analyzing
can fixed
being information
analyzing windows
characterizegen- a match
a match the (?; with
verifying brary
sue suecupancy
new
a by by of
new representing
statistics
full
representing
such map
size. team
model,
analysis based
includingplans
team team
on
revealing
(?; ball
were behavior
spatial
behavior
?). This movement
given
the location
isas as
as
a a spatio-temporal
over
input of
spatio-temporal
distinction
despite the fixed
events
(?). between
use windows
being
of oc- oc-gen-
the plan proach, library. provided Characterize
6Banerjee 6
that et a team fully
al. proposed tacticsobserved as to a team function
formalize trace of
and MAPR the a pla
li-
s, we can characterize the coded proach,
brary
proach, the
brary ?) planof
with
time,
cupancy
dynamic
being of
erated
ofing
provided
of atime,provided
full
library.
full new
which
partial
generated.
by
map size.
team
team
team
which that
model,
we
humansbased
Banerjee that
plans
team
we
call
a
membership
plans a
on
fully
call
play-segments.
(?)
were
were fully
revealing
ball
tracings.
et and al.
observed
play-segments.givenobserved
player
movement
givento theWe
proposedprune
asas By
tracking
input
show
team
input
distinction
By
overanalyzing
team
the to (?).
this
trace
size
(?).
analyzing trace
information
fixed
formalize of by
and a
windows
between the
a
match
and
framing
matcha MAPR
li- (?;a and
the li- the
solve
Sycara solve
proach,
brary
proach, codedplan • MAPR MAPR
Show
of
proposed
• provided
library. the
provided full •7 team dynamic
that
problems Characterize
6problems
another
Banerjee teams that
that plans
5 team 6
a a in have fully
were
recognizing
fully et
in
the team
membership
al.
the uniqueobserved
model
given
observed proposed tactics
model as playing
algorithm using 7 input
team as
team
tousing
to a
prune a function
trace
styles,
(?).
trace
formalize
a
first-cut
that first-cut
the en-
and and size
despiteof
MAPap-
a a th
of
li- lia
od ng s. we to
By overcome
cananalyzing
informationcharacterize this
a (?;a match is-
the newtics. cupancycupancy
of
statistics time,
erated map map
which
by
including based
humans
based we call
oncall
spatialon (?)
ball ball
play-segments.
and movement
movement
location player of we By over
tracking
over
events analyzing
fixed fixed
being information windows
windows
gen- a match (?; • withAutomatically brary adynamic of
new full size. model, team analyze 7
plans revealing an
7awere unseen given the match
distinctionas input in terms
(?). between of tac-
By analyzing match hardness asofas ?) aof
beingplaybooksingle generated. of
and play-segments,
multi-agent plan can characterize
recognition, the
and a the
time, which we play-segments. By analyzing a match 6
ws s,
method we
for canto characterize
tactical
overcome analysis thistheof is- brary with of fullShow team that plans teams were have givenunique as input
playing (?). styles, despite hav- proach,
brary plan of ing a8full
library. provided partial15team size. 7 10plans
Banerjee that
team were
tracings. fully
et observed al. given observed
proposed We asdistinction input
10
show team to(?). (?).
formalize
this trace by and
framin li-ath
MA
ssws for tactical analysis of planbrary library. Inof as
English problem
full
this aa•as
Banerjee new
ateam
paper, playbook model,
as
plans
Premier we et aof team
al.
of
propose were revealing
proposedidentification
play-segments,
Leaguegiven
a tracking
method the
as
to input distinction
formalize
Ballwe task.
toinformationcan (?).
overcome
Action MAPR
characterize abetween
this Data is- theand coded the5 proach, the
brary with provided
of full new team team ×model, that
12 plans 7a 10
membership fully were
revealing given to the prune as team
input the trace
size of and
between the
10
aforspatio-temporal oc- ofof hardness
as ?)time,
as which
playbook
•thisof Show single
we that
call and teams
play-segments.
play-segments, multi-agent have unique
we By plan
can analyzing recognition,
playing
characterize styles,
match the despite × tics. 4 hav-
10 × Reference
ewe can characterize
tactical analysis the
of is- • erated time,by beinghumanswhich generated.
we (?) call and player
play-segments. By analyzing a (?;
match hardness of single 10 and multi-agent plan recognition, and
eethod to overcome this behavior of each team which allows for tactical analysis of
as as a playbook of play-segments, we can characterize the 15
wsior can
for
efficacy
as acharacterize
tactical
and analysis
usability
spatio-temporal the ofoc-
ofwith?) abeing
solve
Show sue new
as English
MAPR
English
hardness bythe
behavior
amodel,
In utilization
ing
representing ofof Premier
problemspaper,
Premier
partial
of single
each
revealing of
team
we
team
team and in
this League
proposethe
League approach
tracings.
which
the
behavior multi-agentmodela
allows
distinction Ball
method
Ball
asBall aWe using
on
for Action
to
an
Action
show
plantactical
between
spatio-temporal
overcome
aiPad, first-cut
this Data
recognition, which
Data
analysisbythe
the
this
oc- ap-
framing is-
1 plan
of and brary brary
library.
thehardness with ofEnglish
ofproblem
full a
Banerjee
• full Show
new
15 team
•ofproblems
team
Show Premier model,
as
single that
15
plans eta10 plans
that team
al. teams were
and revealing
proposed
teams
were
League have
identificationgiven
multi-agent
given
have unique to the
as Ball unique
as
input
formalize distinction
plan
input
playing
task. (?).
Action playing
(?).
recognition, styles,
MAPR between
J.;Data styles, despi and
ohe efficacy and usability solve •as
English playbook
MAPR Premier of
problems play-segments, League in the model we can
aAction
using characterize Data
aanalysis
first-cut the ap- Allen, N.; Templon, McNally,
ent
slish forover
refficacy
as
overcome fixed
atacticaland windows
analysis
usability
spatio-temporal
this is-windows ofoc- as behavior
aas team’s
behaviorIn
agenerated.
sue playbook
Automatically thisby ing
each
paper,
performance.
of each
representing ofpartial team
we
play-segments,
team propose
aaanalyze We team
whichwhich
team show tracings.
allows
aallows
an
behavior method
we
theunseen for
can
efficacy
for
as We
tactical
totactical show
overcome
characterize
match and
spatio-temporal in
usability
analysisthis this
terms aby
of of is-
oc- of framing
5•× tac- solve
Show ×English the
English
MAPR
the utilization20
15
Premier ×20 of
15
in
this League the approach model Ball using
on
20
Actionan aiPad, first-cut Datawhich ap-
for
vement tactical
efficacy
Premierover andanalysis
usability
League
fixed ofsoc- of hardness proach,
shows cupancy
solve English
behavior of ait’s provided
team’s
single MAPRviability
problem
mapof Premier
performance.
eachbased
and that
Problem
problems to
teamas on be
multi-agent fully
used
team
ball
which League
We observed
for
Formulation
identification
show
movement
in the
allows amateur
plan the
model Ball team
efficacy
over
recognition,
for level
Action
task.
fixed
using
tactical trace
and use. and
usability
windows
a
analysis Data
first-cut
and li-
2
of with ap-
5 5× 4a× 545× hardness
new
4 510
solve ×45
English
10
4 × ×
545×
10 8model,
410 510×55
MAPR4
8 ×
10 ×
4×
8 ing
5
15 ×10 4× ×
48 4of
10
15 8× ×84
15
10
ing Premier
partial
Premier
20 10 ×
12 single
8revealing
1015
10 15 8
12
15 ×
10
problems 8×
12
partial
×15 8 ×
15 × 8×
815
12 15team
12 820 and
15×12 15×15 12
teamLeague
15
League
×the
12 × ×
in tracings.
12multi-agent
×12 12
the distinction
12
tracings. Hammond, model Ball Ball We 25We show
plan
Action
usingAction
between
K. 2010.
show this
recognition,
a Data
References byRefere
Referenc
Referenc
the
StatsMonkey
Data
first-cut
this frami
Refer
Refe
by Ref ap R fR
ement
glish
ments.
glish
.efficacy
ish ByPremierPremier
analyzing
patio-temporal
ficacy over
and
Premier
By and League
League
fixed
usability
analyzing
aoc-
usability
League match
windows
aof
soc-
soc-of
soc-
match brary
In English
behavior a
proach,
this
our team’s
sue a team’s
of
paper,
method
tics.
cupancy
full
by ofprovided Premier
performance.
representing
team
we
each problem
on
performance.
map Problem
propose
team
thebased
Problem
plans
that as
which
2010-2011League
We
team
were amethod
aWe
on show
fully
teambehavior
Formulation
allows
show
ball
Formulation
given
the Ball
observed
English tofor
the
movement as
efficacy
identification as
overcome
input Action
a
tactical
Premier
efficacy team
overand
spatio-temporal
(?).and trace
task.
this
analysis
League
fixed Data
usabilityis-windows
usability and ofsoc- of 13a1hardness
oc-
of li-
1 111 proach,
shows 1 1solve
English
11 • 11it’s
English
of
Automatically
1provided
single MAPR viability 25
problem Premier 25 Premier
and
25
that
Problem
problems to as20be
multi-agent
analyze
aa25fully team
League used League in observed
for
Formulation
identification an
the
Allen, amateur
Narrative
plan Allen,
Ball
unseen
Allen,model Ball
Allen,
N.; team
recognition,
Allen, N.;
Allen,
N.;
Action
Writer.
Templon, match
level
Allen,
Allen,
N.; Action
task.
using
Templon,
Templon,
N.;
N.; trace
Allen,
Allen,
N.; In use.
Allen,
Templon,
Allen, in
AAAI
Templon,
Templon,
N.;
N.;
J.;a
Templon,
Data
and
N.; N.; terms
Data
Templon,
J.;
McNally,
Templon,
J.; To
N.;
N.; Fall
first-cut
and aDo
Templo
McNally
Templon
J.;
McNally
J.; li-
Sym
Temp
Temp
J.;McN
J.; Mc MoJ.PTJM
Problem Formulation
our method on the 2010-2011 English Premier League soc- proach, provided that a fully observed 30 team trace and To a D l
solve sueMAPR of our time,
aproach,
team’s
cupancy method which
performance.
provided
problems map on we the
based callin that play-segments.
the
2010-2011 We
on,show show
aanalyzefully
model
ball theaobserved
English
movement using By
efficacy
asPremieranalyzing
aBy
over and
team
first-cut League
fixed a
usability
trace match
ap-
1windows inof242solve
and
soc- a2of 25li- 2 2tics. 30 problem
30Problem Problem
30 as a team Formulation identification task.
we
ents.
hnts, ver can
shPremier Bycharacterize
Premier
fixed League
windows
analyzing
League a the
soc- soc-
match The by
tatrest brary
team’s
,,cer
tas representing
of
our oft2•the
data.
,,data.
cer
of time,
data.
full
performance. paper
, ttAutomatically
method 3,, ton
team
whichon team
tProblem
tAutomatically
4,, tthe
describes
the twe
5,,t2010-2011
behavior
plans
We
2010-2011
6,,ttcall
were the
7, ttplay-segments. t9asthe
Formulation
8, ,tanalyze
method
given an
English spatio-temporal
efficacy unseen
an
of
input
Premier and
unseen
doing
match(?).
usability
analyzing League oc-
match
this,
in aof terms
soc-
match
in353terms 3 The
22
333 3MAPR
brary tac- 2 2 222
of
rest proach,
brary tac-
of
of
full
9
the
of problems team
provided
full paper team
plans
Problem
25
in
describes
13 30
that
plans the were aanalyze model
were
Formulation
given
Formulation
fully the
17Hammond,
given
Hammond, asHammond,
Hammond,
Avrahami-Zilberbrand,
observed
method using input
Hammond,
Hammond,
as2010.
Hammond, K.Hammond,
aof
input
Hammond,
K. (?).
K.
2010.
team
first-cut
doing
Hammond,
Hammond,
2010.
K. Hammond,
Hammond,
2010.
K.
(?).
K.K. StatsMonkey:
2010.
Intrace
2010.
2010.
K.
K.
StatsMonke
2010.
StatsMonk
D.;
ap-
this,
2010.
K. K.
2010.
in
K.
StatsMo
StatsM
K.Stats
Banerje
and Stat
20120120
Fa
2SS
for
we can
tactical
characterize
aanalysis of the
the proach, cupancy
addition t our as
0our
0,provided
cer
braryt t
of
• 1method
method
1to
cer
,
atime,
Show
map
as
t t 2 playbook
of ,
experiments
data.
, as t •full
based
tics.
3 awhich
,ont the
that Problem
team
the
playbook
4 , t a
on
5of
utilization
we
,
play-segments,
fully
ball
2010-2011
t 6 plans
call
which
, of
t 7 movementFormulation
observed were of
play-segments.
play-segments,
, t 8 , t English
English
show 9 this
given we
team
over
the
can
Premier
approachas
Premierfixed
we By
suitability
characterize
trace
input
can windows
analyzing
League League
on
and (?). of
characterize ana li-
a
soc-
these
the
soc-
iPad,
match the 4proach,
which t 3
0
33
,
provided
t 33
1 , 3t2• , t Automatically
•3 , t
Problem
Automatically
4 , t Problem
5 , 30
t 6 , t 7 , t
Formulation
8Table , Formulation
t
analyze 9 Narrative Lyle,
2: an
Narrative
Narrative
Table unseen
Narrative
J.Narrative
Narrative
an
Writer.
Narrative
showingunseen
Writer.
Narrative
Writer.
Narrative match
Writer.
Narrative
In
Multi-AgentNarrative
Narrative
Writer.
Writer.
Narrative
AAAI
Writer.
In
the Writer.
AAAI
Writer.
AAAI
match
In in
In
Fall
In
identificatio terms
Writer.
Writer.
AAAI
InAAAIWriter.
Writer.
AAAI
Fall
Fall
PlanSympo
AAAI
In
InFall
in
AA
Sym
AA
Sym
Fa
ReIn
te
IF
s, analyzing
ows we cantactical
for match
characterize analysis of t0behavior ,We tas
English
0cer 11 , trefer
data. 22 , tof 3 tto
, each 4 , team
Premier
tics. t55team t66of,behaviors,
,play-segments. twhich
77, t88, tallows
League 99 By B, Ball aswe
for short,
tactical
Action observable
analysis Data of 46seg- 444a t4team
4 addition Table t040,brary 4
Table
2:
,t4t•1441Table ,t4Show
over
,to Table of
t22:2experiments
,,,ttaTable t2: 3full
tics. ,,,that
showing
match. tthe
Table
ttto 44,,showing
,team tt55a5utilization , fully
showing
Given
,time tt666plans
the
,which identification
, ttthe 7that
,,observed the 88were
ttidentification
,,ashow of
tidentification
t99team’s this given
and
rate team
the approach
possession
Algorithms.
by
rate as
suitability trace
breaking
rate by input by
breaking and
In on (?).
breaking
string of
AAAI. ana D.; li-
is
these iPad, Btw
ork
7 Avrahami-Zilberbrand,Avrahami-Zilberbrand,
Avrahami-Zilberbrand,Avrahami-Zilberbrand,
Avrahami-Zilberbrand,
Avrahami-Zilberbrand,
Avrahami-Zilberbrand, Avrahami-Zilberbrand,
Avrahami-Zilberbrand, Avrahami-Zilberbr
Avrahami-Zilberb
Avrahami-Zilberbra
Avrahami-Zilberb
D.; D.;
Banerjee,D.;
Banerj
Baner
D.;
D.;D.;
Ba
brary of time,
of cer full
We
tdata. which team
refer
shows
as a weplans
to
3playbook 4
tit’scall
team were
viability behaviors, given ttto
play-segments, , be as B, input
used as analyzing
for (?).
short, foramateur
can a match
observable
characterize level seg-
use. the t 5English
athe , We t
team , t refer over Table
3
Premier team
2: t Table
, t , behaviors,
t showing,
League t the
, t 9plans the B, Ball into
identification as short,
different
Action observable
rate time by
Dataseries
breaking and seg
,ato match.
tGiven ,,that aof team’s possession string is
the plans into different series and the field into differ-
Work fficacy for and usability of t0 ,tas
tasks t0awe1We
,We ,behavior
,English
refer
trefer
2mentioned. 34,to
23,,ttcoordinated t45ofteam tteach
,,PremierWe ,,team
7firstRelated
6behaviors,
, tbehaviors, ttmovement
78,,describewhich
League t B, Workallows asBall
our short,
dataset. tactical
Actionobservable analysis
Data 575of
seg- brary oflength 5full
,We t5t515a2the team
refer plans ttics.
into
plans
,to ateam
different
awere behaviors, time
given series as input and as theshort, (?). field into
observable differ- seg-
0 1 5shows 2 3 4 7 8
an wscharacterize the ments tof to team B, as
and short, action observable
executed seg- 5 555of5
a2Table ,tinto the 2: it’s resulting ,Table t5viability
showing
,number ttto the be used
identification
B, play-segments for J.amateur rate by
for level
breaking
each use.
6by 6a
plans
ork the tactical
efficacy and analysis
usability of of as a, behavior
team’st1playbook • performance.
Show ,of of the5 6 , tteam
play-segments, We
utilization show
Related
89 9we the
of efficacy
this
Work
can approach
characterize and usability
on the an ofiPad, tasks
6which ,6tments t06we ,We ,English
6refer
mentioned.trefer
,team T
tof ,team
athe over
tcoordinated tdifferent
over
plans team tmatch.
Premier We 6into
,areas.
,match. time
6behaviors,
,different
first ttmovement
Given series
Given
describeLeague
ent ,that ttime and
Lyle,
that
quantization Banerjee,
aB, Lyle,
team’s
seriesthe
Lyle, J.
asteam’s
aand
our Lyle,
Ballfield
2010.
Lyle,
J. J.
short, Lyle,
2010.
dataset.
and 2010.
B.;
possession
action areas. into
Lyle,
J.
Lyle,
Multi-Agent
2010.J.
2010.
2010.
Action
possession
the Lyle,
Kraemer, differ-
Lyle,
Multi-Agent
Multi-Agent
2010.
J.Lyle,
J.
The
fieldLyle,
observable Multi-Agent
2010.
2010.
Multi-Agent
J.
executed
J.intoJ.
2010.
Multi-Agent
L.;
results
string 2010.
Multi-Agent
J.
2010.Plan
2010.
Multi-Ag
Multi-Ag
Data
string Plan
andPlan
is Mul
Multi
Recog
are
differ- seg
by
Mu
Mu
Pla
is
LylPl
RRP
th
ork We ,to 45team tbehaviors, as short, observable seg-
6t60ent 3 ,T ,into
1
ments of coordinated each Related
movement which allows
Work and for
action tactical executed analysis by of aiPad, quantization
of 6length66ent •quantization Show3 areas. the the
resulting The results
utilization 7
Thenumber 8areresults the
9
of
of B, trace this
play-segments
are the aof approach the
trace team of for
the on each
team an iPad,
h
tactical Premier
rkefficacy League
analysis of soc- ments
We
ments
team ourWithWe
aThe
referEnglish
method
a refer
of
ofthe
(e.g.
team’s
rest coordinated
to • to
coordinated
pass
on ofShow
team Problem
the Premier
performance.
team
the
from the
paper
behaviors,
2010-2011behaviors,
Related
movement
movement
agent
We
utilization League
Formulation
describes
Ashow
show
to
English Work
B,
as and
agent
the
of
and asthe
short,Ball
efficacy
thisshort,
action
action
B)
Premier method Action
approach
observable
.analysis
A
and
observable
executed
executed
plan
League ofusability
can on
doing
soc- Data
seg-
by an
by
bede-
10 6of
seg-
athis,
de- a
7 777 in
6 possession
ments ments which
1 6
ent
We of quantization
English7of
of
refer
length
the
is length
ent
coordinated
14team
plans
coordinated therefore: to quantization
Show , team the
over
areas.
Premier , the
resulting
the
a7The
different match.
behaviors,movement
resulting
8
areas.
movement
Nutilization
results time
9(T Given
League
recognition.
number The −T
number areseries
and Plan )+1and
results
B, thethat
and
Algorithms.
and
ofand ofshort,
and
and
Algorithms.
Algorithms.
and
trace
Recognition:
as
. of Ball
If
play-segmentsthis
team’s
and
action
short, the
Algorithms.
play-segments
are
action the Algorithms.
Algorithms.
and
of and
the
field
Algorithms.
the
approach
Inand
Action
possession
possession
and
trace and
Algorithms. into
Algorithms.
and
AAAI.
In
In
team
executed
Formalization
observable
executed of
for
differ-
Algorithms.
Algorithms.
Algorithms.
AAAI.
AAAI.
Algorithms.
In In
InAAAI.
In
AAAI.
for
the
each
on
AAAI.
Data
string
AAAI.
In
In
by
each
team by
an
AAA
AAA
InaInI
an
sethiaa
kort and military domains,
he and Leagueusability of behavior shows it’s viability to B, be used for amateur 2
level 7use. recognition. 10 T T 14 N = 1

nglish Premier soc- of

team’s each obvious team
performance. which Related
applications Weallows to for
Work sport
the tactical and
efficacy military and of
domains,
usability of7 7We team
ments 7possession 77The refer
7 (e.g.
7recognition. of ent rest toof
shows•quantization
pass
coordinated islength team
of Problem
1 the from 1
it’s Tbehaviors,
1paper agent
,viability
the
areas. movement resulting Formulation
describes
The A(Tto 18B, to T−T
results be as
agent
number and
)+1
used the
are actionB)the offor method observable
.trace A
play-segments
amateur plan
executed
of of
the candoing
level
team seg-
be
by
for de
aan
ea
use
of soc- English Premier League
Problem Ball
Formulation Action Data
port and military domains, team (e.g.
our methodpass from on the agent
Related 2010-2011 A toWork agent English B) .Premier
A plan can
League bebe 15
soc- recognition. therefore: = Banerjee, Banerjee,
1
Banerjee, .Banerjee,
Banerjee, If
Banerjee, B.;
Banerjee, the possession
ort and military domains, ments
team With
(e.g. of coordinated
the
pass obvious
shows from movement
applications
agent
it’s viability A to to
agent
to and
sport
be B) and
used action
.A Amilitary
for plan executed
amateur domains,
can by
de-
level a10 English
use. string team is (e.g.smaller Premier pass
recognition. in from durationisProblem League agent than A to Ball
Formulation
toT,Plan agent
we Beetz, Action
discard B)M.; .B.;
Banerjee,
Kraemer,
B.;
Banerjee,
Ait.
von planBanerjee,
Kraemer,
Banerjee,
B.;
Kraemer,
B.;
ToB.;
Data Banerjee,
Banerjee,
B.;
Kraemer,
Kraemer,
Kraemer,
B.;
L.;
B.;
Kraemer,
can
Hoyningen-Huene
repre- L.;
and B.;
Kraemer,
L.;
B.;
Kraemer,
be B.;
B.;
and
L.;
and Krae
Kraem
Lyle,
L.;
L.;
de- Kra
Kra
and
L.;
Ly
Ly
Partial Team Tracing from Ball Action Data
−T
acy
port
grown glish and
and usability
military
quite
Premier domains,
substantially
League aments team
fined
team’s cer With of
data.
addition
research (e.g.as coordinated
the
performance. an obvious
pass to
ordered
interest experiments
from We into movement
applications
agent
sequence
show MAPR Athewhichto to ofand
sport
agent
efficacy
has team
grown action
showand
B)and . military
the
behaviors
quite executed
plan
usability domains,
suitability
can
substantiallydescribing
of be by 10 de-
20 a
of
10 1010 these Partial
ments
10 10 ments
team 10
fined
team 10 10 1010 of
addition (e.g.
possession
10
(e.g. as of possession
Team
recognition.
coordinated an coordinated
pass
pass to
orderedshows isfrom
Tracing
from therefore:
experiments therefore:
it’s agent
movement
sequence
agent movement
viability N
fromA AN = T
which
to of =(T
agent
T
and
agent to Ball
−T(T
team and
Tbe
1
show
)+1
action
(T
B)TB)
1
used action
)+1
.Recognition:
Action
behaviors
−T
. .Siles,
If
A .it.
A
the
)+1 the
forIf
executed
plan plan the executed
possession
amateur
suitability possession
Data
can can
describin bebybe leve
de- by
ade
of
6Problem Formulation
our With method the obvious on the 2010-2011
applications to English
sport and Premiermilitary League
domains, soc- string is smaller possession in duration is therefore: than , we
Plan Plan Recognition: discard
Plan Recognition:
PlanRecognition:
Plan Plan
Recognition:
Recognition:
Plan Recognition:
Plan To
Plan
. Plan repre-
Formalization
Recognition:
Recognition:
Plan If Recognition:
Formalization
Formalization
Recognition:
the Recognition
Recognition
Formalizatio
Formalizat
Formaliza
Formaliz
possessi Form
Form
and F aaA
6Problem Formulation
1
cer data. N Gedikli, = S.; F.; Durus, M.; an
rtgrown
and quite substantially
military domains, team
t0fined fined
,With tWith , tresearch as
2(e.g., The tpassan pass
,obvious
ttheordered
rest interest
,from ,of
from
t52010-2011 the
tapplications tsequence
into
7agent
, sequence paper
,MAPRt8A MAPR, tto A of
describes
to has team
agent grown behaviors
the
B) .method
quite A(?). plan describing
of
substantiallycan doing be de- this, tsent team in
,fined each tstring(e.g. as play-segment
The
string 3an is pass rest
ordered
smalleris tsmaller ,of
from the
in tsequence
pduration
inp ,= papertduration {p A,achieve describes
than .of . .than team
,agent p,Twe , }, the
behaviors
wediscard Tthe
.method
discard quantized
it. it.
To of
describing
To
repre- doing
repre-
to76agent to B) A plan can be d
as the an ordered to of sportteam and military domains, tsent ,15 ,step tpass ,string t4ordered ,from tsmaller , sequence , durationtto
rown quite
quitesubstantially .behaviors describing T−1
15 1515 15 15 15 015 15 15 115 15 15
emier
and League
military soc-
domains, team
our fined methodresearch
(e.g.
tasks as on
an we interest mentioned.
ordered into
agent sequence English
We has
agent
first
of grown
Premier
team B)
describe quite
A
behaviorsLeagueplan
our substantially
can
soc-
dataset.
describing be de- 25
team fined (e.g. as
each an play-segment agent 8A 09
agent of Tteam
TB) .},behaviors ,A the plan quantized can describing
be de-
a recipe
1 the Problem
used
obvious
3 4
by a
applications team Formulation to 9toachieve
sport and a goal
military A
domains, team per- 5is in than we discard it. To repr
grown
ering group substantially
actions in or- aWe t recently.
cer
recipe ,research
t data.
, t
,addition used, TheThe
t interest
, t
by significance
rest ,
to tainto ,into
of
team t
experiments ,
the tMAPR
to, of
t
paper ,considering
achieve t has grown
describes
which aa goal show group
quite
the
(?). the actions
substantially
method
AA team
suitability in or-
of
per- doing oft725 these
each
ball a
fined
this,
taA t recipe
time
position
recipe tasks
, t 2
each
sent inas , sent t
,addition an
each used
time
used we Problem
, is
The
at t
eachordered then
eachmentioned.
,
step t
play-segment
by
by rest ,to
used
, is
play-segment ttimea then , of team
sequence
experiments t Formulation
,
the
steppopulate
used t = We ,to paper
is{p
each
t
to ,
then first
,tof
0populate eachBeetz,
.9time .POGAMO:
which
Beetz,
.team
describes
used,Beetz,
describe
pstep
entry. Beetz,
M.;
each a
Beetz,
−1 to goal
M.;
Beetz,
show
M.;
behaviors
is
T von Beetz,
then Beetz,
populate
entry.
M.; von
AutomatedM.;
von
M.;
our
the (?).
the
the
Beetz,
used Beetz,
Hoyningen-Huene,
Beetz,
M.; Beetz,
von
method
the A
Hoyningen-Huen
vondataset.
each team
M.;
Hoyningen-Huen
M.;
von
M.; vonHoyningen-H
M.;
Hoyningen-
von
Hoyningen
von
describing
to M.;M.;
Hoyninge
Sports
populate
quantized
suitability
quantized
vonvon
of von
Hoynin
Hoyni
von pe
Gam Ho
do H
ea
ao
deringgroup groupactions actionsininor- or- t10 aa15 team to achieve aGedikli, goal (?). A team per-
0 1 2 3 4 5 6 7 8 9 20 20 20 20 20 20 20 20 20 20 2020 20
,afined
tresearch , ,treferrecently. tused ,to The
t,ordered
tteam ,tsignificance
,into ,,MAPR
tt16
behaviors, of considering
as group actions inor- or- We each 0, ,trefer 1 time 2 step
,to
3 tteam is 4 then 5 used
behaviors, to 7 populate 8 {p aseach short, entry. observable }, seg-
ashort, observable inseg-
,afined tused p {p = , . . . , p },
t,time ,taof ,, sequence
tt16
30 as an ordered p = of , . team. . , p behaviors describin
rown
ering quite substantially cer research
data.
11 recipe
recently. t12 as an13interest
The by
14 significanceaFigure
15 team sequence
RelatedMAPR to
7of
has
achieve B, of
considering
Work team
grown goal behaviors
quite
group (?). substantially
actions 3describing
team in per- 10
fined 11 recipe
ball
t25 0as
description position
t12 an ,ball tposition a213 sent
each
ordered 3at
of ,14 by
each
timeteach
this sequence
process
teach team
time
play-segment
step is ,step then is
tprocess to ,given achieve
is
used
of
tachieve8A,step
B,
thenGedikli,0ternational
team
in to
tdescription Gedikli,
Gedikli,
used0{p
populate
Figure aas Gedikli,
S.;
behaviors,goal
T Gedikli,
to S.;
Siles,
−1
4. S.;
of Gedikli,
Gedikli,
populate
.Journal
T
each ,Siles,
S.;
−1
Giventhis (?).
Siles, S.; Gedikli,
Gedikli,
S.;F.; entry. A
Gedikli,
Siles,
Gedikli,
S.;Siles,
Siles,
Durus,
F.;
F.;
S.;
Siles,
ofS.;
each team
Durus,
Durus,
F.;
Computer
Adescribing
process
}, the S.;
Siles,
Siles,
S.;
F.;
F.; S.;
Durus,
S.;
M.;
F.;
is Siles,
Siles,
Durus,
quantiz
givenper
Siles
Durus,
F.;
Siles,
Durus
M.;
F.;
M.;
and Du
Du
M
F.
Sc aL MF
Work
han
wn
dering k n a aquite group
sequential substantially
sequential
actions
process
process
in or- of
of
fined
There
mentsa
aforming
forming
0 asrecently.
trecipe
der
recently.
recipe
der
We der
has
of
1an
to
to, trefer
tasks
interest
to
aa2ordered
isolate
used
been
coordinated
used
isolateThe
3 ,to
group
The
addition
isolate
group we bytteam
an
aof tsequence
4significance
team
mentioned.
significance
by
team team
of a
5team
to plans,
these
explosion
these
, 3:
plans,
movement
team
plans,
tto
6behaviors,
experiments plans
Related
plans
of
to
rather
ofhas
,ratherof
tachieve
Figure 8
We
rather
in
considering
achieve
9
to
grown
tteam
,considering
to 3.
than
first
the B,
and
than thanachieve
achieve as
which
Work
quite
abehaviors
sequential
goal
describe
a a
interest
action
a
group
short, show
sequential
group
goal
sequential a
substantially
(?).
amajor
major
in our
executed
Adescribing
actions
actions
(?).
process
observable
the team
dataset.
Aprocess
live-sport
goal
process
goal
suitability
in
team
per-
or-
25
(e.g.
byor-
of
of(e.g.
ofseg-
252525
a t
per- , t
25 entry.
ments of
There
,
the
25
a
aforming
t
forming
25
forming entry.
recipe
recipe
,
25
these
possession t
2525 A We
ball
of
25
has
,
11
t
A
1description each
tasks referused
description
a
coordinated
,
a used been
t ball
group
group
addition
position
group ,
string A t we toby
positionby , an
description
4 step
at
of
t team
mentioned.
of
of a
,
of
this515
at t
team
explosion
is
to these
these
team
these
this
each then 6
movement
at
, behaviors,
experiments
process
ttime
each
of
to 7used
time
.to
plans
plans
plans
this
step
.the .time
is
We in
,achieve
19
to
aprocess
p
givenis
is to
POGAMO:
possession to
to
populate
9
first
the
}step
Cohen,
=
and
given
thenB,
POGAMO:is
achieve
POGAMO:
shown,
achieve
achieve
in aPOGAMO:
which
thenis is
0goal
used in
Figure
describe
interest POGAMO:
POGAMO:
action
a
.short,
each
POGAMO:
then
given
P., goal
.Figure
used
Automated
string we
and
to
a
POGAMO:
POGAMO:
a
paentry.
(?).
show
usedmajor
Automated
Tmajor
populate
Automated
in firstto
4.POGAMO:
in observable
−1 4.
Given
our
POGAMO:
executed
(?).
major
Levesque, POGAMO:
POGAMO:
Automated
Automated
FigureAutomated
to A
team
Given
the
populatedataset.
live-sport
goal
Automated
Sports goal
suitabil
each
Automated
Automated
populateSports
goal
4.Sports
team
H.
per-
each
(e.g.
Autom
Automa
Game
by
Given
(e.g
Auto
Sports
Auto
Sportseg
Spo
(e.g
1990. Ga
Spo
Gaa
pe
ea
nan ring
ngaagroup group
sequential actionsprocess in or-
of t , t , t
forming
recently. , t We
recognizing, t ,
toarefer
The t group , t ,
significance
to
plans t of
team , t these ,
ofplans, t of plans
behaviors,
individual considering toagents.achieve group
aas aamajor
short,
Outside major
actions observablein
oftactics.
the goal or-sport (e.g. seg-
0 abreak
1 recipe winning
forming 2 We
3 the
Partial used 4 refer
possession
aof aentry.
A 5
groupmatch), bypossession
description 6 to awe ateam
7
of string
bycan
=
these 8 of {a
to9
be 0 ,=
behaviors,
this achieve
plans said process .to be
is afield
B,
achieve goal
given as
employing short,
shown, in (?).
a13 Figure
major
a A observable
we= team
4. {a
first
tactics.
goalGiven , . .
per-.
(e.g.
, a s
aTeam Tracing ,from Ball Action D
30 30 30 30 30 30 30 30
Aments 30 30 3030
description 30 ofstring this aprocess 0is , . .0given in Figure 4. Given
a recipe der used aisolate by team
aplans to rather
achieve thanafrom goal sequential
(?). A process
team per- of coordinated movement }and
,action executed by
13
actions in or- B,and Table 2: Table showing the the identification possession rate breaking a {a , . . . , a shown,} we first 0 1
forming
winning Partial aof group
match),
aTeam
of these
can plans
Tracing
be
Related said to achieve
be
Work employing Ball Action goal (e.g. Data Aentry. entry.
0 1 2 3 4 5 6 7 8 9
ents.
Work sequential
Outside of process
the sport of ments recognizing taskscoordinated we mentioned.
of movement
individual We agents.first action
describe
Outside executed
of our the dataset.
sport by awhich team winning has the
(e.g. field
created
description aapass the tasks
up
match), into
from of aof
demand grid
can
this mentioned. =
ofbe
process
string
{a
said
for break and
to
isreal-time We, aternational
ternational
4.given
the be
Commitment. ternational
then
13
0first
ternational
employing
13ternational
international
.,vectorize ternational
.up Journal
statistics
Figure ternational
describe ternational
intoJournal
Journal ternational
ternational
Artificial 4. ternational
Journal
aternational
itof
Journal
Journal
shown, Journal
Computer
of
tactics.of
Given
grid ourJournal
and
Journal
Computer
Computer
ofweof4of
Intelligence
of Journa
Journal
Comput
of
Compu
datase
vi-
Journ
Journ
Comp
Scien
Com
of
first5of C
an SC S
aagent A to agent }B) .in
With the aobvious applications We refer to team behaviors, × as short, .aobservable aA plan can seg- be de-
toto sport and .military domains, forming aAthe group these 4grid plans a5of to {a .5achieve a}Cohen, major itgoal × (e.
gents.
tnts.
naussed and
a Outsideof
military
sequential
Outside ofthe thesport
domains,
process sport
of which
We team winning
refer
forming
der
winning der has
(e.g.
to
realm,
ments
to
recognizing to
recognizing
With
isolate
created
isolate apass
team
a
of most match),
group
match),
the teamteam
from
behaviors,
plans
of
coordinated plans
obvious demand
of
Related can
plans,
agent
of
plans,
this can these
of workbe
individual
be rather
applications Asaid
rather
for
B,plans
Work
said
individual
movement has as to
than
to than
real-time
agent be
short,
agents.
to be
focussed
agents.
to employing
aemploying
B)
aachieve sequential
sequential
sportand statistics
observable
Outside A
on
Outside
andaction plan
a major
dynamic tactics.
of
the process
process
military canand
seg-
the
plans
tactics.
of the
executed be
sport
goal
into
teams
domains, ofvi-of
sport de- (e.g.
different
by forming
a thetime winning
winning ments break
team series
possession
Ideally, break A the
ashowing the
and
group
(e.g. of
description match),
field possession
match), the
we description
pass
field
coordinated
string up
field
of
would up
into
fromthese can
into
canof intostring
athis = grid
differ-
like
of be be
agent
athis
plans {a B,
said
movement
a of
process
said to
=,4A
0process to ×
to
.recognize
=
to
.grid{a
to
Cohen, 5achieve
,is ×
be 0be and
aagent
Cohen, Cohen, is.employing
given
,and
employing
.and
Cohen, then
P., given
Cohen, shown,
and
Cohen,
P., B)
13 then
P.,team aball action
}vectorize
Cohen,
and
Cohen,
and
shown,
major
Figure
in
.
P., A
Levesque,
P.,
P., vectorize
and we
tactics
Figure
Cohen,plan
Levesque,
P.,
Levesque,
and
Cohen,
andCohen,
P.,tactics.
and
P.,4.
tactics. we
executed
it
goal
first
Levesque, Given
4.P.,
can
P.,
Levesque,
Levesque,
and
and
P.,
H.
first
P.,
Levesque, for
and (e.g.
Given
Levesque
1990.
H.and
Levesqu
H. be
and
and 1990
Leve
H. soc
Lev
1990H. de
Le
Le
H.
InbHt
19
sequential
ents. Outside
on process
dynamic of the of
sport
teams formingwinning Ideally,
team a (e.g. group
a match),
we pass of
would these
from can be
like plans
agent saidto A toto achieve
recognize
to beagent employing a
team
B) major
. A tactics
plan goal
tactics. can for(e.g. be soc- de- to give quantized
aPartial
ball positions.
Team At each
to
Tracing give time quantized step, the quan- positions. At each
.,from Ball Action 4D
sport and military domains, ments ofteam the possession
coordinated break A the
string
description movement field up of into {a this and a process Hervieu, of 13 4 } is
× shown,
A., given
5 and and then
in we
Bouthemy, vectorize
Figurefirst 4.P. Givit
201
,action executed by a Intellig
realm, Table
TableTable
Table
2:
Table
2:Table
Table
2:Table
Table
2:Table
Table
2:TableTable
Table
Table
2:
showing
Table
Table2: Table
showing
Table2: showing
Table
2: showing
2:
Table
2:showing
Table 2:
showingTable
Table
the Tableshowing
the Table
showingthe
identification
showing showing
showing
the
identification
the
showing
identification
the identification
theidentification
identification
the theidentification
the the
identification
the rate the
identification identification
identification
rate identification
rate byidentificationrate
by breaking
rate
by breaking
rate
a by
breaking
by rate
by = rate breaking
by
breaking
raterate
breaking
byrate rate
breaking
by breaking
by ,by
breaking
by
. by
.breaking
.breaking
breaking
, breaking
a
,most of
,this work has focussed onwith dynamic ateams sualizations. Ideally, Due we would
to the like
difficulty to recognize associated team with tactics tracking for soc-
fined as an break ordered the field sequence up into a grid of of
team and then vectorize it
cussed ondynamic
dynamic teamsments sualizations.
fined research
of recognizing
realm,
Ideally,
tcoordinated
as
With ,an interest
aPartial
most
theDue we
ordered plans
obvious of
would
to into
this
the
movement Team
of
sequence MAPR
individual
work
like
difficulty
applications toTracing
has has
recognize
,andof grown
agents.
focussed
associated
action
team
to sport from quite
Outside
team
behaviors
and on
executed
onmilitary
substantially
Ball
dynamic ent
tacticsof Action
quantization
the
tracking
by
describingfor
domains,sport
teams soc- Data
areas. The
break winning to
results
Ideally, give
tseries
the to
the give
,series
the quantized
are
tpossession
field
(e.g. wequantized
match),
the ,time
possession ttheup
pass trace would ,series
into ball
tcan of
fromstringball
can
,apositions.
the like
tdiffer-
grid
string positions.
team be
agent,atheistpositions.ofto said At
,into 0
recognize
t{aA At
each to
to
Commitment. each 4be
and
Commitment. time
agent
Commitment.
×
13
abehaviors
time
.employing
Commitment.
5 step,
then
Commitment.
Commitment. team
Commitment. B) step, the
vectorize
shown,
Commitment.
Commitment.
Artificial the
quan-
tactics
.Commitment.
A
Commitment.
Artificial describing
quan-
plantactics.
we
Commitment.
Artificial
Commitment.
shown, Artificial
Artificial
Artificial it
Intelligence for
can
first
Artificial
Intelligenc socbe
itwe
own quite substantially Intelligen
Artificial
Artificial Arti
Artifi
Intell
first Ar
InteAr
Art
Int
nts.
tion port
s.
eave
iguresssed
ussed Outside
Outside
grown on
and
and
3. ofof
onmilitary
dynamic
join
quite the the
teams sportsport
teams
domains,
teams
substantially over winning
winning
With
recognizing
cer. team(i.e.
Ideally,
Ideally,
fined
the
recently.
realm,
However,
10 where
research
awhere
obvious
(i.e. (e.g.tmatch),
as 11where
The we
wematch),
an
plans
mosttpassProblem
individual
12 would
as
ordered of
applications
significance
of
,individual
would
interesttcan from
soccer
13 can
thisindividual
tlike
be like
into
14 ,Formulation
be
agents
work
agent
said
sequenceis
of
tto
Figure to
to
15 said
MAPR has
to tA can
recognize
recognize
low-scoring,
agents sport
considering 16 to
agents.
3:
beto
of
can has
and be
leave
focussedagent
Figure
employing
team employing
Outside
grown
military
leave team
group
and
team
3.B)
continuous
behaviors
and
join
quiteof
dynamic
tactics
.domains,
A the the
the
tactics.
tactics
join
actions
tactics.
teams
recognition.
teams
sport
substantially
plan
plans
the
the 4plans
plans
the
for
and teams
for
thecan
plans
into
describing
in
the
or-
over
plans
into
the
plans
into
the
soc-
over soc-
com-be
the
different
plans
the
into
the
plans
different
into the
plans
into
team de-
plans
different
plansplans
players
winning
8different
different
into plans
into
a
tized
different
timeinto into into
(e.g. cer.
into
different
time different
cer.
recipe break
time into
different
series
tized
ball
Ideally,
finedtime However,
different
different
time
and
10
different
time
tized
pass
However,
aball
different
series position
time
12 the
and
used
as match),
series
time
to
athe
11
series
and
timeball
from
to
and
time
the
give wean
field
series
the time
position give
series
and
ball,
12
time
and
by
is
series
Partial field
the
andordered as
would
up
series
possession
quantized
position as
used
series
field
and
agent thequantized
the
a
soccer
series
field
the
and 13
into
is
soccer
datainto
and
team
the
field
into
and
16
used
toand
field
into
the
is be
and
Alike
a14
the populate
sequence
and
field
ball
Team
used grid
into
the
to
to
containing
said
the
differ-
into
field ball
into
differ-
is
field
string to
the15
into =
field
to
populate
agent low-scoring,
atized
field
differ- of
differ-
field
differ-
into
low-scoring,
achieve
4 to
positions.
field =
recognize a.
differ-
populate a
×
16
4into
20 be
into
Tracing
Using
ball
differ-
into
× {a
of
video
0differ-
=B)
5
into a. 5employing
.differ-
At
this
0.differ-
a
,.each
team
position
.and
differ-
{a
Using At
a.differ-
5A
T
0using
goal ,.13
Using .,=
team
each
then
plan
continuous
information
}behaviors
continuous
.atime Tfrom
.13 10,
(?).a}vectorize
,isplayers
time
= Tused
we
tactics
step,
can =
10,tactics.
}step,
A
to
shown,
the
Ball
we
be
10, team we
and
and
for
describin
populate
the
trajectories.
quan-
de-
is
quan-
we
Actio
com-
still
com
soc-
per-
a.fi
realm, most Partial
of this workTeam has Tracing
focussed onfrom dynamic Ball teams Action to
Data give Ideally, quantized
break There the we field
has ballwould up
been positions. into like an athe grid to
explosion At recognize
of each time in and step,
the team then the
interest tacticsquan-
vectorize in forit live- so
13
leave and jointeams teams team
over players (e.g. cer. (i.e.research
pass
However,
and the from individual
interestasagent soccer into A agents
toMAPR
is agent
low-scoring, can has
B) leave .grown
A and
plan
continuous join
quite can teams
substantially
be and de- over
com- then cer. fined chunk However, break
as into an theN as
ordered = fieldsoccer (14 up
− 10 into is
sequence + low-scoring,
1)/10 a
then grid Hervieu,
chunk = of
4teamHervieu,
of ×
Hervieu,
5the play-segments
4 Hervieu,
team Hervieu,
×Hervieu,
A.,
into 5 continuous
Hervieu, A., and
and
A.,
Hervieu,
Hervieu,
behaviors and
A.,
and then
Hervieu,
Hervieu,
Bouthemy,
A.,
A., Hervieu,
Hervieu,
A.,
and
Bouthemy,
Bouthemy,
and
andA.,
A.,
and and
vectorize
Bouthemy,
Bouthemy,
Bouthemy,
and
A.,
andA., A.,
Bouthemy
P.
A.,
describ com
and
and
Bouthe
Bouth
2010.
P.
P. and
and it
20
Bo
20BP
P.
dssed military domains, tball, data containing this information is still
a cer. realm,
recipe most
used of
by this work has focussed on dynamic teams Ideally, we tized would ball position
like to is used
recognize to populate team Using
tactics for soc- we
taby team achieve aquitegoal (?). Arecognition.
team per-
ing group actions inteams
or- (i.e. where individual entent quantization
ent
ent
quantization
ent
quantization
ent
quantization
ent
quantization
ent ent
quantization
ent
quantization
ent
ent
quantization
areas. ent quantization
quantization
quantization
areas.
quantization
areas. quantization
areas.
Theareas. areas.
The The areas.
results The
areas.results
Theareas.
results
The areas.
areas.
areas.The
results
results
are t
areas.
The
results are
The results
the
The
are ,
The
The t
results
the are The
results
aretrace
the are
results ,
results
results
tracethe
are
the t
trace results
the
are
of are
the ,
trace
ofthe
trace
are t
the
ofare
arethetrace
the are
team
the
the
of
of ,the
tracethe
of t
trace
team the
team
of
tracetrace
the ,
trace
ofthe t
trace
team
of
team team the of the of
team ,
of tZhang,
of
the
team the the team team ateam team L.; and N = Jones, (14− G., 10+ eds., 1)/10Intel
tagents can leave and join teams over cer. However, as soccer is − low-scoring, 4continuous and com-
a. T = 10,
ed avegrown on on
and dynamic
join
quite
dynamic substantially
teams over researchIdeally, aIdeally,
fined However, There
recently.
interest taswe
10 ,used we
an wouldhas
as,would
The
ordered
into been
soccer significance
MAPR,like
at13 like,anis
sequence
to has explosion
to
low-scoring,
recognize recognize
of ,considering
Figure
,toautomatically
t15 t16 of ateam
3: inaFigure
team the
team
continuousgroupinterest
behaviors tactics
actions
3.recognition.
tactics forand in for
describingin
soc- live-sport
com- or-soc- fined as to athen
an recipe
give then
toordered
chunk tized
quantized
10 chunk break used a into ball 11 a the
sequence
position
isby
into Nball
12 = Naball
field 13 team
positions.
(14 =up is used
(14−10+1)/10
14
of into
10 to +At
,team
to
15 achieve
a.video
1)/10 populate
each
grid 16
behaviors = of time 5a = a. goal
× 5step,
play-segments Using play-segments
5Using and (?).
the
describing
T = A
quan-
then 10, team we
vectorize pe
eave and join teams over cer. However,
recipe as soccer team is ofgrown
low-scoring, achieve substantially
continuous
goal (?). A and team com- per- give quantized topositions. At each
forming
tized ball ato position group of ,these
used plans populate to achieve Using atime aarmy step,
Tmajor
the goalwequan- (e.g.
sidering group actions in or- Inas der
this topaper, isolate we team
are plans,
looking rather
to than sequential answering process the of scarce. resulting Most in five of
give play-segments
the quantized data collected ,is video video
via shown.
using
video videoanusing using
video players
video
using
video using
using
5players
players
video
usingvideo video
5video
using
players
using
of using
trajectories.
players
players
players
using
human using
trajectories.
trajectories.
using
players
players
trajectoriplaye
trajector play
trajecto
players
trajec
play .tra
In
tra
(i.e. where individual agents can leave and join teams over cer. However, as asoccer ,ball positions.
,.At ,each time step, the quan-
{p is low-scoring,
resulting . (14−10+1)/10
.− p a. }in five continuous
play-segments T = 10, and 0 , com
etically 3:
quite
adering Figure 3. process
substantially
sequentialanswering the of finedscarce.
forming (i.e. an recently.
Mostwhere ordered
der a of
group
to
11
The
individual
the
isolate of
12
significance
sequencedata these
team agents
collected
14
of
plans
plans, team
can considering
to
rather leave
is behaviors
via
achievethan andan a group
join
army
a describing
major
sequential teamsof
recognition.
recognition.
actions recognition.
human
goal
recognition.
over
process
recognition.
(e.g.
recognition.
recognition.
inrecognition.
recognition.
or- of
recognition.
recognition.
cer. aforming
However,
tized resulting which
recipe resulting ball then in then
There
to has
used
position
tinto
a chunkfive
as give in
group ,chunk tcreated
soccer
five by
has
play-segments
quantized is into a12
aused
play-segments
oftto into
been isteam
N
these ta− N
low-scoring,
to= demand
ball ,an0 = tto
populate
(14 {p
plans positions.0,achieve
explosion
{p ,10 t.ysis 4
.0Zhang,
to + for
.,Zhang,
,a. .and
1)/10
p t{p
achieve
.L.; real-time
continuous
4Using}AtpL.; Understanding.
in
shown. each
}= goal the
shown. =
a time
= major 10,statistics
play-segments
(?).
interest
play-segments
Using and
step,
Using A
we Springer
{p
goal team
com-
thein .qua
liv
(e.gan
.Bp ,
ave
than e and and
a joinjoin
group
sequentialteamsteams
actions over
process over
in or- of cer. cer.
recently.
following a
However,
recognizing
forming However,
which
recipe The There
tactical has
t10
a used
significance
as
plans
group as
created
, soccer by soccer
has
tquestions ,ofoft12a been
of team
is a is
, tlow-scoring,
individual
these for an
consideringlow-scoring,
demand to
, plans achieve
explosion
Figure
14 ,agents.
trather
soccer: t15 to for
group real-time
3: a
continuous
, achieve
t16 Outside incontinuous
goal
actions
Figure the a of (?).
major statistics
interest
in
3.the
can or-
and
look A and
sport
goal team
com-
at in
the com-
(e.g. and
live-sport
a
short-term per- recipe
annotators vi- thisbehavior
then
winning process,
thisused
chunk tizedprocess,
sualizations.
this a tized
who by
resulting
of
process,
which
we
ball
match),
a resulting
a a
10
team can
ball
labelwe position
team has
N in get
can
we
11
position
from can =
five
Due in
all
play-segments
can get
created five
(14
a is be
actions used
achieve
play-segments
partic- is
play-segments
get to
13
said
play-segments used
10
the
play-segments
a
to
this + 14
demand to
thatpopulate
to a
1)/10 Zhang,
from
process,
difficulty be goal
15
populate
Hess, {p occurfrom
all
employing =
0 for
Zhang,
,
16 Zhang,
R.,.a.
from
the
(?).
we . 5 Zhang,
,
all
and
. L.;
4
Using
around
and
. a.
associated
, . pZhang,
and
Zhang,
posses-
play-segments
can .
the
all
real-time AL.;
and
Jones,
L.;
L.;
Using
, Fern,
}
4a.thep get Zhang,
Zhang,
and
L.;
Jones,
teamT
shown.
posses-} Zhang,
Jones,
Zhang,
and
and L.;
and
L.;
G.,
Jones,
=
tactics. Jones,
shown.
the
A. T
posses-
Jones,
and
L.;
play-segments L.;
and
G.,
eds.,
G.,
Jones,
per-
10,
= L.;L.;
ball
Using
2009.
with
and
statistics10,and
Jones,
eds.,
G.,
Jones,
eds.,
weG.,and
Intellig
and
G.,
Using G.,
eds.,
Jone
-we
Discr
Jon
eds.
Inte
Inte
ed
tra
Jo
GJo
Gef
der to isolate 11 team plans, 13 than a sequential process of forming
then chunk tized a a group
into
There ball N of
position
= has these
(14−10+1)/10 been is plans
used an to
.,explosion to populate achieve
= 0 5 play-segments in
4
aUsing
wethe major T
interest goal
= 10, (i w
han
group
ure ts. Outside 3: actions
Figure
a sequential of3. intheor- sport ofa recipe
process winning
annotators
der used
toforming
realm, isolate
winning
recognizing
sualizations. awho by
which
most match),
team a team
aaoflabel
group
ofThere
match), plans,
has this
plans
can
Dueall toProblem
created
of be achieve
actions
rather
work
has can
ofsaid
to
these been
be
individual
theathan
has to
that
demand
plans
said
aFormulation
abe
focussed
an
goal
difficulty occuragents.
sequential
to
toexplosion
(?).
employing
for
beaachieveemploying
Aprocess
Outside
around
onassociated
real-time
dynamic
team
acan
inagoal tactics.
ular the
major
the
of per-
region.
of thegoal
ball
with
statistics
teams
interest
tactics.
sport
By -can
tracking
not forming
(e.g.
in and chunking
which live-sport
sion winning
resultingvi- Ideally,
strings
athey
sion
resulting into then
group
sion strings in
then
this
play to
call
chunk
strings
we five
then
in
aof
represent
this chunk
process,
segments,
tofive match),
would
aball
process,
theseatointo
play-segments
represent
chunk awe
represent
actions.
play-segments
a team’s
into N
plans
this
like
can
we can a=
into N localcan
team’s be
(14−10+
get
ateam
ato
= behavior.
toteam’s
The
sion
get said achieve
play-segments
{s behavior.
(14−10+1)/10
recognize
ysis
0strings
play-segments
F24
to
,behavior.
ticle
ysis
We ysis
..1)/10and
.be ,Filters
soccer
ysis
and
sdo
,.aWe
and
ysis
Understanding.
to4employing major
team
ysis
this
}representand
Understanding.
=
Wefrom
do
Understanding.
and ysis
shown. ysis
for
shown.
and
asplay-segments
5from
this
=do
data all
tactics
ysis
Understanding.
5goal
Complex as
this
ysis
Understanding.
ysis
Understanding.
and andysis
aall
the
and
Understanding.
and
team’s
Using
weas
5feed
and
Understanding
Understanding
and
Springer
tactics.
the
play-segments(e.g.
Understan
posses-
we
for
play-segmen
Using
Understa
Understand
Springer
Understa
Springer
posses-
Multi-Obj
col-
Spring
behavio
soc-
Spri
Ber
SprSp
agents.
rmulation
quential Outside
processofof the sport forming 1. What a recognizing
group realm,
type mostthese plans
playing of plansof
this
style individual
to
workProblem
can achievehas
we agents.
focussed
expect major
Formulation Outside
from on can
dynamic
team? of
can
behaviorlook
can
look
(e.g.the
canlook
can
look
atmaybe
can
at
looksport
the
can
look
atthe
teamscan
at can
look
short-term
the
can
at
lookthe
lost can
short-term
atlookthe
look look
short-term
at
look
the as look
short-term
attheshort-term
at can
atat
short-term
the
we
this at
theshort-term
behavior
the the
atwinning
the
will look
short-term
behavior
process,
can
theplayers
short-term
behavior short-term
short-term
short-term
behavior
resulting
Ideally,
seelook
atshort-term
behavior behavior
ofsualizations.
the
in
sion
resulting
behavior of athe
behavior
of
sion team
which
we
atresulting theand
ashort-term
behavior
astrings match),
behavior
of
team
of
in
we
nextbehavior
behavior
team
of
strings
can aabehavior
from of
five
short-term in the
ateam from
would of
team afrom
has
subsection.
to get ofteam
five
a
arepresent
of aof Due
ball,
from
from
play-segments
to can
behavior
partic-
team
ofafromaaof
team apartic-
created afrom
team
like
represent
play-segments
play-segments
team
partic-
behavior
aafrombeto
data
Nateam
from
partic-
partic-
canaaof
=
partic-
from
to the
said
from
team’sa{s
partic-
from containing
(14−10+1)/10
aaaalook
from team
partic-
Hess,
recognize
of
partic-
0a{s
demand
team’s difficulty
to
aHess,
aHess,
.behavior.
partic-
apartic-
at .,R.,
apartic-
0team from be
from
partic-
Hess,
the s.R.,
Hess, and
.R.,
4
behavior.
},short-term
employing
Hess,
from saand
all R.,
for this
Hess,
and
team Fern,
Hess,}associated
partic-
R.,
R.,We the and
R.,
a.Fern,
Hess, =
Hess,
Fern,
and
and
real-time
shown.Wedo information
Hess,
Hess,
R.,
A.R.,
andFern,
Fern,
tactics
partic- Fern,
A.
do
this
posses-
shown. R.,
2009.
and
R.,
A.
and tactics.
Fern,
behavior R.,
R.,
2009.
A.
Using
thisas and
2009.
and
Fern,with
Fern,
A.A. and
and
Discrim
2009.
A.
for
as
we
Using Fern
Fern,
2009.
2009
statisDisc
of
weA.Fer
Disc
A.
200
Fer
socatriD 2
sed
gents. on dynamic
Outside of teams
the sport which(i.e.
recognizing Ideally, they
winning players
where
In this we
callmost
plans
sualizations.whichandwould
ball
aindividual
paper,
of
match), the actions.
individual
has
we like
ball,
Dueagents
cancreated
are to beThe
data
to
looking recognize
agents.the
can
said a F24
containing
leave
demand
to soccer
Outside
difficulty team
and
toautomatically
be employing
forof data
this
join tactics
the
associated teams
ular
real-time feed
sport
information
ular ular
aansweringfor
region.
ular
=region.
ular col-
60with
over
tactics.
ular
region.
region.
ular soc-
region.
ular
By
aregion.
ular
By
statisticsular
ular
region.
ular
not
By
aregion.
winning
By
2the
is
tracking
ular
not 9lected
,region.
By region.
region.
not
chunking
region.
By notstillcer.
region.
chunking
By
not ular
chunking
not
By
49and
chunking
,By not
achunking
By this ByBy
chunking
not a
However,
region.
for
into
not By
chunking
not
9ascarce.
match),
can
notprocess,
not
into
chunking
6vi-
into the
not
chunking
aplay
look By
chunking
chunking
into
chunking
play
playersinto ainto chunking
play ,English
segments,
intonot
play
9Most
at
play can
aswe
segments,
intoplay
segments,
,into
the
aplay soccer
chunking
into
and can
into
segments,be
segments,
into
play short-term
segments,
,playof in
intothis get
segments,
play said
Premier
play
the play five
segments,
this
the play is
segments,
thislocal play-segments
into segments,
segments,
ball,local
this to
play-segments
,low-scoring,
segments,
this local
data this behavior
segments,playbe
this
local
ular local
,data League
local
this this employing
segments,
local
ticle
collected this
region.
Hess,
this
local this
local
ticle of
this
ticle
containing
{s
local
Filters locala
from
local
ticle R.;
0 team
local
Filters
ticle
By
,
continuous
(EPL){s
Filters
.
aticle this Fern,
.
for
Filters
is
not all
,4tactics.
. .
Filters
ticle
Filtersfrom
,
ticle for
s
. the
local
Complex
viaFilters
for 4
by
ticle
this
chunking ,}
A.; s
ticle a
Filters
Complex
for ticle
Filters
Complex
ticle
for
anfor and
partic-
posses-
} Opta shown.
and
Filters
4 Filters
Complex
for Mortensen
Filters
Complex
Filters
Complex
Multi-Object
for
for
Complex
information
army into com-
(?)
for
Multi-Ob
Multi-Ob
Complex
for
Comple for
for Usi
Multi
ofCom
play Com
MulMu CT
Co
M
h s
m ocussed
Outside pect from
Formulation on a dynamic
of theteamsteam? teams
sportover2: winning realm, realm,
afor Ideally,
match),
most (i.e. ofwhere we
can
this ofwould
individual
be of
work thissaid has work
liketo agents
beto
focussed has focussed
recognize
can
employing ontheleavedynamic on
team
and
tactics. dynamic
tactics
join
teams teams{a ,teams
for 15, soc-
over a 3 , asion
Ideally,
5ular
cer.,strings Ideally,
this ,region.
ular
we However, 9can
this
7 ,process,
region.
would to can
8
this
look
sualizations.
process, By
arepresent we
9look
process, Byat
not
14 10
like
we
as would
the at not wecan
achunking
15
soccer
11 the short-term
to
,achunking
we
acan get
12 15 like
short-term
team’s
arecognizeDue can
aget
isplay-segments
13into 12
inget
toa behavior
to
behavior.
play-segments
low-scoring,into 14 play recognize
}behavior
12 the
Parts
play-segments play team
12difficulty
segments, of
Pictorial We
segments, from
of 10
tactics
0team
afrom
used
continuousteam team
thisall from
Structures
from
the
this
all
this
for tactics
from
associated
local posses-
alocal
the
all
partic-
ap-
soc-
aand
for posses-
the
for
partic- com
Object
posse wi s
ve
automatically
ncussed
and join
leave on dynamic
and join answering
Figure
teams teamsover the
(From cer.Are
lected
2. However,
the
cer. scarce.
(i.e. there
Ideally, the
XML
following
However, players
where any In as
English
Most
we
sualizations. soccer
areas
feed,
this and would
tactical
individual
as of
paper, Premier
the
soccer the
we isball,
the we
questionsdata
can
like
Due field
agents
is Problem
low-scoring,
are League
data
totocollected
which
infer
looking
low-scoring, containing
recognize
for
the
can they
soccer:to Formulation
continuous
(EPL)
difficulty
leave is tendviathis
automatically
and team
continuous by
join
behavior
an
to behavior
Opta and
behavior
behavior
utilize
atactics
associated
teams
behavior
information
army com-
behavior
maybebehavior
maybe
behavior
(?)
answering
and
maybe
of
for
over
behavior
behavior
maybe
behavior
behavior
maybe
lost
a,with
com-
behavior
maybe
lost
humansoc-
maybe
lost
as
isis maybe
lost
aswe
maybe
lost
,the
behavior
maybe
as maybe
astill
maybe
lost
we
tracking
proach
as
will
lost
we
,good
maybe
as
sion lost
will
aswe
,10abehavior
acer.
we
lost
will see
as
we
lost
asion
,14a,,a6annotators
lost
lost
will
as seemaybe
will
we
strings as
,2a25However,
,3as
behavior
lost
in
see
will
example
we
as
scarce.
ular
,sionit
as as
in
we
will
the
see
,a336strings
see
we
in
ular
maybe
allows
,sion
aswe
will
see
the
we will
next
region.
44strings
itplayers
lost
thein
to seewe
in
will
maybe
will
innext
will
seethe
the
region.
next
,a6a9,a5,astrings
see
represent
to
inas
will
the
subsection.
Most
,,lost
who
see
of
as
,7us
see
innext
the
seewe
subsection.
Byin
,6represent
,to
next
,lost
see
subsection.
the in the
,,7,this.
as
soccer
to
and
in
next
By
will
in
not
label of
we
,represent
a,to
in
the
subsection.
subsection.
thenext
subsection.
analyze
,as
the
next
not
the
see
the
subsection.
,the
next
team’s
chunking
will
we
,represent
next
,The
subsection.
next
aall
,ais
next
subsection.
behavior
willa,data
team’s
chunking
ball, see aa14
the
subsection.
subsection.
actions
low-scoring,
,the }F24
,see
,team’s
subsection.
subsection.
behavior.
Hess,
in
a,a14
next
into
In
data
Hess,
collected
,the
,short-term
team’s in
Hess,
maybe
,behavior.
into
aaICCV. data R.; subsection.
Hess,
play
anext
}the ,behavior.
Hess,
that R.;
play
containing
We
Fern,
R.;Hess,
next lost
is
segments,
,subsection.
behavior.
Fern,
Hess,
R.;Fern,
Hess,
We,is
occur
used
R.;
R.; A.;
aa
asR.;
segments,
continuous
behavior
Hess,
Fern,
subsection. via
We
Hess,
Fern,
Fern,
we
time
used this
Hess,
A.;Hess,
R.;
A.;
and
R.;
Fern,
We
anA.;
and
will
this
around
used
this
ap-
R.;
Fern,
R.;
and
Fern,
Mortensen,
A.;
A.;
this R.;
R.;
A.;
and
coded
army
this
of
used
Fern,
Fern,
see
local
this
informa
Fern,
Mortense
Mortense
and and
A.;
Fern,
A.;
ap-and
Morte
inMort
local
and the
ap-
this
and
A.;
Mo
ofA.
andMA
the
co tEA
aO
h
ng on
m dynamic
Formulation
they tend
to automatically to teams
utilize answering Ideally,
is (i.e.possession
a good
morewhere we example
than would
individual
scarce. others? like
Most of agents to
this.ofthe recognize
can
the The dataleave
F24 team
and
collecteddata tactics
join
is teams
a time
is via this for an
a=
over
Entropy soc-
a={aa=
coded
army
{a
=
a {a
0 ,Maps
a a{a
=
0 ,=
a0a
1 {a
,,aa{a
=
01 a=
01a
2
ofcom-
,{a
,aa=
cer.
0 2a
1 {a
,a=,31a
==
02
human
a{a 3a,a{a
,1,0{a
2 ={a
,402a
13aa0,2,10{a
However, ,4a
3,a
0 a ,a513a
24aproach ,=
3 a
21
5 2
4,35a{a,1a
2
4 a
2
6
proach
,,a
7
3
04,6a
5 a a
2,behavior
3
5a4aas
as3
7
annotators
,,84a
5,1
6 7,aa ,a
3
6 4a a
5
behavior8,a
soccer
as
4
7 ,9,57a
6a
8,2 ,a,a7,4a
allows 5
it
,10
55
8,8a
6 79a ,3maybe
a a8,a56aa
allows
a
,10
6,a
6 a
9
7810a ,4a
is11
maybe9a
us
6 a
7
who
,a
8 7 a
810
11
9 aa
,5toa
lost
712
11
108
us
a9,8,aa
low-scoring, 9
8a
,analyze
10 a11
a,12
11
lost
6 aa
to
label
89,a
as
12 ,a13
11
10 9 9a
,a,10
a a,7we
11
analyze
a12
9,13
10
as12 ,a
10
13
12
11 ,a,11
10
a we
all
a,10
a
will
8 12a ,a14
a
the
13
= 11a
Parts
13 11
13
12 a}912
,11
will {a
actions
a}Parts
,see
a
11
13
the
a ,14
continuous
Parts
12 a
short-term,14
14Pictorial
12
13
10 12
13
,see
}14
Parts
in ,aParts
12 ,aa
short-term
a 13}a
Pictorial,a
Pictorial
the 13
,14
in
11 a13
Parts
14
a,that
}13
,a}Parts
2the
aPictorial
,next
,a
Pictorial14 a14
Pictorial
Parts
Structures
12
a11
}14
14
3next
}Structures
behavior
Pictorial }
Parts
occur a
}
and
,,Structures
a
Parts Parts
Pictorial
Parts
Pictorial
subsection.
13 ,a com- aof
aPictorial
,Structures Pictorial
Structures
Pictorial
Structures
for }Structures
Pictorial
,Structures
subsection.
behavior 14
around
for
Objects
Structure
for
6a,14 aof ,Stru
Objec
for
Struc
Objec
for Str
aStr
for
Ob
fo
8},e
th
w
ball position
overcer. the and at
following every tacticaltime step
questions (solid for soccer: feed -Tthe that lists {a all 0 player action 5events
}
awithin the 11game with
a = , a , a , a , a , a , a , , a , a , a , a , a ,
soccer:
leave and joinover teams cer. annotators However, players In who
this asof and label
soccer
paper, weall
isball, are actions
low-scoring,
data
looking that automatically
containing occur
continuous around and
information the ball a team �
isaMaps which
still over a proach = a 0 particular
they
{a
scarce.
1 as1call
, a 2 it,asaallows
, a 2 ball
3
Most
, region,
a 3 us
4 ,
12, a23of
a 4 actions.
5 , a to which
6
4the
, analyze
a 7 ,
,data
a IBM 8 maybe, The 0
a 9 SlamTracker.
the
67collected
, a1
9F24
10 lost
short-term a , otherwise.
soccer
a 4,2012.
12
a,is
a behavior
5
via
13
adata
, a 13 feed
www.aus
,an
, 7a
army ofa
in-to answering = team 10
=Tover proach =aa particular 1it ,aaallows region, aa3us ,awhich Tto
6Inanalyze ICCV.
7In
,,In maybe ICCV. 8ICCV.
,In In a7ICCV. 8the
In
,ICCV.
ICCV.
,aalost In
ICCV.
In
10 short-term
, ICCV.
aICCV.In ,InIn
otherwise. InaICCV.
ICCV. ICCV.
ICCV.
12 behavior 14
nd s for join teams
soccer: lines and dots feed However,
are thatscoring 1.lists
annotated, What as all
annotators soccer type
player iswho
dotted low-scoring,
playing
action label lines style
events
all are continuous
can
actions withinwe expect
that the and
occur from
Entropy
Entropy
game com-
Entropy
Entropy
aroundaEntropy
Entropy
with team?
Maps
Entropy
Entropy
Maps Maps
Entropy
Entropy
Entropy
Maps Maps
H(X)
the Entropy
Maps
aMaps Maps
ball =
player, MapsMaps Maps
Maps
s� -�
T a{6,
p(x)
team, team awhich
10
log =a.over
team p(x)
10 event
{a
over=
they a0{a particular a0type, ,particular
call
, a (1)
ball
minute
,,region, ,region, a
actions. 45,=
aen_AU/ibmrealtime/index.
awhich
and 5610 awhich ,a
The maybe
second maybe 89, alost
F24 910
for soccer
lost
10 11, a,otherwise.
otherwise.
each 1112 a, 12
data
a13
ac-
, a, 13
fe r1
king to automatically answering 3. the
In which they
scarce.
1. following
What
situations call Most
type ball
wheretactical
ofof actions.
the
playing
are questions
data
their The
style strengths F24
for
collected
can soccer:
we soccer
and is
expect via
weak- data
froman army
a feed team? col-
of � p
human
0 = lected
sp� 00� � = � 5,
{6, T� for
aannotators
� .team
= 5,
.�
�T
,� 10
.
15}
= �
. the . over
,
10 15} English a particular
who label Premier region,
IBM all IBM IBM SlamTracker. IBM
actions League
which
SlamTracker.
IBM SlamTracker.
IBM SlamTracker.
IBM maybe
SlamTracker.
SlamTracker.
IBM SlamTracker.
thatIBM IBM(EPL)
IBM
SlamTracker.
2012.
IBM
SlamTracker. lost
SlamTracker.
2012.
occur www.austr by
otherwis
SlamTracke
SlamTracker.
2012.SlamTracker
2012.
2012.
2012.
2012.
www.au aroun
www.au Op
2012
2012
www.
wwwww w2
n we expect from ferred). a team? a player,2.team, tion.
x∈X
ps Each =H(X) {5, s
p
event lected
00 9, s= . . {6, .s=0has , 12} for5, . .
a5,
.
the , 15} English (1) (1) (1) (1) (1) (1) Premier(1) Intille, (1) (1) (1) (1) S.,
(1) (1) League and Bobick, (EPL) A. 1999.
by O A
which they call ball actions. The F24 soccer data feed col- . ,series ofEntropy qualifiers describing it.a Ev-
H(X) H(X) H(X) H(X) =
H(X) H(X)
= H(X)
= H(X) = H(X)
=p(x) = H(X)
H(X)
p(x) = p(x)H(X)
log = p(x) = p(x)
log =
p(x)
log= = p(x)
p(x) log= log
p(x)
p(x)
log p(x) p(x)
log
p(x)
p(x) p(x)
p(x)
log p(x)
log log log
p(x) log
p(x) log
p(x) p(x)
p(x) p(x) en_AU/ibmrealtime/index.ht
en_AU/ibmrealtime/index
en_AU/ibmrealtime/index
en_AU/ibmrealtime/ind
en_AU/ibmrealtime/i
en_AU/ibmrealtime/in
en_AU/ibmrealtime/ en_AU/ibmrealtim
en_AU/ibmrealtim en_AU/ibmreal
en_AU/ibmrea
en_AU/ibmrealt
en_AU/ibmrea
from a team? nesses? lected Are for event
there
annotators theany type,
English areas
who minute ofPremier
label the and fieldsecond which
League for
they (EPL) eachby
tend to ac- utilize
Opta (?)
1
is
sp11 as= good
{5, p0 9, 0
which .p
. {6, .=example. ,.{6, 12}
.they
.5, . . ., .15}
. . ., call
. ,of 15} ballthis. actions. The Maps F24
S.,The
data F24 issoccer timeda Vic
ons
yle for we
can
strengths soccer:expect
and weak- ofallthe actions that occur around the 2ball -is .{5, nizing Multi-Agent Action from
p = {9, p 19, s= .1x∈X ,012} 9, 12}
lected
more 2.1.than Are What
for there
the
others? type anyofareas
English playing Premier style field canwhich
League we they tend
Analysis
expect
(EPL) Using
fromby toOpta utilize
aEntropy team?
ery event
x∈X
(?)
sx∈X 2 Maps
x∈X
spx∈X
feed collected
x∈X x∈X
=
1x∈X
{9,
that
pa
x∈X 1 x∈X 9, sgood
= x∈X
x∈X
x∈X
. x∈X
11lists {5,
.= by
. ,.{5, 9,
12} example
Opta
all
.9,
player .12} .for , 12}a ofaction given this.
Intille, Intille,
Intille,
match TheIntille,
S.,
events
Intille, S.,
and
Intille, F24 Intille,
and
Intille,
S.,
and
isBobick,
S.,
S.,
within data
Intille,
andIntille,
S.,
Bobick,
Bobick,
and
listed Intille,
andIntille,
S.,
and
S.,
Bobick, is
Bobick,
A. S.,
Bobick,
and
S.,
andBobick,
A.
S.,
1999.
within
the
S.,
A. andaand
Bobick, time
and
Bobick,
and
1999.
A.
1999.
A.
game
Bobi
A. Bob
AA.
199Bo
Bo
19 F1A
ld which they tend to utilize tion. Each isdefine event
a good which has
moreexample a they series
than call
others? of
ofofball qualifiers
this. The
actions. F24describing
The data F24 Determining
it.
isAnalysis
soccer Ev-
aAnalysis time data
Parameters coded feed
ps3 col- 3 =
spMaps
22{9, s
p
=
2 2
feed
9,
{9,
=
sp22 9,. lected. {9, .p
s=
,
that
12}
. Maps
9,
{9,
.Maps
=.of ,.lists
. for
9, .9,
12}
. , 12}
. .all the., .12} . ,player English nizing Kim,
action nizing Premier
nizing K.;
Multi-Agent
nizing nizing Multi-Agent Grundmann,
Multi-Agent
nizing
events H(X) nizing
Multi-Agent
nizing League
Multi-Agent
Multi-Agent nizing
nizing
Multi-Agentnizing
nizing
Action
�
Multi-Agent
within
= Multi-Agent
Action M.; (EPL)
Multi-Agen
Action Multi-Age
Multi-Agent
Multi-Age
Action
from Shamir,
Action
the
p(x) Action
Action
from
fromVisua
Acti
gam
log Act
from
frp
V V bf
he
style fieldcanwhichwe expect they tend
from to a utilize
team? We is a a team
good as
example a set of agents this. having
The F24 a sharedAnalysis
data Analysis ob-
Analysis
Analysis
is Using
a Analysis
Analysis
UsingAnalysis
Using
time Analysis
Analysis
Using Analysis
Using
Entropy
Using
Entropy Using
this Entropy
coded Using
EntropyUsing
Entropy
Using Using
Using
Entropy
feed.
p Maps Using
Entropy Entropy
Maps
= a Entropy
Entropy
33{9, Entropy
Maps
Entropy
The
player, s
p Maps Entropy
Maps9, =Maps . type .Maps
{9, .p Maps ,22Maps
team,10} 9, Maps
{9, . . events
, event
12} 12}listed
type, are:
minute
gins, J.;goals, and and Essa, shots, second I. passes
2010. for Motioeac
tion. Each event has ery event a feed series collected
3.andInthat scoring
lected2. by
Are
listsmental Opta
there
all the
situations
for player for
any a
where
English given
areas
action of
are match
the
events
Premier their field is listed
which
within
strengths
League within
they
theand
(EPL) tend
game
weak- by to with utilize
Opta
4
p
(?) =
paMaps
3 3
4 player,
{9, sp is 9, = a . {9,
. good
. , 9,
10}
team,
. . . example
, 12}
.event 12} type, of
Kim, Kim, this.
Kim, K.;
minute Kim, Kim,K.;K.;
Grundmann,
Kim, The Grundmann,
K.;Kim,
Grundmann,
Kim,
K.;
K.;
and K.; F24
Kim,
Grundmann,
Kim, Kim,
Grundmann,
Grundmann,
Kim,
K.;
K.;
Grundmann,
secondM.;data
K.;
K.; K.;
Grundmann,
Grundmann,
K.;
M.;
M.; Grundma
Grundman
Shamir, is
Grundm
Grundm
M.;
Shami
Shami
M.;
for M.; a
M.;
Sh S
eaA MS
aof qualifiers describing it. This 3 3 s33 9,
p = .{9,
jective shared state (?). As we are their dealing How
Determining discriminative
with
Determining
Determining
Determining
Determining
Determining
Determining
Determining
Determining
Determining
Parameters
DeterminingDetermining
Parameters
Parameters are
Parameters
(with Parameters
Parameters
Entropy our
Parameters Parameters Parameters
start/end representations?
4
Parameters
Parameters
Parameters Parameters = {9, .That ,.9,
.tackles, 10} .is, . ,do Evolution in Dynamic x∈X
Sports Scene
feed 3.that Inmore scoring
lists all situations
player where
action are
events strengths
within the andgame weak- with Entropy tion. Each pMaps 4point), = {9, event 9, . .has , 10} a series clearances, gins, gins, ofJ.;
gins, gins, qualifiers
gins,
J.;
andJ.;gins, cards,
and and
J.;Essa,
gins,
J.;
gins, and
Essa,
J.; Essa,
and
gins,
and gins,
J.;
I. free
gins,
and describing
gins,
J.;
Essa,
Essa,
Essa,
2010.
and
I.and
I.
J.;J.;
Essa, kicks,
J.;
J.;
2010.and
2010.
and
I.Essa,
Essa,
I.I.and
and 2010.
MotionEssa,
I.2010.
Essa, I.Essa
2010.
Moti
2010
I.Essa
Moti 202MiI
having a shared type ob-
of data is this
currently feed. a each The
player,
used nesses? type
is ateam,
for the of real-time than
events
event others?
listed
type,fixed
online are:
ofminute
visual- goals, and shots,
second
teams passes
have for unique each styles ac- Entropy of play? Entropy Maps And
tion. feed if p
s Each Maps
so, = that
can {9, event we 9,lists . detect
. . ,
has all 10} player
a series action of qualifiers events within
describing the
rethe fieldstrengths
which they tend to utilize good example this. The F24 data is adiscriminative
time coded
soccer, agent isteam, permanently to one team. How Unless
How How
Howdiscriminative
How
discriminative
Howdiscriminative
How How
discriminative
How
discriminative
How
discriminative
How
How How
discriminative
discriminative
discriminative
are discriminative
are discriminative
are
our are our Entropy
are ourare
representations?
ourare
representations?
our are Entropy
representations?
our are our are
representations?
are representations?
are
our Maps are
representations?
our our
representations?
our our 44
our
representations?
representations? Maps
representations?
representations? representations?
representations?
That That That is,ThatThat That
is, dois,That Laviers, K.;Dynamic inSukthankar, G.; Molin
their andof weak- nesses? corners, ery offsides, event substitutions
collected by ado is,
That
dois, That
is,
Analysis
Opta and do
That
is, do That
do Thatis, That
is,
Evolutiondo
for stoppages.is,
do
Evolution is,
do
Evolution is,
Using is,
aEvolutiondo do do
Evolution
Evolution
given indo
Evolution Evolution
in Evolution
Entropy Dynamic
Dynamic
match An Evolution
Evolution
in inin Evolution
Evolution
Dynamic
in
Dynamic
Dynamic
Sports
Dynamic
inin
example
MapsSports
isSports
Dynamic
Dynamic
in in inDynam
Scenes.
Sports
in
Dynam
SportsDyn
Sport
Dyn
Scen
Spor
Scen SS
islisted w
The information content of random discrete variable,
asapost-analysis player, event type, minute and second for each ac-
when they deviate from this style Entropy and aofevent
quantify Maps
player, it? � team, event type, minute and X
second fM
here we are their
are dealing with
strengths
izations andevents, (with
weak- as well
an start/end
agent tion. has Each feedpoint),
3. In
been event
that scoring
dismissed tackles, has
lists for afrom clearances,
series
prominent
situations
all player the of qualifiers
tele-
where
action
match, cards,
are events
each their free
describing
teams
teams
team
teams
teamskicks,
have
teams
strengths
within
have
teams
al-
have
teams
have
unique
teams
have
unique
teamsteams
teams
have
teams
unique
the
have
and it.
teams
unique have
unique
styles
unique
have
game Ev-
styles
have have
unique
styles
weak- with
have
unique
styles
of unique
styles
unique The
unique
unique
styles
of play?of
with
a
unique
styles
play? The
styles
ofplay?
of information
styles
probability
of And
styles
play?
The
Entropy
ery
styles
play?
styles
of information
play?
And styles
of
And play?
if
Entropy
play?
And
of And
information
H(X) so,
if of
play?
of And
if of
play?
so, play?
And
canplay?
so, ifMapscontent
distributioncollected
play?
ifAnd can
if = And
so,
can
we if
so,
Maps
And content
AndAnd
we if
so,
can And
detect
we
if
can so, if
�
content
p(x)
of
if
detect
so,
can wewe
detect
ifwe so,can if
so,
canaby
so,we of
detectso,
detect random
detect
logcan we cancan
we Opta
adetect
of
can
p(x) random
we
Laviers,we
detect 2009.
detect
a
we we
Laviers,detect
random for
detect
Laviers,
discrete
detect
defined
detect
Laviers, Improving
adiscrete
Laviers,
K.; Laviers,given
K.; K.;
Sukthankar,
Laviers,
Laviers,
discrete by
variable,
K.;
Sukthankar,
Sukthankar,
Laviers,
K.;
K.; Offensive
match
variable,
Laviers,
(1) Laviers,
Laviers,
K.;
Sukthankar,
Shannon Sukthankar,
Sukthankar,
K.;
K.;
X
Sukthankar,
G.;
variable,K.; K.;
Sukthanka
Sukthanka
K.; Perform
X
K.;
G.;
G.;listed
Sukth
Molinea
Suktha Suk
G.;
Sukt
Mol
Mol
G.;GG
Wetion. define
EachWe a
define team
event a as
has team a a set
series
as of
a setagents
of of agentshaving
qualifiers when
when
having a
when
when
theyshared
describing
when
they
when
they
when
deviate
they
a when
deviate
they
when
they
shared when
deviate
when
when ob-
deviate
they when
deviate
they
fromdeviateof
it.
they
from
they
deviate
they
fromdeviate
ob- the
Ev-
they
this
from
deviate
from
deviate
thisdeviate with
deviate
from this
style data
deviate
from this this
style from
this style
this
withfromandafrom from
style
this
from
style
andfeed
feed.
probability
this
style from
and
quantify
a this Entropy
style tion.
this
and this
quantify
and
style
this
probability quantify
and is
this
style The
style
and style
quantify
quantify
style
H(X) given
Each
quantify
andit? style
and quantify
it?
and Maps
type
distribution
and
it?
quantify
and and
quantify
= it? eventin
quantify
quantify
distributionit?
quantify of
quantify
it? Figure
�
it? p(x)was
it?events
Determining
p(x) has �
it?
p(x)was it? it? it?a2009.
log Modeling. 5(a). listed
series
p(x) defined This Parameters
defined are:
of In by
AAAI type
qualifiers goals,
Shannon
by (1) of
Conference
Shannon data
shots,descri
Xon pO
to one team. Unless vision and newspapercorners,
ways has ery offsides,
entities event
the asame
(e.g. substitutions
collected
nesses? ESPN,
anumber by The
of by Opta and
agents
Guardian).
forstoppages.
(11). a given An
match Mutualexampleis Information
listed within (?) this
with a feed. probability TheH(X) type =
distribution of events
p(x) p(x)was
2009.
� log2009. Improving
2009. listed
p(x) 2009. Improving
Improving
2009. 2009.
Improving
2009.
defined are:
Improving
Improving 2009.2009.
Improving
Offensive goals,
2009.
2009.
Improving
Improving
Offensive
Offensive
by Improving
Improving
(1)Improving
Offensive
Offensive
Improving
Offensive
Shannon shots,
Performan
Offensive
Perfor
Offens
Perfor
Offen Pe
O P
aWe refer to team
Figure 3: Given a possession string, x∈X we break p(x)was it up (1)
jective ery player,
and
event andteam,
shared
collected mental event Optastate type, for(?). minute
As
given weAs and
are
match second
dealing is listed with for iswith
within each
currently (?)(with
� ac-
(?) (with used start/end
The ery for
information
event the point), H(X) real-time content
=
tackles, Modeling. of online
p(x)
Interactive
Modeling.
Modeling.aclearances,
log
Modeling.
Modeling.
Modeling.
random
p(x)visualizations
Modeling.
In Digital
Modeling.
AAAI
Modeling.
In aIn AAAI Modeling.
Modeling.
AAAI
In
discrete Entertainment.
cards,
Modeling.
In Modeling.
Conference
In AAAI
In
AAAI
AAAI
Conference
Conference
AAAI
InInrepresenta
Conference
AAAI
AAAI
variable,In
Conferen
In of
free
ConfereIn
Confere
InAAAI
onAAA AA
Con
AAA
Con
Art
on
on lik
thecollected by Opta for given match is
x∈X
jective aused, shared mental state (?). we (1)
where are their strengthsthoughand ofweak- ofare dealing start/end point),
H(X) How =tackles, discriminative p(x) clearances,
log p(x) are our cards, free
match, agents each
having team
Even al-
a shared ob- this thedatadata
behaviors, this
has
soccer, feedbeen
feed.
B, tion. as
each
widely
isThe
short, given
Each agent type in
observable
eventis of Figure there
events
permanently
has segments
a
are
5(a).
listed
series
no
fixedThis are:
of
of to type
one
Mutual
goals,
coordinated
qualifiers
Mutual
into Mutual
Mutual
team.
Mutual
N Mutual
Information
dataMutual
shots,Information
Mutual
Information
equal Mutual
Unless
describing
Information
Mutual
Mutual
H(X) Mutual
Information
Information
passes Information
Information
chunks
events,
=Information Information
−Information
Information
Analysis can
it.
� Information
� can
as
�
Ev-
be
corners,
p(x)
� and
well
� can
(?)
be
Using
� log
� be
with � as �p(x)
use Entropy
offsides,
� � �post-analysis
a � probability quantized
Maps substitutions
(2) x∈X
distribution for
x∈X Laviers,
prominent
Interactive Interactive ball
Interactive Interactive
and Interactive
Interactive K.;
Interactive
Digitalpo- Sukthankar,
Interactive
stoppages. television
Interactive
Digital
Digital Interactive
Interactive
defined Interactive
Digital
Entertainment.
Interactive
Digital
Digital
Digital
Entertainment.
Entertainment.
Digital
Digital
by G.;
An and
Entertainme Molin
Digital
Entertainm
Entertain
Digital Digita
Entertai
Digita
Shann Ente
EnteE
exa
set of agents having
systems a shared ob- thisorfeed. soccer,
data TheWe define
each
like type agent a team is
of events as
permanently a set of
listed are:sition agents
fixed to having
one
goals,values team.
shots, a shared Unless passes ob-
The Analysis information corners, can this
Using be feed. Entropy
offsides,
content, The Maps
namely type teams
substitutions ofx∈X have
(Shannons) events unique
p(x)was and listed styles
stoppages.
entropy are:of play?
goals,
of AnAnd sh
ex i
1). We refer to team which withiswith
use this data
currently used for the this for
real-time automatic
aonline visualizations H(X)
ofdealing H(X) H(X) H(X) =
H(X) H(X) = −H(X) =− H(X) −
x∈X
= H(X)
=H(X) =− The −
H(X)
H(X)
p(x) =−H(X)p(x) = −
p(x) =information
log − = p(x) =−
p(x)
=
log −
p(x)
log = −− p(x)
p(x) −
log
log
p(x)p(x)log p(x) p(x)
log
p(x) p(x)p(x)log content,
p(x)
log logp(x)log
log
p(x) log(2)
p(x) p(x)
p(x) (2)
p(x)(2)
namely (2) (2) (2) (2) 2011. (2) (2)Laviers,
(Shannons) (2) (2) (2)
Improving (2) entropy Offensive of Perform
l (?). stateAs (?).weAsare wedealing movement (with an (with agent and
start/end action
has been executed
point), dismissed tackles,by fromteam clearances,
the (e.g.match, pass cards,from
each teamfree al- kicks, as
newspaper our
Determining �of Analysis
play-segment The
the entities
Analysis (?) information
Parameters
data Using (e.g. Using
feed Entropy ESPN, iscontent,
representation,
Entropy given when Maps
The namely Laviers,
Maps
in they Laviers,
Guardian). Laviers,
Figure (Shannons)
deviate Laviers,
K.; Laviers, K.; K.;
Sukthankar,
Laviers,
Laviers,
s.
from
5(a). K.;
Sukthankar,
Sukthankar,
Laviers,
K.;
K.;
Even Laviers, entropy
Laviers,
Laviers,
K.;
Sukthankar,
thisSukthankar,
Sukthankar,
ThisK.;
K.;
Sukthankar,
G.;
K.;
though
style K.;
Sukthanka
Sukthanka
K.;
type of
K.;
G.;
G.;
and Sukth
Molinea
Suktha Suk
G.;
Sukt
Mol
Mol
G.;
quG
of G M
tacticalare dealinganalysis. ery anjective event
agent
start/end and collected
has abeenshared
point), by
dismissed mental
tackles, Opta for
state
from
clearances, a(?).
the given As
match, match
we
cards, are
each isH(Y
free team listed kicks, al-
a with within
discrete
x∈X
a� Determining
x∈X
discrete
x∈X
aDetermining
x∈X
discrete
x∈X x∈X of
random x∈X The (with
the
Analysis
random
x∈X x∈X x∈X information
random
x∈X
x∈Xx∈X data
Parameters variable
x∈X start/end Using
variable feed variable X content,
is
Entropy
(3) X thatpoint),
that given X that has 2011. namely
hashasModeling.Maps tackles,
in
2011. a
2011. Figure
probabilitydistribu-
a probabilitydistribu-
Improving
2011. 2011. (Shannons)
Improving
Improving
2011.
a probabilitydistribu- In
2011. clearances,
Improving
2011. 5(a).
Improving
Improving International
2011.2011.
Improving
Offensive 2011.
2011.
Improving
Improving
Offensive entropy
This
Offensive Improving
ImprovingImproving
Offensive
Offensive
Improving
Offensive cards,
type
Performan
OffensiveJoint
Perfor
Offens
Perfor
Offen ofPe
O P O
ments
aly fixed
set ofoffixed
coordinated
to
agentsonetohaving
team. Unless
awork
shared events,agent ob-as Atheways
corners, well
to 2010-2011 agent has
thisas
ways post-analysis
B).
the
offsides,
soccer, feed. has These
same the
each The behaviors
number
substitutions
same
agenttype fornumber
isof prominent
of are
agents
permanently
events observed
and
of agents (11).
listedtelevision
stoppages. from
We
(11).
fixed are: par-
refer
toWe one An
goals, and to
refer team.team
example toH(Y
shots,
) = −�
this
team Unless How
tionpX data
passes
�
is
p(y)
has
discriminative � � a�
currently
=(p1,...,pn)isthendefinedas:
log
Determining � been
discrete � p(y)
� can � � � widely� be are used
random Parameters our Parameters for used,
representations?
variable the there real-time
X(3) Intelligence
that are hasThat no aonline systems
is,
Conference. do
probabilitydistribu- which
visualizatio
anently onethis
For team. Unlesswe used corners, offsides, EPL seasonsubstitutions F24 Opta and stoppages. An H(Y example
H(Y )H(Y =H(Y
) H(Y = −=
)teams H(Y)− )H(YtionpX
= −
y∈Y How
)=
H(Y −
H(Y
= −
H(YtionpX
H(Y
)have
p(y) =)− H(Y
p(y) )= −
p(y) )is
=log)− )=
unique p(y) =−
p(y)
log
= currently
corners,
=(p1,...,pn)isthendefinedas:
discriminative =(p1,...,pn)isthendefinedas:
Determining
)p(y) −
log = − − The
p(y)
p(y)
log −
log
p(y)p(y)log p(y)
styles information
p(y)
log
p(y)
p(y)p(y)p(y) log log
p(y) are
of
used
offsides,
logp(y)log
log
p(y) our(3)
Parameters
play? log
p(y) p(y)
p(y) for
(3)
p(y) (3)representations?
Mutual
content,
And
the
substitutions
(3) (3) (3) Modeling.
if so,
real-time
Modeling.
(3)Modeling.
Information
(3) namely can
Modeling.
(3) Modeling.
Modeling.
(3) (3) Modeling.
(3) we
In Modeling.
That Modeling.online
and
International
InInModeling.
(Shannons)
detect
Modeling.
is, stoppages.
Modeling.
International
International
In
Modeling.
InIn do visualizat
International
InInternational
International
In
InternationaInInIn
In Internati
Joint
entropy
In
Internati Inte
Joint
Inter
JointCoIn
In JoA
om eam
tal from
ed the (e.g.
match,
state (?). passAs
the match, from
each
we which team
are
eachdealing al-
team
tial
newspaper
with
al-
spatio-temporal
of behaviors,
the entities data
behaviors,
(with an (e.g.
B,feed
agent tracings,
start/end as ESPN,
short,
is
has
B, given
as which
been The
observable
short,
point), in in
dismissed this
Guardian).
Figure
observable
tackles,
case
segments
from refers
5(a). Even
segments
the
clearances, of
Thisto
match, ball
though
coordinated
oftype coordinated
each
cards, of teamuse data � this In al- �
teams [85],
In data
events, How [85],
In tionpX
How
events,
have or
Shannon[85], discriminative
of as
Shannon data
unique discriminative
the well
Shannon as like
=(p1,...,pn)isthendefinedas: uses well
styles
data as
uses thisprobability
are
post-analysis
as
uses of
feedprobability are for
our
post-analysis
play? probability is our automatic And
given
Li,
Intelligence
representations? Intelligence
theory
representations? if
theory R.,
Intelligence so, in
theoryand
Intelligence
for Intelligence
Intelligence
for can tactical
Intelligence
to to Figure
Chellappa,
Conference.
prominent Intelligence
Intelligence
model Conference.
Conference.
That
prominent
we
model
to Intelligence
IntelligenceThat
detect
model analysis.
R.Conference
Intelligence
Intelligence
Conference.
Conference.
Conference.
Conference.
infor-
is,
infor-
5(a). do2010.
Conference
televisio
is,
infor- do Confer
Confere
televisi
This
Gro
Conf
Conf
ty
=free kicks,
of 380ofgames the data feed is given in Figure 5(a). This type �
feed, consists and more than 760,000 Y ) of −data when they deviate a discrete from this random style variable
and quantify X that it? has a probabilitydistrib
y∈Y y∈Y y∈Y y∈Y y∈Y y∈Y y∈Y y∈Y y∈Y y∈Y y∈Y y∈Y
y∈Y y∈Y
H(X, teams p(x, Iny) How
have log[85], discriminative
p(x,
unique Shannon y) styles (4)
uses of areplay? probability our Using And representations? a
ifaby Spatio-Temporal
so, theory can to
we That
model
detect is,
Driving
infor- do F
movement infer from hand-annotated ball-action a data (see mation sources, teams havei.e., unique
the adata styles produced of play?
Li, Li,
R., by
Li, R.,And
andLi,
R., aLi, source
and R.,
and
Li, if
Chellappa,
R., Li, so,
and
Chellappa,
R.,
Li, Chellappa,
and
and R., Li,
and
R.,can
is
Li, Li,
Chellappa,
Chellappa,
Li, we
treated
Chellappa,
and
R.,
andR.,
Chellappa,
R. R.,
R., andanddetect
Chellappa,
2010.
Chellappa,
R.
R. and
and 2010.
Chellapp
2010.
R.Chellap
R.Chella
Chella
Group
R.
2010.
R.
2010
201
Gro
Gr
20
RRT
remanentlyobserved from par- For our work we used the 2010-2011 EPL season F24
when they deviate from this style and quantify it?
ismovement movement and actionand executed
action executed byof a agentsteam
by (e.g.
team pass
(e.g. passfrom from � � � � � � � � �
� � � � � � � � � � � � � � � � −
gents of (11).fixed
agents Weevents.
(11). refer
toWe one to
refer team
team.
Each toteamthis
Unless
team data
plays has currently
38 is games been
currently
corners,ways widely used
each, has used the
offsides, used,
for
which same
forthe there
the number
real-time
substitutions
corresponds are
real-time no systems
online online
and
chunk
(11). whichWe
visualizations eachrefer
visualizations
stoppages. H(X,H(X, H(X,H(X,YH(X,
possession
)Y
H(X, An
= Y
)H(X, Y=to
−Y=
)H(X, −
)H(X,
YH(X,
as
team
of
example
−
H(X,
)= H(X,
x∈X )Y=of =−
mation
Y)a−Y
H(X, )−
=
y∈Y Ynewspaper
YY−
)=
random
string,
mation
when
)p(x,
= ))− Y=p(x, = −
sources,
newspaper
when
mation
= )p(x,− =
y)− − tionpX
teams
they
p(x, is
y)
p(x,
sources,
−p(x,
log y)
o,
they
log
variable. currently
deviate
p(x, sources,
p(x,
logy) y) entities
p(x,
into
y)
i.e.,
have
p(x, p(x,
logp(x,
y) y)
log
p(x,
i.e.,
p(x,
p(x,
y)
the
entities
=(p1,...,pn)isthendefinedas:
deviate p(x,
log
y) y)unique
from
p(x,
y) log i.e.,
y)
The y)
log
set
data
the
p(x,
y) y)
y)(e.g.
used
from(4)
logy) this
log
p(x, p(x,
log the
data
y)log
of
produced
(e.g.
styles
(4) (4)
p(x, this
p(x,
informationy) style
p(x,data
y) ESPN,
p(x,
play-segments
for
(4)
produced
(4) (4)
y) of
style
Using
y) ESPN,
y)(4) and the
produced
Li, y) play?
(4)
Using (4)
Using
by
and
C.;
The
(4)
quantify
aUsing real-time
(4) (4) And
H(X)
The
(4)
source
quantify
Spatio-Temporal
Using aManya,
aUsing
content,
aGuardian).
Spatio-Temporal
Spatio-Temporal
by Using
source
aUsing aif
Guardian).
it?
=
aso,
Spatio-Temporal
is
it?
aSpatio-Temporal
source
F.;
treated
aUsing
Spatio-Temporal
UsingUsing
is
aonline
can
Spatio-Temporal
Using
namely
treated
we
aaDriving
Mohamedou,
p(x)
Even
Spatio-Tem
Spatio-Tempora
Spatio-Tempora
Even
visua
detect
Spatio-T
Spatio-Tem
aisaSpatio-Te
Driving
Driving
treated
log
Drivi
Forc
Driv
th
Dr D
N
his next subsection).
agent A agentto agentA to B).
agent These B). behaviors
These behaviors are observed
are observed from par-
from Opta par- Mutual as feed, a
as random
Information a which random variable. consists variable. The of The information
380 information games content, andcontent, namely
more namely than
ssedcase refers tocoordinated
ball use this data orof data like this for automatic tactical S =
analysis. {s , . . . , s } of
this equal this
data data
In length,
has [85], has been Shannon been where widely widely uses N used, is
probability used, the there total
there are
theory are no
x∈X
no
to systems
modelsystems info w
ble servable segments
from segments
thewithofmatch, of
each coordinated
eachteam team playing al- events, each events, as
other behaviors,
well
the as team well
data as post-analysis
astwice
feed as is
B, post-analysis short,
(once given observable
home forin for prominent
Figure segments
prominent 5(a). television
This of coordinated
0
television type
N
and x∈X −1
and
(Shannons)of
x∈X x∈X y∈Y Mutual
(Shannons)
x∈X
data
y∈Y
x∈X x∈X
y∈Y
(Shannons)
x∈X
y∈Y
as
y∈Y x∈Xy∈Y x∈X
aevents,
when
Information
entropy
y∈Y
x∈X x∈Xx∈X
x∈X
y∈Y
random
entropy
x∈X
y∈Y y∈Y y∈Ythey
y∈Y
y∈Y
entropy ofas
y∈Y deviate
ofay) � well
variable.
discrete
a of discrete a
from as The
discrete post-analysis
random this
randomLi,Exploiting style
information
Li,
random C.;
Li, C.; Li, C.;
variable and
Manya,
Li,
variable Li,
Manya,
C.; Manya,
C.;
variable
quantify
Cycle
Li,Li,
C.; for
Manya,
F.;
Manya,
Manya,
X content,
Li,
C.;
X Li,
C.; prominent
Manya,
Li,
F.;Li,
that
thatX
it?
Structures
C.;
Mohamedou,
F.; C.;
Manya,
Manya,
� C.;
Mohamedou,
C.;
F.;
Mohamedou,
F.;
Manya,
F.;
has
has
that Manya,namely
Manya,
Mohamedo
F.;
Manya,
Mohamed
Mohame
hasF.;
F.;
MohaminMohtel
N.;
Moh
F.; M
F.;FF N
ball-action data (see A team plan tial can tial
spatio-temporal bespatio-temporal
movement defined as
and anaction
tracings, ordered
tracings,which
executed sequence
which
in this by incase a ofteam
this
number team
case
refers refers
(e.g.of to ball
play-segments
pass to 760,000 ball �
from
�
use events.
Mutual use
for
Mutual
this mation
H(X) this
a Information
data Each
possession, data
Information
p(x, =sources,
or − team
or
data data �
i.e.,
p(x) like plays
and the
likelog this data
M
Exploiting
p(x) this 38
tional for
Exploiting is produced
Exploiting games
for the
Exploitingautomatic
Conference
Exploiting
Exploiting automatic
Exploiting
Cycle to-
Exploiting
Exploiting
Cycle by
Cycle each,
(2) a
Exploiting
Exploiting onsource
tactical
Exploiting
Structures
Cycle
Exploiting
Cycle
Cycle which
tactical
Theory
Structures
Cycle
Structures
Cycle
Cycle
Structures is
StructuresCycle
Structure
Cycle treat
ana
and
Cycle
Structur
in Cycle
Struc
Struc
MAX
in
in an S MpM
i
by
cuted a team
by a (e.g.
team
ber of agents (11). We refer to behaviorsand pass
(e.g.
once from
pass
away). from For
The
team movement ournewspaper work names
newspaper ismovement we entities
and
currently used ranking
entities the
(e.g.
used 2010-2011
for
(e.g. ESPN, the
ESPN,
forThese the The
2010- The EPLGuardian). season
Guardian). I(X; F24
Even Even
Y ) though
= though − a probabilitydistributionpX
a probabilitydistributionpX
a (Shannons)
probabilitydistributionpX
p(x, y) newspaper log entropy entities of a
=(p1,...,pn)isthendefinedas:
(5) discrete
=(p1,...,pn)isthendefinedas: (e.g.
=(p1,...,pn)isthendefinedas: ESPN, random H(Y The
variable
) = −Guardian).
(2) X that
p(y) has
log E
by areal-time online visualizations of H(X) = − � p(x) � log p(x)
infer from hand-annotated ball-action � � � � � � � � �
� � � � � Mutual
�
as � � � � � �
a each � � �
random � Information variable. The information content, name
)data )(see
describing infer aIn recipe
from used
hand-annotated team total
ball-action achievenumber data (see
of corresponds with For ap(x,
our p(x, p(x, y) work team
p(x,
p(x,
y) p(x,
y)over p(x, y)
y) p(x,
y)
we playing
p(x, y) p(x, p(x,
y) p(x,
usedy) p(x, tional
y) y) bility
y) tional
y)
each
tional
the Testing.
Conference
tional tional Conference
2010-2011 other
Conference
tional tional
Conference
tional Conference
Conference team
tional
tional
Conference
on tional
tional
Conference
Conference
on
Theory
on EPL twice
Conference
Conferenc
Conference
Theory
on
Theory
Conference
on on Theory
onand
Theory
Theor
on
Theo
on
seaso and
and
Ap TTo
)play-segments for team aof match. Given
p(x)p(y)
par-Opta feed, which agent A to agent B). behaviors are observed from par-
)Ywhere For 0 a=p(x, log our
probabilitydistributionpX
this = 0 datawork and H(X) H(X) x∈X
has
the we =log base
(5) been − =used
(5) (5) − (5) (5)widely
(5) the
p(x) the
=(p1,...,pn)isthendefinedas:
(5) p(x)
(5)logarithm
(5)
log 2010-2011
(5) used,
log
(5) (5)
p(x) (5) p(x) there
determines EPL are (2) no season
(2) sys
viors are observed 2011 EPL
from data is given in Table hasconsists
1. our ofas 380
approach, games to and more than
I(X; I(X; I(X; I(X;
Y I(X;
Y
)I(X; Y I(X;
Y
= )I(X; Y= Y
I(X;)= −
I(X; Y=−
I(X; Y= −
)I(X;= Y − YY−
)= )y∈Y
= )−)Ywhere =− )p(x,
= = = − −
where
p(x, − −
y) − 0 p(x, y)−
p(x,
log log p(x,
y) 0 log p(x, loglog
y) y) =p(x,
y) p(x,
log0 y) log
p(x,= p(x,
p(x,
andy)log 0y) p(x,logand
y) y)
log
they) y)log the
log base log base of the �of thelogarithm logarithm determines determines
observable behaviors are
segments observed of from
coordinated par- a goal this (?). data
this Aevents, data
next
team hasbeen
subsection).as
performing been
well widely widely used,
post-analysis
a knowgroup used, there
ofwhichthere
these areare
for nonosystems
prominent
plans asystems
to television which which
x∈X
and Opta (Shannons) p(x)p(y)
feed, p(x)p(y)
p(x)p(y) entropy
p(x)p(y)
p(x)p(y)
which
�
p(x)p(y) p(x)p(y)
x∈X
p(x)p(y) p(x)p(y) of
p(x)p(y)
p(x)p(y)
consists Tp(x)p(y)
p(x)p(y) a− ,discrete
bility Li, bility bility Testing.
R.; of bility bility random
Testing.
Testing.bility
Chellappa,
380 Testing.
bility
bility
Testing.
Testing. Testing.
games variable
bility
bility bility
Testing.
bility
Testing.
y∈Y
Testing.
R.;Testing.and Testing.
Testing.
and XZhou, that
(2)S.
mor h
analyze the tactics of a next teamsubsection). we tial are spatio-temporal
required to tracings, the inthat
thistacticalcase team’s
refers = possession (once tox∈X
H(X) the ball
x∈X x∈X unit,
the
y∈Y home
+ x∈X Opta
y∈Y
x∈X the
H(Yunit,
x∈X
y∈Y e.g.
x∈X
y∈Y y∈Y string
unit,
x∈Xy∈Y )and e.g.
x∈X where
−feed,
aifuse
y∈Y
x∈X x∈X
H(Y x∈X
x∈X H(X,
y∈Y base
e.g. if
x∈X
y∈Y once y∈Y
base
this
y∈Y )is
y∈Y0if
y∈Y = log
which
Y2base
y∈Y of )the
− away).
2data =length
the 0the
H(X)
2measure and
consists
�
measure
orp(y) The
measure
data
x∈X the
=log x∈X isbase
1 team
is
like
p(y) inthe of
in isbits,of p(x)
bits, in
this 380 names
result-
the bits,if log
if for logarithm
its games
its p(x)
if the and
its
the
automatic
(3) � natural
the
naturalranking
determines
andnatural
� more
tactic
gs, dichsequence
xecuted which
in this incase
by aof
this teamteam(e.g.
case
refers refers
to ball to760,000
pass ball from use events. use
this this
data
newspaper Each
A data
planor team
or
data can data
entities like
be plays like this
defined(e.g. 38
this for as games
for
ESPN, automatic
an automatic
ordered each, tacticalwhich
sequence analysis.
analysis.
of team probabilitydistributionpX H(Y ))Y = − p(y) � Li, log Modal Li, R.; =(p1,...,pn)isthendefinedas:
Li, p(y) R.;Li,
Chellappa,
R.; Li, Densities
Chellappa,
R.;
Li, Chellappa,
R.; Li,Li,
R.;
Chellappa,
Chellappa,
Chellappa,
R.;
Li,
R.;Li, on
Chellappa,
R.;Li,Li,
Chellappa,
R.; (3) Discriminativ
Chellappa,
R.;
and
R.;R.;
R.; Chellappa,
Chellappa
Chellappa,
and
R.;
and
Zhou,
Chellappa
R.;R.; R.;
and
Zhou,
Zhou,
and
and
R.;
R.;
and
S.ZhoZh anZa2R SS
achieve aAmajor goal (e.g. winning ahand-annotated
match), at The can Guardian).
beball-action
saidof team Even though the760,000 unit, e.g. events. if base 2Each the measure team plays
is in bits, 38 if games
its the each,
natural
ing number of play-segments
fornumber the 2010-2011 e then for its each
EPL in in nats, possession
data is � given is in therefore: Table 1.
=etc. The term log 1/pi indicates
plan canhas be defined as of an H(X, Y ) = − p(x, y)
itordered sequence 760,000number − − −ewhere then
events. − − − −
0its
− − in− Each
− − nats,
− etc.
team The plays term log
38 1/pi
games indicates each, Dw
x∈X
position of the ball and which movement
team infer
possession from = =
data = H(X) = H(X) = H(X)
=
(see H(X) =number
+ H(X) = H(X) +H(Y= H(X)
+
= =
H(Y = H(X)+
H(Y +
H(X)=
) + H(X)
H(Y e then
H(X)
H(Y
H(X)
) +H(Y ) H(X)
H(X, + H(Y +
H(X,
) ) H(Y
For
H(X,
+ ) +
H(Y + its Y H(Y
) +
H(Y
H(X, H(Y)
H(X,) Y
our
H(Y H(X,
)
log
H(Y)nats,
)
H(X,
y∈Y Y
) H(X,Y
work
) )H(X,) YH(X,
)− H(X,
0etc.
)H(X,
Y
=and Y )
− ) Y
we The
Y Y )
the
)Y
Modal)
used
)
Manifold term
Modal
p(y) Modal
base Modal Modal
log log
the
Densities
of Modal
Densitiesfor
Densities
the
p(y) 1/pi
Modal Group
2010-2011
Modal Densities
Densities
Densities
Modal
logarithm indicates
Modal
on Modal
Densities
Modal Activity
Densities
on
Densities Densities
Discriminative
on Densities
Discriminati
on
Discriminat
on
(3) EPL
Densities
onDensities
on Recog
Discrimi
Discrim
(3)
determin Discri
on
on
Discr Do o
yse otatedabehaviors
d-annotated team to are
ball-action achieve
ball-action data
observed data
(seefrom (see
corresponds
topar- beevery employing For
behaviors withFor our
thisbehaviors each
our work
tactics.
data
describing workteam
has we
describing
However, we
been playing
used
a recipe used athe
as
widely each
the
recipe
soccer2010-2011
used used, other
2010-2011
used
by a by
is low-scoring,
there team
N EPL
team a=areEPL
team
(Ttwice
to no −season
1season tosystems
achieve T )achieve + 1. F24 F24
the
To Ifwhich amount
the
analyze the the
amount
corresponds
number
possession corresponds
amount of the
the uncertainty
ofunit, e
of uncertainty
uncertainty
tactics
H(Y
then
with e.g. string with
ofits) =
ifassociated
each a
base
iny∈Y
iseach
associated
team, nats, associated
smaller
2team
=the
p(y)etc.
team
we
Manifold
�
with
measure with
Manifold
log The
require
Manifold
playing in playing
with
Manifold
p(y)
the the
Manifold
Manifold term
dura-
for Manifoldiscorresponding
the
corresponding
for
Groupin to
Manifold
for
Manifold
each
log
Group
for each
corresponding
know
bits,
x∈X
Manifold
Manifold
Groupfor
for 1/pi
Manifold
Manifold
Activity
Group
for Group
Group
ifforother
y∈Y indicates
where
Activity
for
Group
otherActivity
its for
Group
for
Activity
Group
Activity
Recognit
the for
for team
Activity
team Grou
Activit
Group
RecoGro
Reco
Gro
naturAct
AcR
every time step (i.e. second). next subsection).
To do this, we infer the Summary and Future the Opta
amount Work feed, ofalso uncertainty
H(Y which )y∈Y y∈Y − consists
associated
Molineaux, p(y)
with of
log M. 380
p(y)
the
2008. games Working
corresponding
(3)
and Spe
up ngs,ofwhich theseinplans to refers
this location
case (once thehome
tocontinuous
ball Opta a goal Opta and feed,
and a(?). once
goal
feed,
complex which
Aplan away).
(?). which
team Aconsists
due team The
consists
to
orperforming the team
performing
of this
various of names
a380 380 agamesgroup
multi-agent and
games
oftion ranking
of
and and
than these
in- more Tmore ,plans we the than thanoutcome.
discard to ball outcome. is outcome.
and it. It(once can It
To
who
number can also
It home
representcan also
has be
eWork be viewed
and
possession
then beits
viewed viewed
once
each in asas nats, the away).the asy∈Y
play-segment
ofMolineaux, amount
Interface. theamount
Molineaux,
it Molineaux,
etc. at amount
Molineaux,The
Molineaux,
Molineaux,
every
The Molineaux,
M.of of
Technical team
Molineaux, information
Molineaux,
term M. of
information
2008.
M.time information
Molineaux,
Molineaux,
2008.
M. names
Molineaux,
Molineaux,
2008.
M.M.
log Working
M.
2008.
2008.
2008.
report,
step M.
Working
1/piWorking
M.
2008. and
M.
Working
2008.
M.
2008.
Workin
Knexus
(i.e. M.
Specifi
Work
M.
indicatWork
200200
Sp
Sp
20
W 2W r
ball from data use
feed.
achieve
this
AWe data
a
describe
majorcan data
begoal
our like
defined method
(e.g. asgroup
winning
for
an automatic
ordered
a
these
match), sequence plans
tactical
Summary
can Summary
Summary
Summary
be
toSummary
Summary of said
analysis.
Summary team
Summary
gainedand Summary
Summary
and Summary
Summary
gained and (once
Summary
Future
and andFuture
gained
by and Future and
outcome.
by Future
observing and
Future home
Future
and Work
760,000
by and
observing and
Future and
Work Future and
Work Future
observing Future
It and
Future
Work Futurecan thatFuture
Work that
once
Work
events.also Work that
outcome. Work Work
outcome. beWork
Work away).
viewed
outcome. Each Thus,
Interface. Thus, as
Interface.
The
team
Interface. Thus, the entropy
Interface. entropy
Interface.
Interface.
team
amount
Technical
Interface. plays
entropy
Interface.
Technical
Technical
Interface. isis
names
of
Interface.
Interface.
Technical 38
merely
merely
Interface.
Technical
Technical
Interface.
report,
Technical games
isinformation
merely
report,
Technical
report,
Technical
and
Technica
report,
Knexus Techni
Technical
report,
report,
Technic
Knexu
report,
Knexu
ra
rep
rep
Kn Re
K
as
match), an ordered
can be sequence
said of team
forteractions,
tthe 2010-2011 760,000 EPL events. data Each
islocation
given team plays 38 To
games Do{s each, which for the 2010-2011 EPL data is given in Table 1.
ain Table 1.
ordered sequence of team ball labeling and segmenting the game into = }, the quantized the amount ball of uncertainty
position is Opta.associated tabulated 2012. with the correspondi
ToaToDo series
s , . . . , s www.optasports.
and-annotated ball-action 2. data 0(see 760,000
achieve events.
aFor major our Each
goal
work (e.g. team
we winning used plays 1 . a38
tthe match),
2010-2011games can 0each, be
EPL said which
T every
season−1
asecond). for
F24 theaverage Tobyaverage
2010-2011 do this, ofEPL we thatinfer data isOpta.the given ball in location Table 1. from
tobehaviors describing recipe used atby aDo team to achieve gained observing outcome. Thus, entropy iswww.opta
merely
via Figure At the is passed to the at corresponds with
cipe used be employing tactics. However, aseach soccer ToTo
Do To Do
isTo DoTo Do
ToDo To To
low-scoring, To
Do
To Do DoTo Do
Do Doa statistical statistical a statistical outcome.
To average analyze ofItof uncertainty
can uncertainty
the uncertainty
also tactics beeach or Opta.
viewed orRiley,
of information.
Opta. aorteam
information. 2012.
team, Opta.information.
Opta.
as 2012. 2012.
Opta.
the playing Opta.
2012.
Opta.
we 2012.
2012.
www.optasports.co
amount 2012.
Opta.Opta.
www.optasports.
www.optasports
require Opta.
Opta.
2012.
2012. each
www.optaspor 2012.
2012.
www.optaspo
www.optasp 2012.
2012.
www.optasp
of www.opt
informatiother
to2002. www.
www.o
know www
www
a by a team toaction
achieve P., and Veloso, M.
occer
used is low-scoring,
by team The to
next achieve of plans
To
labeled analyze to isiscorresponds
corresponds beat extremely
thetOpta4atactics
employing goal
where with
feed, awith
difficult.
(?). of
tactics.
eacha
player A
which
each
team, Hence
team takes team
However,
team we
consists require
playing
performingon playing
recognizing anas of to
soccer
each
a
380 team
know
group is each
other
games
other
tac-
where
low-scoring,
of
time
team
these team
and
step. the
twice
plansmore
twice dataAn
to than
example
feed. To a statistical analyze The
gained(once
ofmethod the
by home average
the observing
description
tactics
� of
and of doinguncertainty
� once of
that a
Riley, this is
team,
outcome. Thus,
tic away).
Riley, Riley, Opponent
given
P.,Riley, is
or
Riley, P.,
andwe
P.,
Riley, best
information. The
and
Riley,
and
P.,
Riley,
in
require
Veloso,
P.,
P., described
and
Riley,
MovementP., team
Riley,
Veloso, Riley,
Veloso,
and
andRiley,
P.,
and
P.,Veloso,
entropyM.
Veloso,to
Veloso,
and
and
P.,P., names
Veloso,
M. P.,
P.,
M. know
and
2002.
and is
Models. and
Veloso,
Veloso,
and
M. in
2002.
M.
2002.
Velos
M.Velo
M.
mere Ve
Ve
200
Re20 MawM2
rming a group of in- these plans to (once continuous
home and and once complex away). due The to the team various names
Figure multi-agent
3. and ranking in- the ball is and who � has possession of it at every time st
rious d. awinning
asgroupmulti-agent
an of opposition
ordered these plansplayer.
sequence thetics
toof Asusing
ball
team iscontinuous
(once
nothing and thewho home MAPR
occurred
760,000 achieve
teractions, has and
and framework
possession
complex
a once
between
events. major
labeling away). we
due
the
goal
Each of
and
described
it
to
time
(e.g.at
The
teamthe
segmenting every1various
team
twinning
plays
previously
time
thenamesa
38 step
multi-agent
match),
game games
isand
(i.e.
into can a in-
ranking
each, Figure
series be said whichevery second). 5(b).the ball a for is
statistical H(X) and
the H(X) who
H(X)
2010-2011 = average =has
To
= p(x)
do possession
of p(x) �
EPL p(x)
tic
log
uncertainty S.;
log tic
Opponent
ticp(x) log
and
data Opponent
p(x) Opponent
tic p(x)of
ticTadorokoro,
Opponent
tic
isOpponent
Opponent
or it
ticMovement
Opponent
tic
given at Opponent
Movement
tic
information. ticevery
Opponent
Movement
tictic Opponent
Opponent
in (1)
Opponent
Movement
Movement
Opponent
Movement
S., time
Movement
Models.
(1) (1)Move
Movemen
eds.,
Table Movemen
Models.
Models. Mov
Mode
RoboC ste
Mo
Mod
Mo
1. In
MoM
a match), caninfer beevery said for the 2010-2011 From 1.our set of play-segments for a given = this,
team, Swe or infer Sp(x) ,and the ball locatio (1)
weEPL data isand given in Table
impossible without these S.;World S.;
and S.; and S.;
and Tadorokoro,
S.; and
S.;
Tadorokoro,
Tadorokoro,
and S.;and
S.; Tadorokoro,
Tadorokoro,
Tadorokoro,
S.;
and S.; S.;
Tadorokoro,
S.; S.,
and and and
Tadorokoro,
Tadorokoro,
and
S.,
eds.,
S., Tadorokoro
Tadoroko
Tadorokoro,
eds.,
Tadoroko
S.,
eds.,
S.,
RoboCu
S.,eds.,
S.,eds.,
Robo
Robo
eds.
S.,
eds
S.,R
ening game into
a used
match), aby
andseries
cant4 , we be said the second).
ball for teractions,
was the To
dribbled toplans
2010-2011 be labeling
do this,
in
employing alabelled
EPL and
straight infer plans.
segmenting
tactics.
data lineisthe However,
given ball theinlocation
gameas
Table soccerinto1.from a series
is low-scoring, every second). To feed. analyze ToThe
H(X) x∈X do x∈Xx∈X
the this, tactics we
p(x) A
infer Cuplog V. B
the Springer
ball Verlag.
location
recipe
owever, as soccer a isteam to
low-scoring, achieve of
corresponds is extremely with each difficult. team Henceplaying recognizing
we each
can buildother team a team tac- twice
distribution theto data characterize method
the World
expected � World ofof
World CupWorld doing
WorldaCup Cup team,
World V.be- World
Cup
Springer
World
V. Cup
V.
Cup this World we
Springer
Cup World
Springer
V. is
World
World
V.
Cup
V.
Cup require
Springer
V. best
Cup
Springer
Springer
Verlag.
Cup
Springer
V.V.
CupCup
Verlag.V.V.
Verlag.
Springer descr
Springer V. to
Sprin
Verlag.
V.
Verlag
Sprin
VerlaSpk
Ver
Spr V
ecognizing
er, as soccer team atac-
at low-scoring,
is uniform velocity the data To between overcome
ofToplans
feed. ToThe analyze
isthis
these
tics extremely
method
continuous
using issue,
two the
the tactics
we
locations.
of
and
MAPR doingquantize
difficult.
complex ofthis aHence
We
framework team,
ais
due domatch best we
torecognizing
we require
into
described
thedescribedvarious equal to
team know
ineach
multi-agent
previously tac- where Measuringis in- the Analysis data the Team
feed. Using ball The
isEntropy andTactics methodH(X) who Maps
x∈X
=via
has ofGiven doing
possession Entropyp(x) log thisp(x) isMaps
of itbest at every describ tim (
due forming to the a group
various of these
multi-agent plans in-
temporal to t5 the chunks
analyze
(once which homethe tactics
we and
theuse
of
onceto
a team,
away).
describe
we The
team
require team
havior
behavior.
to namesknow
in where
and location. Analysis
ranking Analysis WeUsing
Figure Using do Entropy
5(b). this Entropy as Maps
follows: Maps the
scribed previously the same is thing between tics andusing ball t6 ,the isbefore and MAPR who has
framework
ball possession
is passed we described ofplans.it at everypreviously time step isa series (i.e. Figure Analysis 5(b). Using Entropy doMaps
isteractions,
impossible labeling
without these and segmenting
labelled the game into x∈X
thewinning
various Figure 5(b). Determining we every second). Parameters theTo this, of we infer the ball loc
egmenting .g. the amulti-agent
match),
game
from the into cana
location
in- beAs
series atsaid t6 the
these to impossible
ball
tsegments
every
7 . Using for and
the
second). do
without
the who
2010-2011
not same has
describe
To
these
possession
do
procedure, EPL
this,
labelleda data
methodwe
of
plans.
we inferis ofitgiven
at every
achieving
the in
observations
ball
time
Table a
location
step
1. ofPlay-Segment (i.e.
aDetermining
from
team, Determining
Determining
determine Representation
Parameters Parameters
Parameters
subset play-
nting . the game into a series every of To plans overcome is extremely this issue, difficult. we Hence
quantize recognizing
a match into team equal tac- How Measuring
the
discriminative
Analysis data feed.
Using are Team The Entropy our method Tactics
representations? Maps of doingvia ThatThat Entropy
this
is, is, is do bestMd
thesecond). notTo do this, we ofinfer the ball location fromin
However, as soccer is low-scoring, Todofeed. analyze the tactics a team, wesegments S , which How How originated discriminative discriminative in are
quantized are our our representations?
representations?
area d. From That do
do
cult. Hence recognizing team thetac- specific goal, we
data call
The them
method plans, of but
doing play-segments.
athis isrequire
best intodescribed to know Given where isthe partial spatio-temporal tracings of the team (i.e.
d
can estimate ball position To overcome
and
tics
temporal team
using this
chunks issue,
possession
the MAPR which we for
we quantize
framework the
use to we match
describe described team equal
previously
behavior. teams Measuring How have Figure discriminative unique 5(b). Team
styles are of Tactics play?
our And
representations? via
if canEntropy
so, can we we That detect is,Ma do
eework
Hence a match
due towethedescribedinto
recognizing variousequal
remaining team times Measuring
tac-
multi-agent
previously -Weisasin- use thethe these
temporal data
stoppages the Team
feed.ball
play-segments
chunks are The Tactics
issegments
and which
tagged method
who
to form inhas
we via
of Entropy
doing
a possession
use
the libraryto
data describe this
or of
playbookisteam
the Maps
best
it at every
ofdescribed
play-segment
behavior. time ballwhen
teams
stepin
vectors
movement),
teams
(i.e.
have
when
havewithin
Play-Segment unique
Determining
they
unique
we deviate
S styles
d
chunk
, styles we
from
of ofplay?
Parameters
keep Representation
this this
play? a
signal
style
And countAnd
and
if ifso,
upquantify
so,
of
into
canthe we
discrete
detect
detect
it?can weseg-
Figure As 5(b).
these
impossible without do not
these describe
labelled aplans.
method of achieving a when they they teams deviate deviate have from unique
from this this styles
style style ofand
and play? quantifyquantify And ifit?so,
it? detect
kcribe segmenting
abelled team
we described plans. behavior.
thefeed.previously
game It is intoworth isa noting
series
play-segments, Figure
here that 5(b).
every all=second). To do this, awe where infer
locations of where thePlay-Segment
isaachieving
the the ball location aments from - called
ball when
travels
How play-segments.
they discriminative
deviate
or occupies Representation from this are Westyle
during ourrepresent representations?
and quantify
these it? That is,
theseofplay-
Play-Segment As these specific Todata
Psegments
Representation {p1is
goal,
overcome , do normalized
pwe2 ,not pdo 3 , describe
this .not
. issue,
. ,callpm onto},
them
we method plans,
quantize mof but
play-segments. play-segments.
match into As equal this Mutual
representation Given Measuring
Information the partial will spatio-temporal yield Team aofvery Tactics high- tracings via Entropy the tea
ethod
d
ficult. plans.
we quantizeof achieving
Hence recognizing
a a
a match intoteam
field of size 100
equal number
×tac- specific
100, of unique
with the Weallgoal, play-segments
positions
use
data wethese
feed. do given
play-segments
notThe within
call for
methodthem teamsthe to playbook.
form
plans,
of doing a
but library Using
play-segments.
this isor playbook
best described segments Mutual
of Mutual Given inas Information
ball an teams
Information
the occupancy
movement), partial
have unique map
spatio-temporal we styles which
� chunk play?
describes
this signal
tracings
And ifwhich
upso, can
of into areas
the
we dete
discre
team
s,mework but play-segments. Given the partial Measuring temporal chunks
spatio-temporal Team which Tactics
tracings
we usevia
of
to describe
the Entropy
dimensional
team
teamMaps
(i.e. feature behavior. vectorMutual (D when
Play-Segment
2
),H(X) Information
they
we H(X) woulddeviate �
=−
� from �
prefer
Representation
this p(x) the style log dimen- and
p(x)quantify it? (2)
quantize e use to we a match attacking
described
describe into equal
left
teampreviously to
behavior. is We this
right. playbook Measuring use we
Figure As thesecan:
play-segments,
these play-segments
5(b). Team Tactics
segments P = dotonot
{p 1form
, p , pvia
2describea3 library
, . . Entropy
. , p
asionality
m },
method where
or playbook of
to Maps m
achieving is of
beofD for visualization the the field
aball movement), the
ments ball H(X) - was called
purposes.
= during
= − we − play-segments.chunk that
p(x)
To
p(x)
achieve
play-segment.
log this log p(x) p(x) signal this, We up represent into (2)
(2) discrete these
(2)
edescribe brary
to
labelled or playbook
describe aplans. team of
method behavior.
of achieving ball amovement), play-segments, Play-Segment number we chunk
specific ofP unique
goal, =Representation
this we signal
{pplay-segments
1do , p2not , p3call ,up into
. . .them , within
pm }, discrete
plans,
we can butseg-
the playbook.
where m is either
play-segments.
use the Using ments
the mean,Mutual
segments Given -median, called as the Information anH(X)
occupancy
play-segments.
partial
mode,
x∈X
�
x∈X
�
=− x∈X
� mapWe
spatio-temporal
total
p(x) log p(x)
count
which
or represent describes
tracingsthese ofwhic th
a where m4. OCCUPANCY
isaachieving
the aments Play-Segment
MAPS this playbook Representation
we can: of the field H(Y the ) ball = − was
x∈X p(y)
during � log that p(y) play-segment. (3)
ibe m }, method of equal- number called play-segments.
of unique play-segments We represent within thethese playbook. play- Using segments as an )occupancy map which describes (3)which
(3)
all them plans, but play-segments. GivenWe the use partial
these spatio-temporal
play-segments to tracings
form a library of the or team
playbook (i.e. of ball H(Y movement), = − p(S we p(y) � chunk log p(y) this signal up into (3)d(
H(Y )= −H(X) p(y) log p(y)
ue, we quantize match into Measuring Team Tactics via
even Entropy
an entropy Mapsmeasure to describe each
H(Y )y∈Y =− y∈Y = d ),− which p(y) log p(y) p(x) will log p(x)
m
ts theto playbook.
plans,
formtobut a describe
library Using
play-segments.
orteamplaybook segments
ofball Given thisasball an playbooktheoccupancy partial
play-segments,
movement), we can: map wePwhich
spatio-temporal chunk = {pinfor- describes
1 , p2 tracings
this p3 , . .which
,signal , pof
up
.yield into
}, areas
the where teamm(i.e.
discrete is
seg- or of
the the behavioral field
mentsthe- ball called y∈Y was during
play-segments. that play-segment. theWe represent
we use Given we have behavior. tracking and team possession m an occupancy team map. Vectorizing y∈Y
x∈X

ot orm , pa3 ,library

p2describe . . . ,apm where
or},playbook
mation,
method of mcan ofisofthe
weachieving the field
partition aballthe ments
movement),
the Play-Segment
ballnumber
ball - was called
tracking during
of we unique chunk
data Representation
that
play-segments. play-segment.
play-segments
into thispos- signal Wewithin uprepresentinto
occupancy the discreteplaybook. these map,seg- Using
play-gives us oursegments D-dimensional as an occupancy spatiotemporal map which describes w
ments
,
3call them . . . , within
p }, the
where
session playbook.
m plans, but play-segments. m is
strings theUsing(i.e. ments
continuous segments - this
called
movementplaybook
as an occupancy
play-segments. we
of can:
the ballmap We
for which describes
represent feature these
vectorwhich play-
x. areas The mean ofoccupancy the field the maps ball using was during entropy that play-segmen
Given the partialwhere spatio-temporal OAwhich
tracings of the team (i.e.
sents within the
to form a {o
a single team
playbook.
library
A Using without segments
or Aplaybook ofB = {oB
turnover of the orfield
as
ball movement),
stoppage),
an B
the
occupancy ball was during
map
we chunk this signal
=that play-segment.
describes to describe which
up intoofdiscrete
each areas area is shown in Figure 4 and can give a
seg-patterns. As can be seen the top
0 , . . . , oI−1 } and O 0 , . . . , oJ−1 } refer to the pos- indication redundant
1 , p2 , p3 , . . . ,session pm }, where stringsm is theof with
associated
the field ments
each team
the- ball calledand I −1
wasplay-segments.during and J −1
that play-segment. We teams represent have higher theseentropy play- over most of the field compared
egments within arethe the playbook.
number Using of possessions. segments We then asquantizean occupancy the field map which to the describes lower teams which - this areas gives an indication that these teams
into D bins where D = l × wofequal the field size areas the ball andwas vectorize during thatutilize play-segment. more options (i.e. less predictable) which is intuitive
the field via the columns. As the possession strings vary in as these teams normally have more skilled players. As the
length, we apply a sliding window of length T to quantize or frequency counts incorporate temporal information (more

1369
 7

66

(1) Manchester United (2) Chelsea (3) Manchester City (4) Arsenal (5) Tottenham

55

44

(6) Liverpool (7) Everton (8) Fulham (9) Aston Villa (10) Sunderland

33

22

(11) West Bromwich Albion (12) Newcastle (13) Stoke City (14) Bolton (15) Blackburn

11

0
0

(16) Wigan (17) Wolverhamption (18) Birmingham City (19) Blackpool (20) West Ham United

Figure 4: The mean entropy maps for each of the twenty English Premier League teams characterizing their
ball movement patterns. The maps have been normalized for teams attacking from left to right. The bright
red refers to high entropy scores (i.e. high variability) and the blue areas refer to low entropy scores (i.e very
predictable behavior).

counts, means the ball is moving quicker), we used the to- % Teams Correctly Identified
tal count of entires as the entropy measure normalizes this 50
information. 40
30
20
5. STRATEGY ANALYSIS 10
0
5.1 Discriminating Team Behavior Statistics Occ Maps Combined
Evaluating team strategy is a very difficult task. The ma-
jor hurdle to overcome is the absence of strategy labels. But Figure 5: The identification rate for correctly iden-
given we know the team identity, and assuming that teams tifying home performances using event-labeled data
exhibit similar behaviors over time, we can treat the task as (i.e. no location information), occupancy maps (i.e.
an identification problem. We can do this by answering using no event information, just location), and a combina-
only ball movement information, can we accurately identify tion of the two.
the most likely team?
We model team behavior using a codebook of past per-
formances. If a team’s behavior is consistent, then previous knowledge of where teams operate could boost the discrim-
matches will be a good predictor of future performances. inating power.
For the experiments, we used 380 games of the season and To conduct the experiments, we compared our occupancy
used a leave-one-match-out cross validation strategy to max- representation, to twenty-three match statistics currently
imize training and testing data. Before we investigate the used in analysis (e.g. passes, shots, tackles, fouls, aerials,
difference between home and away performances, we first possession, time- in-play etc.). We also combined the two
need to obtain the best possible representation. To evalu- inputs by concatenating the feature vectors. For classifica-
ate this, we wanted to see how effective event-labeled data tion, we used a k-Nearest Neighbor approach (with k = 30)
was in discriminating between different teams, and if having and all experiments were conducted using D = 10 × 8

1370
 1 18
Exp Event-Labeled Occupancy Maps
HvH 19.26 38.79
HvA 16.09 30.08
Actual Team Ranking AvA 13.98 36.41
AvH 16.36 30.34

Table 2: The hit rate accuracy of experiments which

tested home (H) and away (A) models against home
and away matches (e.g. HvH refers to home model
tested on home matches and HvA refers to the home
model being tested on away matches).

0 do this, we simply subtracted the home occupancy maps

20 from the away maps and divided by the away occupancy.
1 Predicted Team Ranking 20 The difference maps for all twenty teams is given in Figure 7
(a) Home vs Home and it makes for compelling viewing. To make it easier to
quantify the difference in occupancy, we calculated the dif-
Figure 6: Confusion matrices of the team identifica- ference with respect to certain areas of the field. Specifically
tion experiments using the combined representation. we calculated the difference: for the whole field (W), the at-
tacking half (H), and the attacking-third (T) - these values
are listed below each difference map. As can be seen from
spatial areas (heuristically, we found those values of k and the difference maps, spatially, nearly all the teams (18 out
D gave the best performance). To reduce the dimensionality of 20) had more possession in the attacking half and prob-
but maintain class separability for the spatiotemporal rep- ably more telling is that 19 out of the 20 teams had more
resentation and the combined representation, we used linear possession in the attacking third. Seeing that the shoot-
discriminant analysis (LDA). We used the number of teams ing proficiency is essentially the same (10% vs 9%), we can
as classes (C = 20), which resulted in a C − 1 = 19 dimen- point to the observation that the more possession in the at-
sional input feature, y = WT x, where W can be found via tacking third leads to more chances, which in-turn leads to
solving more goals. A potential statistic to back this observation
up is that Chelsea - who were the only team to have less
WΣb W occupancy in the attacking third - had the smallest discrep-
)
arg max Tr( (1)
WΣw W W ancy between home and away shots (only 12, the next was
where Σb and Σw are the between-scatter and within-scatter Arsenal with 45).
matrices respectively. To test the various representations, With the absence of labels to compare against it is im-
only identity experiments on home performances were tested possible to say whether this was actually this case as other
(home and away comparisons are be done next). Results for factors may have contributed to the home advantage (i.e.
these experiments are shown in Figure 5. From this fig- referee’s [22], shooting chance quality, game context (i.e.
ure, it can be seen using spatiotemporal data greatly im- winning, losing, red-cards, key injury, derby matches etc.).
proves the discriminability between different teams (19.26% However, through the use of spatiotemporal data, we can
vs 38.79%), and fusing the two together boosts performances provide evidence of behavioral differences which can aid in
again (46.70%). This result makes sense as the location of the analysis of performance and decision making. This ap-
where teams play in addition to what they do should charac- proach can also be used to flag and predict individual team
terize their behavior. The confusion matrix of the combined performances, and in the next section we show methods in
representation shows that teams who play a similar style of- which these can be applied.
ten get confused with each other (e.g. the top 5 teams and
teams 13-15) and from viewing the entropy maps in Figure 4, 6. PRE/POST GAME ANALYSIS
we can see that these teams look similar. Given a coach or analyst is preparing for an upcoming
match, having a measure of how variable a team’s perfor-
5.2 Comparing Home vs Away Behavior mance is would be quite beneficial. For example, the coach
If a team plays in the same manner at home as they do or analyst may have viewed a previous match and formed a
away, the home model should be able to yield similar perfor- qualitative model based on their expert observation. How-
mance in identifying away matches as they do to the home ever, this model is only formed by a single observation and
matches. To test this theory, we used our home models to may be subject to over-fitting. Having an measure which
identify away performances and our away models to identify could indicate how variable a team’s performance is would
home performances. From the results (see Table 2), we can be quite useful. Given they have a feature representation of
see that there is a drop in the hit-rate of the occupancy map each of the previous performances of a team, our approach
representation – 8.69% for the home model tested on the could be a method of determining the performance variance.
away matches and 6.07% on the reverse case. Even though To do this, the distance in feature space between each of the
not excessive, this drop in performance suggests there is a past performances, y, and the mean, ŷ, can be calculated
change in the spatial behavior between home and away per- where the mean is
formances. M
To explore this aspect further, we visualized the difference 1 X
ŷ = yi (2)
in occupancy between the home and away performances. To M i=1

1371
 100

(1) Man United (2) Chelsea (3) Man City (4) Arsenal (5) Tottenham
W=3.8%, H=7.6%, T=10.7% W=1.1%, H=2.5%, T=-0.5% W=1.5%, H=3.1%, T=8.7% W=3.9, H=4.2%, T=9.5% W=0.1%, H=3.0%, T=8.9% 50

(6) Liverpool (7) Everton (8) Fulham (9) Aston Villa (10) Sunderland
W=1.8%, H=3.5%, T=9.5% W=-2.9%, H=-0.3%, T=5.6% W=-0.2%, H=1.6%, T=4.5% W=0.1%, H=0.6%, T=0.2% W=2.6%, H=7.1%, T=8.4% 0

(11) West Bromwich Albion (12) Newcastle (13) Stoke City (14) Bolton (15) Blackburn -50
W=0.6%, H=3.1%, T=9.0% W=5.2%, H=6.7%, T=0.5% W=-3.7%, H=2.1%, T=1.9% W=6.2%, H=5.4%, T=8.0% W=8.1%, H=3.6%, T=7.3%

-100
(16) Wigan (17) Wolverhamption (18) Birmingham City (19) Blackpool (20) West Ham United
W=8.1%, H=12.4%, T=8.1% W=-4.3%, H=0.4%, T=4.1% W=5.0%, H=8.3%, T=13.8% W=-1.0, H=3.4%, T=11.9% W=-2.2%, H=-0.7%, T=3.6%

Figure 7: The normalized difference maps between home and away performances for all the 20 EPL teams. In
all maps, teams are attacking from left-to-right and a positive value refers to a team having more occupancy
in home games, while a negative value refers to more occupancy in away games. Percentages underneath
each team give a value on this difference (Key: W is whole field, H refers to forward half and T refers to the
attacking third.)

120
home and away performance example, we can show the vari-
Distortion Variance

ation in performance for each team in the EPL by finding the

80
variance in distortion (Figure 9). As can be seen in this fig-
40
ure, each team’s home performance has quite a low variance
which gives an indication that when team’s play at home
0
they do not vary their approach too much. Conversely, it
0 5 10 15 20
(a) Home Performances seems that the away behavior is quite random so forecasting
120 away performance may be unreliable.
In terms of post-match analysis, a similar approach could
80 be used to see if a team’s performance was within the expec-
tation range (i.e. ±σ). A good example was Fulham’s away
40 performance against Manchester United. In this match, they
lost 2-0 and conceded both goals in the first half (12th and
0
0 5 10 15 20 32nd minute). As can be seen by comparing both occu-
(b) Away Performances pancy maps, in their match against Manchester United they
occupied a lot more possession in the middle of the field
Figure 8: Variance in distortion for (a) home and then normal. This highlights the importance of context, as
(b) away performances. after scoring two early goals Manchester United sat back
and allowed Fulham to have the majority of possession in
non-threatening regions (52% of overall possession) [6]. To
and where M is the number of previous performances. A counter this, we would have to normalize for match con-
distance measure such as the L2 norm could be used given text (i.e. score, strength of opposition etc.). However, this
that the input space has been scaled appropriately (as is the is a major problem as we would limit the amount of data we
case in our work), which generates the distance measure via would have to train our model. Future work will be focussed
distm = kym − ŷk2 (3) on clustering styles unsupervised to maximize the amount
of context dependent data.
where m refers to the game of interest. Returning to our

1372
 Distance from Mean 60

40

20
(a) (b)
0
0 5 10 15 20 Figure 10: Occupancy maps of the: (top) mean away
Opponent Number performance for Fulham, and (bottom) their perfor-
mance against Manchester United - in this match
Figure 9: Example of the distortion for each away they lost 2-0 and conceded early in the match.
performance of Fulham in the 2010-2011 season. In
match 16 they played Manchester United, and the
performance on the day was far different from their 8. REFERENCES
other away performances (they lost 2-0 on this oc- [1] S. Ali and M. Shah. Floor Fields for Tracking in High
casion). Density Crowd Scenes. In ECCV, 2008.
[2] N. Allen, J. Templon, P. McNally, L. Birnbaum, and
K. Hammond. StatsMonkey: A Data-Driven Sports
7. SUMMARY AND FUTURE WORK Narrative Writer. In AAAI Fall Symposium Series,
Most sports analytics approaches still only use event-labeled 2010.
statistics to drive analysis and decision-making despite the [3] BBC-Sports. Footballers may trial wearing microchips
influx of ball and player tracking data becoming available. to monitor health.
The reason why this new and rich source of information is www.bbc.co.uk/sport/0/football/21460038, 14 Feb
being neglected stems for the fact that it is continuous and 2013.
is extremely difficult to segment into categories which would [4] M. Beetz, N. von Hoyningen-Huene, B. Kirchlechner,
enable high-level analysis (e.g. team strategy labels). The S. Gedikli, F. Siles, M. Durus, and M. Lames.
emerging field of sports spatiotemporal analytics attempts ASPOGAMO: Automated Sports Game Analysis
to use spatiotemporal data such as ball and player track- Models. International Journal of Computer Science in
ing data to drive automatic team behavior/strategy analysis Sport, 8(1), 2009.
which would be extremely useful in all facets of the sports [5] P. Carr, Y. Sheikh, and I. Matthews. Monocular
industry (e.g. coaching, broadcasting, fantasy-games, video Object Detection using 3D Geometric Primitives. In
games, betting etc.). In this paper, we gave an overview ECCV, 2012.
of the types of sports analytics work being done both in [6] ESPNFC. http://espnfc.com/us/en/report/
industry as well as academia. Additionally, we gave a case- 292849/report.html?soccernet=true&cc=5901.
study which investigated possible reasons for why the home [7] K. Goldsberry. CourtVision: New Visual and Spatial
advantage exists in continuous sports like soccer. Using spa- Analytics for the NBA. In MIT Sloan Sports Analytics
tiotemporal data, we were able to show that teams at home Conference, 2012.
play have more possession in the attacking third. Coupled
[8] A. Gupta, P. Srinivasan, J. Shi, and L. Davis.
with the fact that the shooting and passing proficiencies are
Understanding Videos, Constructing Plots: Learning a
not significantly different, this observation can partially ex-
Visually Grounded Storyline Model from Annotated
plain why home teams have more shots and score more, and
Videos. In CVPR, 2009.
in-turn win more at home compared to away matches. Using
[9] Hawk-Eye. www.hawkeyeinnovations.co.uk.
our feature representation, we also showed examples where
pre and post game analysis can be performed. Specifically, [10] D. Henschen. IBM Serves New Tennis Analytics At
we were able to show the variation in home and away perfor- Wimbledon. www.informationweek.com/software/
mances for each team, as well as the ability to flag anomalous business-intelligence/
performances. ibm-serves-new-tennis-analytics-at-wimbl/
Our work also highlighted the importance of match con- 240002528, 23 June 2012.
text and the limiting factor it could have on training ex- [11] A. Hervieu and P. Bouthemy. Understanding sports
amples. In our future work, we will look at unsupervised video using players trajectories. In J. Zhang, L. Shao,
methods which cluster playing similar playing styles which L. Zhang, and G. Jones, editors, Intelligent Video
can enrich our training data set, without effecting its dis- Event Analysis and Understanding. Springer Berlin /
criminating power. Additionally, we are looking to extend Heidelberg, 2010.
this approach to focus on using player tracking information [12] S. Intille and A. Bobick. A Framework for Recognizing
to discover team formations and plays. Predicting team in- Multi-Agent Action from Visual Evidence. In AAAI,
teractions and subsequent performances and outcomes, es- 1999.
pecially when they have not played each other is another [13] K. Kim, M. Grundmann, A. Shamir, I. Matthews,
area focus of our research. As reliable high-level labels are J. Hodgins, and I. Essa. Motion Fields to Predict Play
almost impossible to obtain, predicting match outcomes as Evolution in Dynamic Sports Scenes. In CVPR, 2010.
our evaluation tool seems to be the best indicator of im- [14] M. Lewis. Moneyball: The Art of Winning an Unfair
proved team modeling. Game. Norton, 2003.

1373
[15] R. Li and R. Chellappa. Group Motion Segmentation [26] Opta Sports. www.optasports.com.
Using a Spatio-Temporal Driving Force Model. In [27] S. Pellegrini, A. Ess, K. Schindler, and L. van Gool.
CVPR, 2010. You’ll Never Walk Alone: Modeling Social Behavior
[16] R. Li, R. Chellappa, and S. Zhou. Learning for Multi-Target Tracking. In CVPR, 2009.
Multi-Modal Densities on Discriminative Temporal [28] M. Perse, M. Kristan, S. Kovacic, and J. Pers. A
Interaction Manifold for Group Activity Recognition. Trajectory-Based Analysis of Coordinated Team
In CVPR, 2009. Activity in Basketball Game. Computer Vision and
[17] W. Lu, J. Ting, K. Murphy, and J. Little. Identifying Image Understanding, 2008.
Players in Broadcast Sports Videos using Conditional [29] Prozone. www.prozonesports.com.
Random Fields. In CVPR, 2011. [30] B. Siddiquie, Y. Yacoob, and L. Davis. Recognizing
[18] P. Lucey, A. Bialkowski, P. Carr, E. Foote, and Plays in American Football Videos. Technical report,
I. Matthews. Characterizing Multi-Agent Team University of Maryland, 2009.
Behavior from Partial Team Tracings: Evidence from [31] SportsVision. www.sportsvision.com.
the English Premier League. In AAAI, 2012. [32] STATS SportsVU. www.sportvu.com.
[19] L. Madden. NFL to Follow Army’s Lead on Helmet [33] Statsheet. www.statsheet.com.
Sensors in Attempt to Prevent Head Injury. [34] D. Stracuzzi, A. Fern, K. Ali, R. Hess, J. Pinto, N. Li,
www.forbes.com/sites/lancemadden/2012/07/16/ T. Konik, and D. Shapiro. An Application of Transfer
nfl-to-follow-armys-lead-on-helmet-sensors-in/ to American Football: From Observation of Raw
-attempt-to-prevent-head-injury/, 16 July 2012. Video to Control in a Simulated Environment. AI
[20] R. Masheswaran, Y. Chang, A. Henehan, and Magazine, 32(2), 2011.
S. Danesis. Destructing the Rebound with Optical [35] X. Wei, P. Lucey, S. Morgan, and S. Sridharan.
Tracking Data. In MIT Sloan Sports Analytics Sweet-Spot: Using Spatiotemporal Data to Discover
Conference, 2012. and Predict Shots in Tennis. In MIT Sloan Sports
[21] V. Morariu and L. Davis. Multi-Agent Event Analytics Conference, 2013.
Recognition in Structured Scenarios. In CVPR, 2011. [36] C. Xu, Y. Zhang, G. Zhu, Y. Rui, H. Lu, and
[22] T. Moskowitz and L. Wertheim. Scorecasting: The Q. Huang. Using Webcast Text for Semantic Event
Hidden Influences Behind How Sports Are Played and Detection in Broadcast. T. Multimedia, 10(7), 2008.
Games Are Won. Crown Publishing Group, 2011. [37] Zonalmarking. www.zonalmarking.net.
[23] NBA Shot Charts. www.nba.com/hotspots.
[24] D. Oliver. Basketball on Paper: Rules and Tools for
Performance Analysis. Brassey’s, Incorporated, 2004.
[25] D. Oliver. Guide to the Total Quarterback Rating.
espn.go.com/nfl/story/_/id/6833215/
explaining-statistics-total-quarterback-rating,
4 August 2011.

1374