“Win at Home and Draw Away”: Automatic Formation Analysis
Highlighting the Differences in Home and Away Team Behaviors
Alina Bialkowski, Patrick Lucey, Peter Carr, Yisong Yue and Iain Matthews
Disney Research, Pittsburgh, PA, USA, 15213
Email:{alina.bialkowski, patrick.lucey, peter.carr, yisong.yue, iainm}@disneyresearch.com
Abstract
In terms of analyzing soccer matches, two of the most important factors to consider are: 1) the formation the team
played (e.g., 4-4-2, 4-2-3-1, 3-5-2 etc.), and 2) the manner in which they executed it (e.g., conservative - sitting deep, or
aggressive - pressing high). Despite the existence of ball and player tracking data, no current methods exist which can
automatically detect and visualize formations. Using an entire season of Prozone data which consists of ball and player
tracking information from a recent top-tier professional league, we showcase an automatic formation detection method
by investigating the “home advantage”. In a paper we published recently, using an entire season of ball tracking data we
showed that home teams had significantly more possession in the forward-third which correlated with more shots and
goals while the shooting and passing proficiencies were the same. Using our automatic formation analysis, we extend
this analysis and show that while teams tend to play the same formation at home as they do away, the manner in which
they execute the formation is significantly different. Specifically, we show that the position of the formation of teams at
home is significantly higher up the field compared to when they play away. This conservative approach at away games
suggests that coaches aim to win their home games and draw their away games. Additionally, we also show that our method can
visually summarize a game which gives an indication of dominance and tactics. While enabling new discoveries of team
behavior which can enhance analysis, it is also worth mentioning that our automatic formation detection method is the
first to be developed.
1 Introduction
As chronicled in Jonathan Wilson’s Inverting the Pyramid [1], the many tactical and strategic revolutions that have occurred in soccer
over the last century can be observed in the variations in formations utilized by coaches and managers over this time. For example,
in the very first international match in 1872, Wilson notes that England played what looked like a 1-2-7 (one full-back, two midfielders and seven attackers), while Scotland played a 2-2-6. As the game evolved through various rule-changes (i.e., off-side rule)
and professionalism (i.e., players could train full-time), so too did formations where coaches/managers set up their team’s
formation to best maximize the chances of their team winning while trying to minimize the chances of the opposition. Prime
examples of such formations are the dour and stifling “Catenaccio”, often employed by Italian teams where the emphasis is getting
behind the ball waiting for the opposition to make a mistake and hit them on the counter-attack, or the dynamic and fluid “TotalFootball” introduced by Rinus Michaels through the Ajax and Dutch teams in the 60’s and 70’s which has evolved to the “TikiTaka” style of football Barcelona play today using a 4-3-3 formation (N.B. Lobanovskyi used a similar style with the Dynamo Kiev
team at the same time with similar success). While the differences in these formations are readily evident to the trained eye and
despite the fact that most professional soccer teams are extremely sophisticated in terms of their knowledge of team tactics and
strategy, analytical measures which can quantify such team behaviors are lacking. This is understandable though as compared to
other sports, soccer is a dynamic, continuous game with events (e.g., shots, goals etc.) occurring sparsely. Coupled with the fact that
players continuously swap position or role within a formation, making meaningful comparisons is very difficult.
While challenging, some objective measures are starting to emanate from new data sources that contain ball-events (e.g., Opta [2]),
or ball and player tracking information (e.g., Prozone [3]). In The Numbers Game [4], the authors highlight some recent measures
which debunk some commonly-held beliefs. Notable examples include, “a team is most likely to concede a goal just after it has
scored”, and “the greater the number of corners a team has, the more likely they are to score”. In both these examples they show
that this is in fact not the case, with teams least likely to concede after they score a goal in the first example, and that the relationship
between goals and corners is essentially zero in the second example. Contrastingly though, another commonly-held belief in soccer is
that of the “home advantage” - where teams are more likely to win at home compared to away. In Scorecasting [5], Moskowitz and
Wertheim uphold this belief by highlighting that the home advantage exists in all professional sports and suggest that referees play a
significant role by giving home teams favorable calls at critical moments. Specifically for soccer, they show that event statistics such
as the amount of injury time, number of yellow cards and number of penalties awarded to home-teams reinforce their hypothesis.
2014 Research Paper Competition
Presented by:
ICCV 2011 Submission #525. CONFIDEN
#525
864
ICCV
ICCV
#525
which burden
positions
switch
each
other.
Forand
basketball
(topretriev
can occur,
as with
well
as efficient
effective
a
heavy
currently
required
for such
analysis.
866868
864
865
866
867
868
869
870
871
872
873
ICCV
874
#525
875
876
877
878
864
879
865
880
866
881
867
882
868
883
869
884
870
885
871
886
872
887
873
ICCV
888
874
#525
889
875
890
876
891
877
892
878
864
893
879
865
894
880
866
895
881
867
896
882
868
897
883
869
898
884
870
899
885
871
900
886
872
901
887
873
902
888
874
903
889
875
904
890
876
905
891
877
906
892
878
907
893
879
908
894
880
909
895
881
910
896
882
911
897
883
912
898
884
913
899
885
914
900
886
915
901
887
916
902
888
917
903
889
904
890
ICCV 2011 Submission #525.
of a single soccer team. The results are shown in Figure ??.
865
864 In terms of formation analysis, we are interested in
of a single soccer team. The results are shown in F
866
#525
which positions switch with each other. For basketball (top
865
ICCV
In terms
of formation#525.
analysis,
we are inteR
ICCV
ICCV
2011 Submission
CONFIDENTIAL
867
row), we can see that .... . For soccer (bottom row), we see
ICCV 866
#525
which positions switch with each other. For baske
868
#525
that....In terms
pre-game
planning,
a tool or met#525 867
ICCVof2011
Submission
#525. such
CONFIDENTIAL
REVIE
869
ICCV
2011we
Submission
#525.
CONFIDENTIAL
REVIEW
row),
can
see
that
....
.
For
soccer
(bottom
row
ICCV
2011
Submission
#525.
CONFIDENTIAL
ric can give an objective measure of how a team tends to
868
870
that....In
offertile
pre-game
planning,
such
a too
864
operate.
This
aterms
futureare
areainof
research
as semiICCV
[
of a singleICCV
soccer
team.
Theisresults
shown
Figure
??.
869
871
ric or
canunsupervised
give an objective
measure
of how
aplays
team
865 864
#525 Inofterms
supervised
learning
good
and
bad
#525
[6]
K
a
single
soccer
team.
The
results
are
shown
in
Figure
??.
of
formation
analysis,
we
are
interested
in
870
ICCV
864
872
864
ICCV
2011
Submission
#525.
CONFIDENTIAL
CO
operate.
This
isICCV
aresults
future
fertile
area
ofREVIEW
research
865
of aoccur,
single
soccer
team.
The
are
shown
in
Figure
??.
866
Submission
#525.
CONFIDEN
can
as
well
as
efficient
and
effective
retrieval
which
F[
[6]
K.DO
K
singlepositions
soccer
team.
The
results
are
shown
in
??.
In
terms
of
formation
analysis,
we 2011
areFigure
interested
in
ICCVof awhich
871
#525
ICCV
2011
Submission
#525.
CONFIDENTIAL
REVIEW
COPY.
switch
with
each
other.
For
basketball
(top
865
873
supervised
or
unsupervised
learning
good
and
b
In
terms
of
formation
analysis,
we
are
interested
in
865867 866
will
greatly
help
coaches
and
commentators
alike
due
to[7]
the
ICCV
2011
Submission
#525.
CONFIDENTIAL
REVIEW
COPY.
DO R
N
Fore
which
switch
other.
For
basketball
(topsee
#525 row),
In terms
ofpositions
formation
we
are
interested
866
872
we
can
see
that
....analysis,
.with
Foreach
soccer
(bottom
row),in
we
874
867
row),
we
can
thatICCV
....
. 2011
For
soccer
(bottom
row),CONFIDENTIAL
we see [7] R.
Submission
#525.
REV
864 positions
L[
873
which
switch
with
each
other.
For
basketball
(top
864
875
that....In
terms
ofsee
pre-game
planning,
such
a(bottom
tool
or
met[6]see
K.
Kita
of
a867
single
soccer
team.
The
results
shown
in and
Figure
??.
row),
we
can
see
that
....are
. For
soccer
row),
ICCV
ICCV
of
apre-game
single
soccer
team.
The
results
are
shown
inwe
Figure
??.
will
greatly
help
coaches
commentators
alike
[8]
867869 868
LB
that....In
terms
of
planning,
such
a tool
met865
868 865
a SpJ
Forecas
874
can
see
that
. are
For
soccer
row),
weweor
see
In
terms
of....
formation
analysis,
weplanning,
are
interested
in
876
that....In
terms
of
such
afor
tool
ortomet869
ric we
can
give
an
objective
measure
of
how
ateam
team
tends
#525
864
In
terms
ofpre-game
formation
analysis,
are
interested
inTrw
#525
of868
a870
singlerow),
soccer
team.
results
shown
in(bottom
Figure
??.
spondence.
heavy
burden
currently
such
analysis.
866a single
LCB
of
soccer
team.
The
results
are
shown
in Figure
??.
spondence.
869The
ric
can
give
an
objective
measure
of
how
arequired
tends
to
866
[7]
LiLi
an
[8]toR.#525.
J.Tran
which
positions
switch
with
each
other.
For
basketball
875
ric acan
give
anfertile
objective
measure
of other.
how
a(top
team
tends
877This
ICCV
2011
Submission
#525.
CONFIDEN
that....In
terms
of
pre-game
planning,
such
a
tool
or
met870
865
which
positions
switch
with
each
For
basketball
(top
2
ICCV
2011
Submission
operate.
is
future
area
of
research
as
semitems,
32(1):48
tems, 32(1):486–
LB
In
terms
ofIn formation
analysis,
we
are
interested
in
870
terms
ofwe
formation
wesoccer
arearea
interested
in weassee
operate.
This
is that
aanalysis,
future
fertile
of research
semi864 867
RLB
a Spatio
867
869
row),
can
see
....
. is
For
(bottom
row),
with
871
operate.
This
asee
future
fertile
area
of Figure
research
as semi876
878an
[6]J[
of positions
agive
single
soccer
team.
The
results
are
shown
in
??.
866
row),
we
can
that
....
.good
For
soccer
(bottom
we[9]
see
871
ric
can
objective
measure
of
how
agood
team
tends
to
868
[4]
S.row),
Intille
supervised
or
unsupervised
learning
and
bad
plays
871with
[4]
S.
Intille
and
which
switch
with
each
other.
For
basketball
[8] J.and
Liu,A.P
supervised
or
unsupervised
learning
and
plays
LCB
which
positions
switch
each
For
basketball
(top
868
RB
865
that....In
terms
ofother.
pre-game
planning,
such
a (top
tool
orbad
met870872
supervised
or
unsupervised
learning
good
and
2Acti
that....In
terms
of pre-game
planning,
such
abad
toolplays
or with
met867
879
877
ti
872
In
terms
of
formation
analysis,
we
are
interested
in
Multi-Agent
869
operate.
This
is
a
future
fertile
area
of
research
as
semico
872
row),
we
can
see
that
....
.
For
soccer
(bottom
row),
we
see
Multi-Agent
A
can
occur,
as
well
as
efficient
and
effective
retrieval
which
869
can
occur,
as
well
as
efficient
and
effective
retrieval
which
m
ric
can
give
an
objective
measure
of
how
a
team
tends
to
RLB
row),
we
can
see
that
....
.
For
soccer
(bottom
row),
we
see
LCM
864
can ric
occur,
as
well
asobjective
efficient measure
and effective
retrieval
which
can soccer
give
an
of shown
how
a S.
team
tends
toLi
868866 −5
871873
880
[9]and
J.[7]
P
873
−5
[1
878
864
870 which
[5]
Khan
2[10]
of
acoaches
single
team.
results
are
in
Figure
??.
positions
switch
each
other.
For
basketball
873
that....In
terms
of
pre-game
planning,
such
aThe
tool
or met870
supervised
or
unsupervised
learning
good
and
bad
plays
will
greatly
help
coaches
and
due
to (top
the
of with
aRB
single
soccer
The
are
shown
inM.aFM
operate.
This
isoperate.
a future
fertile
of
research
asresults
semi[5]
S.
Khan
and
will
help
and
commentators
alike
due
to
the
ICCV
will
greatly
help
coaches
and
commentators
alike
due
to
RCM
865
that....In
ofgreatly
pre-game
planning,
such
acommentators
tool
orteam.
met-alike
This
is
aarea
future
fertile
area
of
research
asthe
semi869867terms
Rigidity
of
881
874
871can
tial
872
[9]
J. Form
Liu
ap
879
865
874
terms
formation
analysis,
we
are
interested
in
874
ric
give
anwell
objective
measure
of
how
afor
team
tends
tobad
row),
we
can
see
that
....required
.of
For
soccer
(bottom
row),
we Rigidity
see
871
supervised
or In
unsupervised
good
and
plays
heavy
burden
currently
such
analysis.
can
occur,
as
as
efficient
and
effective
retrieval
which
of
For
Inrequired
terms
of
formation
analysis,
we
inte
heavy
burden
currently
required
for
such
analysis.
heavy
burden
currently
for
such
analysis.
supervised
oralearning
unsupervised
learning
good
and
badare
plays
866
870868
CM
#525
LCM
ric873
can
give
an
objective
measure
of
how
team
tends
to
tial
patt
[6]
K.
Kitani,
B.
Zie
872
882
875
[10] (top
P.[8]LT
875
866
operate.
This
is
a 880
future
fertile
areaswitch
ofand
research
as
semiwhich
positions
withsuch
each
other.
For
basketball
872
875 will
that....In
terms
pre-game
planning,
adue
tool
or
metcan
occur,
as of
well
as
efficient
effective
retrieval
which
LB
ICCV
can
occur,
as
well
as
efficient
and
effective
retrieval
which
LB
which
positions
switch
with
each
other.
For
baske
greatly
help
coaches
and
commentators
alike
to
the
[6]
K.
Kitani,
B.
ICCV
871
ICCV
2011
Submission
#525.
CONFIDENTIAL
R
867
Forecasting.
In
[11]
IEZ
[10]
P.
Luce
873
LW
LB
m
w
RCM
876
883
operate.
is
acan
future
fertile
area
of
research
as
semi876
supervised
or
unsupervised
learning
good
and
bad
plays
873
and
[1
881
row),
wegreatly
can
see
that
....
. For
soccer
row),
wetosee
867
will
greatly
help
coaches
and
commentators
alike
due(bottom
totends
the alike
874876869This
ric
give
an
objective
measure
of
how
a
team
to
LCB
will
help
coaches
and
commentators
due
the
#525
LCB
and
Y.
872
it
row),
we
can
see
that
.... . For soccer
In
[7] Forecasting.
R.(bottom
Li and R. row
Che
874 LCB
868
burden
currently
for
such
analysis.
#525 870 heavy
RWrequired
877
m
884
can
occur,
as877
well
efficient
and
effective
retrieval
which
874as
CM
supervised
or unsupervised
learning
good
and
bad
plays
heavy
burden
required
for
such
analysis.
Team
that....In
terms
ofm
pre-game
planning,
such
a aw
tool or Team
met882
868
875877
ICCV
2011
Submission
#525.
CONFIDENTIAL
R
iscurrently
a future
fertile
area
ofpre-game
research
as[7]
semiheavy
burden
currently
required
for
such
analysis.
RLBThis
873
RLB
875 operate.
[12]
K
that....In
terms
of
planning,
such
aR.
too
869
ICCV
2011
Submission
#525.
CONFIDENTIAL
REVIEW
R.Spatio-Tempora
Li
and
CB
LB
ST
875
878
885
will
greatly
help
coaches
and
commentators
alike
due
to
the
w
w
w
871 878
LB
[9]
LW
RLB
ric
can
give
an
objective
measure
of
how
a
team
tends
to
can
occur,
as
well
as
efficient
and
effective
retrieval
which
[11]
I. M
869
[1
[11]
I. Matth
Liu,
P. Carr,
Y.
876878
874
or883
unsupervised
learning
good and
bad[8]plays
876 supervised
o
RB
RB LB
870
ric
can
give
an objective
measure
of
how
aited.
team
aJ. Spatio-Temp
864
876
879
heavy
burden
currently
required
for
such
analysis.
886
LCB
w
wis
wIn6[
872 879
LCB
operate.
This
aalike
future
fertile
areainofFigure
research
as m
semiRW
ofICCV
aRB
single
soccer
team.
The
results
are
??.
ited.
will
greatly
help
coaches
and
commentators
due
toshown
theretrieval
with
context-con
870
ICCV
875
877
LCM
can
occur,
as884
well
asLCB
efficient
and
effective
which
LCM
877879
871
operate.
This
is a future
area
of
research
880 877
GK
#525
[8]
J.
P.K.
Carr,
865873 880RLB
887
RLB
8.
Summary
and
Future
Work
mresults
m fertile
[12]
Mur
864
[10]
supervised
or analysis.
unsupervised
learning
good
and
bad
2 Liu,
#525
ST
878
876burden
885
871
[13]
S
K.
M
heavy
currently
required
forRLB
such
In greatly
terms
of
formation
analysis,
we
are
interested
in[12]plays
[6
of
single
soccer
team.
The
are
shown
inCONFIDENTIAL
Figure
??.
LCM
RCM
ICCV
RCM
will
help
coaches
and
commentators
alike
due
to with
thegood
ICCV
2011
Submission
#525.
REVIE
878
872
881a878
LB
880
881 865
supervised
or
unsupervised
learning
and
864
order
ob
context-co
RB
888
866
ICCV
2011
Submission
#525.
CONFIDEN
[9]
J.
Liu
and
Y.
Liu
m
w
m
can
occur,
as
well
as
efficient
and
effective
retrieval
which
874
879
RB
[6]
K.
K
of
a
single
soccer
team.
The
results
are
shown
in
Figure
??.
877
N[
ICCV RB
886
872
orde
[1
In
of
formation
analysis,
we
are
interested
in
CM
#525
ICCV
2011
Submission
#525.
CONFIDENTIAL
REVIEW
COPY.
DO
positions
switch
with
each
other.
For
basketball
(top
GK
879terms
CM
RCM
882
heavy
burden
currently
required
for
suchasanalysis.
In
thiscan
paper,
we
proposed
aefficient
method
of
tracking
sports
873
687,
LCB
882which
879
LCM
occur,
as
well
and
effective
2tial
889
865881
patterns.
In19
CN
866
867
will
greatly
help
coaches
and
commentators
alike
due
toretriev
the
T
878
LCM
875 880
ICCV
2011
Submission
#525.
CONFIDENTIAL
REVIEW
COPY.
DO[7
Fore
#525
687,
In
terms
of
formation
analysis,
we
are
interested
in
887
873
which
positions
switch
with
each
other.
For
basketball
(top
LW
880
8.
Summary
and
Future
Work
LCM
883
row),
we
can
see
that
....
.
For
soccer
(bottom
row),
we
see
[13]
S.
Pelle
LW
LB
players
using
the
team
formation
as
our
contextual
feature.
874
RCB
883
LB
RCM
CM
[10]
P.
Lucey,
A.
Bia
will
greatly
help
coaches
and
commentators
alike
880
890
867
[9] we
J. Liu
and
Y.
L
866882
[11]
879876 881
heavy
burden
currently
required
for
such
analysis.
864
[14]
RCM
868
ICCV
2011
Submission
#525.
CONFIDENTIAL
REV
881
[7]
R.
row),
we
can
see
that
....
. For
soccer
(bottom
row),
see
RW
888
874
[6]
K.
of
a switch
single
soccer
team.
The
results
are
shown
Figure
884
S.
PLP[
which
positions
with
each
other.
ForThe
basketball
(top
864
RW
We
depart
from
previous
formation
tracking
work
byNever
alRBLW
875
that....In
terms
of
pre-game
planning,
such
a inare
tool
or??.
metand
Y.[13]
Sheikh.
884
RCM
CM
LCB
891
882
heavy
burden
currently
required
for
such
analysis.
of
aCM
single
soccer
team.
results
shown
in Figure
??.
868
881
tial
patterns.
In
865LCB
880
[15]
FT
883
867
LB
[8
w
In
this
paper,
we
proposed
a
method
of
tracki
Target
882
Fo
869
877 row),
that....In
terms
of
pre-game
planning,
such
a
tool
or
meta
Sp
ST
In
terms
of
formation
analysis,
we
are
interested
in
885
889
865
875
Nev
[1
Team
LCM
we
can
that
.... LW
. are
For
soccer
(bottom
row),
we
see
lowing
non-rigid
which
iswemore
indicative
of
ST
876
LWsee
885
ric
can
give
an
objective
measure
of
how
a team
tends
toBehaviors
883
864
892
In
terms
ofdeformations,
formation
analysis,
arespondence.
interested
inTr
869
866
881
of882
aRCB
single
soccer
team.
The
results
shown
in
Figure
??.
RLB
LB
[10]
P.
Lucey,
A.
B
RW
[12]
[14]
883
[7]
R.ti
884
LCB
ofCM
aRCM
single
soccer
team.
The
are
shown
in
Figure
??.
spondence.
Tran
wresults
mm
868
players
using
the
team
formation
as
our
contextua
which
positions
switch
with
each
other.
For
basketball
(top
ric
can
give
anbehavior
objective
measure
of
how
asports.
team
tends
to
886
866
870
878 that....In
[11]
I.
Matthews
[8]Prozon
J.and
Li
Targ
876
RW
player
in
continuous
team
model
the
[1S
877
886
884
terms
of890
pre-game
planning,
such
a each
tool
or
met865
positions
switch
with
For To
basketball
893
870
882
operate.
This
isThis
awhich
future
fertile
area
of
research
semitems,
32(1):48
RW
a AS
RB
[15]
F.(top
Ram
tems,
32(1):486–
LW
884
RB
LCB
In
terms
of
formation
analysis,
are
interested
in research
and
Y.
Sheikh
In867
terms
of
formation
analysis,
we
are
interested
inother.
883
row),
we
see
that
....
. For
soccer
(bottom
row), as
we
864
RLB
ST
887
885
867
depart
from
previous
tracking
wo
operate.
isWe
awe
future
fertile
area
assee
semiited.
Internationa
869
8.can
Summary
Future
Work
with
871
CM
891
ST
[
877
885
local
player
behaviors
within
a of
team
formation,
we
use
the
887
878
[14]
Proz
[6]
of
soccer
team.
The
results
are
shown
information
Figure
??.
866879 ric
row),
we
canand
see
that
.... a.good
For
soccer
(bottom
row),
we
see
868a
8.single
Summary
and
Future
Work
883
can
give
an
objective
measure
of
how
team
to
[16]
M
894
871
m
tive
rob
[9
[8]
J.
ST
[4]metS.plays
Intille
and
A
885
supervised
or
unsupervised
learning
and
bad
plays
[4]
S.
Intille
and
which
positions
switch
with
each
other.
For
basketball
(top
that....In
terms
of pre-game
planning,
such
atends
tool
or
Team
Behavio
LCM
RLB
which
positions
switch
with
each
other.
For
basketball
(top
868
LCM
RW
888
[12]
K.
Murty.
An
al
RB
supervised
or
unsupervised
learning
good
and
bad
865
884
lowing
non-rigid
deformations,
which
is
more
ind
886
870
LW
886
2
880
888
player
role
asterms
our
representation
and
we
show
ain
novel
and
892
872
that....In
ofproposed
pre-game
planning,
such
aMulti-Agent
tool
or Autono
met869In 895
878
867
879
884
wi
[15]
F.Act
Ra
terms
of
formation
analysis,
we
are
interested
872
886
In
this
paper,
we
awe
method
of[11]
tracking
sports
operate.
This
isoccur,
apaper,
future
fertile
area
of
research
as
semirow),
we
can
see
that
....
.as
For
soccer
(bottom
row),
see
ric
give
an
objective
measure
of
how
aretrieval
team
tends
to
Multi-Agent
AD
can
occur,
as
well
efficient
and
effective
which
order
of increasi
869
[13]
889
I.which
Matthews
an
can
as
well
as efficient
and
effective
retrieval
RB
row),
we
see
that
....
.can
For
soccer
(bottom
row),
we
see
RCM
LCM
In
this
we
proposed
aassigning
method
of
tracking
sports
RCM
866
ST
player
behavior
in
continuous
team
sports.
To
m
887
885
RW
887
8.
Summary
and
Future
Work
870
ric
can
give
an
objective
measure
of
how
a(top
team
tends
to
868
871
889
2[1
881can
accurate
method
of
this
higher-level
feature
[16]
M.
Rod
880
[9]and
J.by
Li
8.
Summary
and
Future
Work
893
879
885
[7]
[17]
L
873
[5]
S. feature.
Khan
M.
tive
[10
m
w
873
887
which
positions
switch
with
each
other.
For
basketball
896
players
using
the
team
formation
as
our
contextual
operate.
This
is
a
future
fertile
area
of
research
as
semi8.
Summary
and
Future
Work
that....In
terms
of
pre-game
planning,
such
a
tool
or
met687,
1968.
2,
5,
8
870
supervised
or
unsupervised
learning
good
and
bad
plays
890
[5]
S.to
Khan
and
M
will
help
coaches
and
commentators
alike
to
the
ited.
will using
greatly
help
coaches
and
commentators
alike
due
theInternatio
888
players
the
team
formation
as our
contextual
feature.
CC
RCM
LCM
ST871
867terms
that....In
ofgreatly
pre-game
planning,
such
a constraint
tool
ormfertile
metDriven
operate.
This
is
abehaviors
future
area
research
semilocal
player
aofdue
team
formation,
CM
CM
[9]
J.w
869882
886
890
enforcing
afrom
planar
onwithin
the
input
descriptor.
In
886
Rigidity
of
Form
888
881
tial
872
894
880
888
874
Auto
We
depart
previous
formation
tracking
work
byasal874
supervised
orwe
unsupervised
learning
good
and
bad
plays
897
m
w
ric
can
give
an
objective
measure
of
how
a team
tends
to
row),
we
can
see
that
....
.and
For
(bottom
row),
we
see
[13]
S.
Pellegrini,
A.2p
871
In
this
paper,
proposed
asoccer
method
of
tracking
sports
891
m
can
occur,
as
well
as
efficient
effective
retrieval
which
889
Rigidity
of
For
heavy
burden
currently
required
for
such
We
depart
from
previous
formation
tracking
work
by
alheavy
burden
currently
required
for
such
analysis.
872
[17]
L.
H.
Rn
[12]
K.
Murty.
An
supervised
or
unsupervised
learning
good
and
bad
plays
PG
In
this
paper,
we
proposed
a analysis.
method
of
tracking
sports
870868
CM
In
this
paper,
we
proposed
aresearch
method
of
tracking
sports
player
role
as
our
representation
and
we
show
atia
RCM
891
ric887
can
give
an
objective
measure
of
how
aplanning,
tends
to
887
[6]
K.
Kitani,
B.
Zie
LW
[18]
terms
of
multi-agent
tracking,
we
implicitly
assume
that
LW
882
889
lowing
non-rigid
deformations,
isfeature.
more
indicative
of
can
occur,
aspre-game
well
as
efficient
and
retrieval
which
Never
Walk
Alo
881
875
[10]
P.[8]
LS1R
Summary
and
Future
Work
873
operate.
This
is889
a 895
future
fertile
area
ofteam
aswhich
semi898
w
weffective
w
872
[1
players
using
the
team
formation
as
our
contextual
[16]
M.
875883 8.
that....In
terms
of
such
adue
tool
or
met892
890SG
[14]
2004.
[10]
P.
873
LB
lowing
non-rigid
deformations,
which
is
more
indicative
of
can
occur,
as
well
as
efficient
and
effective
retrieval
which
m
m
w
will
greatly
help
coaches
and
commentators
alike
to
the
[6]
K.
Kitani,
B.
Z
players
using
the
team
formation
as
our
contextual
feature.
871
order
of
increa
m
O
Forecasting.
In
E
888
LW
LB
892
accurate
method
of
assigning
this
higher-level
f
890
869This
[11
C
players
using
the
team
formation
as
contextual
feature.
roles
change
are
relatively
infrequent,
or
temporally
sparse,
CM
Target m
Tracking.
883
operate.
is
aRW
future
fertile
area
of
asour
semiwill
greatly
help
coaches
and
commentators
alike
due
to the
player
behavior
in
continuous
team
sports.
To
model
the[19]
876
supervised
or
unsupervised
learning
good
and
bad
plays
RW
M
888
873
896
and
884
882
We
depart
from
previous
formation
by
al899
[1
890
w
wresearch
w
893
874
ICCV
Driv
891
ric
can
give
an
objective
measure
oftracking
how
awork
team
tends
to
876
[18]
STATS
874
will
greatly
help
and
commentators
alike
due
to
the
[15]
We
depart
from
previous
formation
tracking
work
by
alplayer
behavior
in
continuous
team
sports.
To
model
the
LCB
872
SF
wassumption
w
w
Forecasting.
In
[7]the
R.
Li
and
R.
Che
687,
1968.
2,an
burden
currently
required
for
such
analysis.
889
r5
891
893
RW
[14]
Prozone.
www.p
heavy
burden
currently
required
for
such
analysis.
enforcing
acoaches
planar
constraint
on
input
descr
and
based
on
this
we
show
that
this
can
enhance
local
player
behaviors
within
aplays
team
formation,
we
use
the
884
In
this
paper,
we
proposed
aeffective
method
of
tracking
sports
870 heavy
LCB
can
occur,
as894
well
as
efficient
and
retrieval
which
depart
from
previous
formation
work
by
al874
LW
lowing
non-rigid
which
istracking
more
indicative
of
877
supervised
orWe
unsupervised
learning
good
and
bad
900
897
mm
883
885
#525
M
Te
889
ST
892
891
[19]
C.Team
Sutto
[1
875
[17]
L.
H
875ST
operate.
This
is
a deformations,
future
fertile
area
of
research
as
semiheavy
burden
currently
required
for such
analysis.
lowing
non-rigid
deformations,
which
is
more
of
877
a indicative
Spatio-Tempor
873
local
player
behaviors
within
am
team
formation,
we
use
PF
RLB
890
in
w
wcontextual
892
mwhere
[12
[7]
R.
Li
and
R.
894
[15]
F.
Ramos
and
H.
[13]
S.
Pellegrini,
LB
player
role
as
ourinas
representation
and
we
show
athe
novel
and
LB
ST
player
situations
players
and
875
885
will
greatly
help
coaches
and
commentators
alike
due
to
player
behavior
intracking
continuous
team
sports.
Tothe
model
the
terms
multi-agent
tracking,
wecross-over
implicitly
ass
2011
Submission
#525.
CONF
871 players
878
using
the
formation
our
feature.
[9]
895
[11]
I.CA
M
lowing
non-rigid
deformations,
which
isICCV
more
of
RW
random
RLB
901
893
m
w
m
can
occur,
as
well
as
efficient
and
effective
retrieval
which
898
884
[11]
I.
M
876
M of
890
PG
LB
892
2004
player
behavior
in
continuous
team
sports.
ToLiu,
model
the
[1
[8]
J.
P.by
Carr,
Y.
876
874886 895
orteam
unsupervised
learning
good
and
bad
plays
player
role
as
our
representation
and
we
show
aindicative
novel
and
Csupervised
878
891
893
tive
robots
within
[20]
Y
RB
m
m
LCB
a
Spatio-Temp
accurate
method
of
assigning
this
higher-level
feature
Never
Walk
A
876
local
player
behaviors
within
a
team
formation,
we
use
the
heavy
burden
currently
required
for
such
analysis.
are
occluded.
886
ite
ing,
4(4
roles
change
are
relatively
infrequent,
or
temporal
879
896
894
872 We
LCB
902
depart
from
previous
formation
tracking
work
al877
ST m
player
behavior
in
continuous
team
sports.
Toby
model
the
w
899
885
ited.
will
greatly
help
coaches
and
commentators
alike
due
toonthe
local
player
behaviors
within
a retrieval
team
formation,
we
use[16]
the
with
context-con
887
RB
LCB
875
891
8.
Summary
and
Future
Work
893
accurate
method
of
assigning
this
higher-level
feature
by
SG
[18]
STA
Autonomous
Age
894
892
can
occur,
as
well
efficient
and
effective
which
L
877
879
RLB
enforcing
aW
planar
constraint
the
input
descriptor.
In
896
877
player
role
as
ouras
representation
and
we
show
a novel
and
mTarget
w
GK
Trackin
[12]
K.
[20]P.
Y.
Zhan
[8]
J.
Carr,
887
897LCM
895
880
and
based
on
this
assumption
show
that
this
can
8.
Summary
Future
Work
878non-rigid
[10]
873
903
2InLiu,
player
role
asand
our
representation
and we
we
show
athe
novel
and
RLB
lowing
deformations,
which
is
more
indicative
of
900
RLB
886
local
player
behaviors
within
a
team
formation,
we
use
876
m
w
888
[12]
K.
M
heavy
burden
currently
required
for
such
analysis.
m
m
w
895
[16]
M.
Rodriguez,
J
893
enforcing
a
planar
constraint
on
the
input
descriptor.
LCM
W
8.
Summary
and
Future
Work
[13
892
ICCV
864
894
will
greatly
help
coaches
commentators
alike
due
the [19]
C.orS
SF
accurate
method
of
assigning
this
higher-level
feature
byto
878
terms
of of
multi-agent
tracking,
we
implicitly
assume
that
RB
878
Learnin
LB
880 897
896
898RCM
with
context-co
888
[14]
Prozone.
www.
a and
single
soccer
team.
The
results
are
shown
in
Figure
881
879In 904
References
player
tracking
in
situations
where
players
cross
[9]
J.
Liu
and
Y.
Liu
accurate
method
of
assigning
this
higher-level
feature
by
w
this
paper,
we
proposed
a#525.
method
ofdescriptor.
tracking
sports
RB
m
w
877874
Driven
Crowd
An
901
887
896
ICCV
m
m
wor
894
behavior
in
continuous
team
sports.
To
model
the
orde
RB
[1D
889 player
player
role
as
representation
and
we
atemporally
novel
and
terms
of
multi-agent
tracking,
we
implicitly
assume
that
8.paper,
Summary
and
Future
Work
#525
68
ICCV
2011
Submission
CONFIDENTIAL
REVIEW
COPY.
w
ware
w
865
GK
enforcing
aour
planar
constraint
on
the
input
In
879
[2
[17]
roles
change
relatively
infrequent,
sparse,
LCM
rand
RCM
893
heavy
burden
currently
required
for
such
analysis.
Inenforcing
this
we
proposed
ashow
method
of
tracking
sports
898
895
LCB
897
PF
879
899CM
880
881
2F.
889
882
In
terms
of
formation
we
are
interested
tial
patterns.
C
In
this paper,
proposed
aICCV
method
of
tracking
sports
aw
planar
constraint
onanalysis,
the
input
descriptor.
In
[15]
Ramos
and
H
are
occluded.
905
w
mm
878890
LCM
[17]
L.
H.
Rose.
The
players
using
the
team
formation
our
contextual
feature.
897
875 local
2011
Submission
#525.
CONFIDENTIAL
REVIEW
CO
895
wwe
was
902
888
S.
#525
687,
roles
change
are
infrequent,
or
sparse,
m
mon
w
866
terms
of
multi-agent
tracking,
we
implicitly
assume
that
880
player
behaviors
within
athis
team
formation,
we
use
the
and
based
assumption
we
show
that
this
enhance
accurate
method
of
assigning
this
higher-level
feature
by
RCM
LCM
[1]
S.relatively
Belongie,
J.
Malik,
and
J.temporally
Puzicha.
Shape
and
w
w
mcantive
w
w[13]
ing,
899
LB
players
using
the
team
formation
as
our
contextual
feature.
898
RCB
900
LW
894
LB
881
CM
896
which
positions
switch
with
each
other.
For
basketball
[10]
P.matching
Lucey,
A.
Bia
883
w
880
890
C
using
the
team
formation
our
contextual
feature.
terms
ofInas
multi-agent
tracking,
we
implicitly
assume
that
882
[9]
J.
Liu
Y.
L(
robots
wit
2004.
1and
[11]
906
879876 players
898
RCM
this
paper,
we
proposed
am
method
of
tracki
Ne
896
[14
We
depart
from
previous
formation
tracking
work
by
alICCV
2011
Submission
#525.
CONFIDENT
w
wdepart
w
w
roles
change
are
relatively
infrequent,
or
temporally
sparse,
867
881
and
based
on
this
assumption
we
show
that
this
can
enhance
903
889
player
tracking
in
situations
where
players
cross-over
and
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
CM
[13]
S.
PZ
w
w
w
891
899
player
role
as
our
representation
and
we
show
a
novel
and
[18]
900
enforcing
a
planar
constraint
on
the
input
descriptor.
In
We
from
previous
formation
tracking
work
by
alRB
901
882
m
m
m
m
and
Y.
Sheikh.
RCM
[20]
Y.we
row),
weare
canrelatively
see that
....
For
(bottom
row),
LCB
mm
Fbym.formation
W
roles
change
infrequent,
or
temporally
sparse,
We
depart
fromRW
previous
formation
tracking
work
al- soccer
884
895
LCB
891
897
w
mm
LW
[18]
STATS
SportsVU
881
tial
patterns.
In
Ta
880
Autonomous
A
899
907
883
CM
897
players
using
the
team
as
our
contextua
[15
and
based
on
this
assumption
we
show
that
this
can
enhance
882
868
m
m
m
lowing
non-rigid
deformations,
which
is
more
indicative
of
References
player
tracking
in
situations
where
players
cross-over
and
877
are
occluded.
522,
2002.
5
904
890
LW
900
Nev
[1
Team
883non-rigid
901
902
892 accurate
LCM
wof
wST
massume
w[19]
wC.
lowing
deformations,
which
is
more
indicative
ofA.
mnon-rigid
mthis
[19]
that....In
terms
ofFigure
pre-game
planning,
such
a Behaviors
tool
or
mT
method
of
assigning
this
higher-level
feature
bythis
terms
multi-agent
tracking,
we
implicitly
that
and
based
on
assumption
we
show
can
enhance
Lear
lowing
deformations,
which
is
more
indicative
of
mare
wthat
m
885
864
892
CM
min
m
Sutton
and
RCB
[14]
Pr
900
881
896
a898
single
soccer
team.
The
results
are
shown
??.
spondence.
RLB
898
LW
908
882
[10]
P.
Lucey,
A.
BS
RW
[12]
player
tracking
in
situations
where
players
cross-over
and
[16]
M.
Rodriguez,
883
884
of
aoccluded.
single
soccer
The
results
shown
in To
Figure
??.
sponde
8.of
Summary
and
Future
Work
869
We
depart
from
previous
formation
tracking
wo
m
w
m
are
901RCM
RW
player
behavior
in
continuous
team
sports.
model
the
878 902
[2]
P.team.
Carr,
Y.
Sheikh,
and
I.
Matthews.
Monocular
Object
De905
884
[11]
I. m
Matthews
and
891
w
m
903
Targ
ric
can
give
ansituations
objective
measure
of
how
amodel
team
tends
player
behavior
inthe
continuous
team
sports.
To
the
[1
player
tracking
in
where
players
cross-over
and
893
player
behavior
inare
continuous
team
sports.
To
model
theand
m
m
random
fields.
865
886
enforcing
aIn
planar
constraint
on
input
descriptor.
In
[15]
F.
roles
change
relatively
infrequent,
or
temporally
sparse,
893
901
882
m
[1]
S.
Belongie,
J.deformations,
Malik,
J.In
Puzicha.
Shape
ma
tems,
32(1):4
RW
899
are
occluded.
RB
tems,
3F
LW
897
909
884
899
RB
In879
terms
of
formation
analysis,
we
are
interested
inm we
and
Y.
Sheikh
870
Driven
Crowd
terms
of
formation
analysis,
we
are
interested
in
883
m
864
902
ST
lowing
non-rigid
which
is4(4),
more
ind
885
ited.
Internationa
tection
using
3D
Geometric
ECCV,
2012.
6the
References
885
marea
904
local
player
behaviors
a The
team
use
906
CM
903
892
muse
operate.
This
is
aformation,
future
fertile
ofthe
research
as
se
are
occluded.
ing,
2012.
local
player
behaviors
within
awithin
team
formation,
we
the
local
player
behaviors
within
aPrimitives.
team
formation,
we
use
[14]
Proz
tiv42
ofon
aSTsingle
soccer
team.
results
are
shown
in
Figure
??.
866
m
894 terms
887
902
883
mwe
w
m
894
8.
Summary
and
Future
Work
object
recognition
using
shape
contexts.
PAMI,
of
multi-agent
tracking,
we
implicitly
assume
that
and
based
this
show
that
this
can
enhance
900
ST
[16
[4]
S.An
Intil
In
this
paper,
we
proposed
aassumption
method
of
tracking
sports
885
871
910
[4]
S.
Intille
and
which
positions
switch
with
each
other.
For
basketball
(top
Team
Behavio
903
898
LCM
900
which
positions
switch
with
each
other.
For
basketball
(top
LCM
RW
[17]
L.
H.
Rose.
T
[12]
K.sports.
Murty.
alm
886
865
player
behavior
in
continuous
team
To
884
m
m
m
905
[3]
I.
J.
Cox
and
M.
L.
Miller.
On
finding
ranked
assignments
886
References
Au
LW
supervised
or
unsupervised
learning
good
and
bad
pl
904
player
role
as
our
representation
and
we
show
a
novel
and
880
[20]
Y.
Zhang,
W.
Ge
player
role
as
our
representation
and
we
show
a
novel
and
907
893
W
m
m
w
player
role
as
our
representation
and
we
show
a
novel
and
m
m
867
[20]
903
888
884895 roles
[15]
F. Ra
Inrelatively
terms
ofinfrequent,
formation
analysis,
we
areseeinterested
in
[1]
S.that
Belongie,
J.where
Malik,
and
J. Puzicha.
Shape
matching
Multi-A
895
522,
2002.
5temporally
901
886
m
References
872
row),
we
can
seesoccer
....
.contextual
For
soccer
(bottom
row),
we
change
are
or
sparse,
player
tracking
in
situations
players
cross-over
and
904
911
players
using
the
formation
asw
our
feature.
Multi-Agent
order
increasi
[13]
w(bottom
wofand
887team
[11]
I.Learning
Matthews
an
mmulti-target
m
1w
8.
Summary
and
Future
Work
899
with
application
to
tracking
and 2004.
motion
corre901
906
row),
we
can
see
that
....
.this
For
row),
we
see
RCM
M
RCM
local
player
behaviors
within
aPAMI,
team
formation,
w
866
[16]
M
for
Eve
ST
can
occur,
well
as
efficient
and
effective
retrieval
wh
885
RW
905
In
paper,
we
proposed
ahigher-level
method
of
tracking
sports
accurate
method
of
assigning
this
higher-level
feature
by
References
887
m
m
868
accurate
method
of
assigning
this
feature
by
904
908
881
accurate
method
ofJ.as
assigning
this
higher-level
feature
by
894
w
m
w
m
8.
Summary
and
Future
Work
object
recognition
using
shape
contexts.
24(4):509–
885896
889
[1
[5](top
S.2,
Kha
tive
902
887
which
positions
switch
with
each
other.
For
basketball
896
[17
8.
and
Future
Work
[1]
S.
Belongie,
J.Summary
Malik,
Puzicha.
Shape
matching
and
873
that....In
of
pre-game
planning,
such
aAerospace
tool
or
met905
[2]
P.and
Carr,
Y.
Sheikh,
and
I.
Matthews.
Monocular
687,
1968.
5,
912
888
and
based
on
this
assumption
we
show
that
this
can
enhance
are
occluded.
m
wterms
mm
We
depart
from
previous
formation
work
by
al[5]
S.
Khan
and
DrO
spondence.
Transactions
on
and
Electronic
Sysmm
ited.
Internatio
907
will
greatly
help
coaches
and
commentators
alike
due
[18]
STATS
SportsV
M
[1]
S.
Belongie,
J.the
Malik,
and
J. on
Puzicha.
Shape
matching
and
player
as
our
representation
and
we
show
ato
n8
ST
enforcing
awplayers
planar
constraint
on
the
input
descriptor.
In
900
902
867
that....In
terms
of
pre-game
planning,
such
aconstraint
tool
or
metusing
team
formation
as
our
contextual
feature.
CM
CM
869
905
522,
2002.
5role
886
enforcing
atracking
planar
on
the
input
descriptor.
InA.
886897
Rigidit
888
909
enforcing
a
planar
constraint
the
input
descriptor.
In
895
882 906
m
m
890
888
903
In
this
paper,
we
proposed
a
method
of
tracking
sports
Auto
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
897
906
874
ric
can
give
an
objective
measure
of
how
a
team
tends
to
row),
we
can
see
that
....
.
For
soccer
(bottom
row),
we
see
[13]
S.
Pellegrini,
tection
using
3D
Geometric
Primitives.
In
ECCV,
[1]
S.
Belongie,
J.
Malik,
and
J.
Puzicha.
Shape
matching
and
889
913
[17]
L.
tems,
32(1):486–489,
1995.
2required
w
w
908
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
Rigidity
ofKita
Ff
player
tracking
inmeasure
situations
cross-over
and
heavy
burden
currently
for
such
analysis.
lowing
non-rigid
deformations,
which
iswhere
more
indicative
of
terms
of
multi-agent
tracking,
we
implicitly
that
[12]
K.
Murty.
An
PG
906
In
this
we
atotracking
method
ofm
tracking
sports
870
907
We
depart
from
previous
formation
tracking
work
by
alMpaper,
accurate
method
of
assigning
this
higher-level
Mproposed
In
this
paper,
we
proposed
method
of
sports
[19]
C.
Sutton
and
868
[2]
Carr,
Y.
Sheikh,
and
I.assume
Monocular
Object
De901
903
ric
can
give
an
objective
of
how
aaplayers
team
tends
887
[6]
K.
LW
terms
multi-agent
tracking,
we
implicitly
assume
that
LW
889
887
891
904
players
using
the
team
formation
as
our
contextual
feature.
910
896
terms
of
multi-agent
tracking,
we
implicitly
assume
that
522,
2002.
5P.aof
w
m
w
[18
883
889
Never
Walk
Alo
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
907
Summary
and
Future
Work
875
This
is
future
fertile
area
ofMatthews.
research
as semi898
[1
890 operate.
[16]
M.
20
that....In
ofusing
pre-game
planning,
such
a[6]
tool
or
met898 8.roles
[3]
I.
J.3D
Cox
and
M.
L.
Miller.
On
finding
ranked
asR
522,
2002.
5players
914
909
[14]
[4]terms
S.
Intille
and
A.
Bobick.
A
Framework
for
Recognizing
LB
change
are
relatively
infrequent,
or
temporally
sparse,
w
mm
K.
Kitani,
B.
SG
907
using
the
team
formation
as
our
contextual
feature.
871
order
of
increa
are
occluded.
tection
Geometric
Primitives.
In
ECCV,
2012.
6
908
lowing
non-rigid
deformations,
which
is
more
indicative
of
player
behavior
in
continuous
team
sports.
To
model
the
Forecas
888
W
enforcing
a
planar
constraint
on
the
input
descr
random
fields.
References
M
890
869
players
using
the
team
formation
as
our
contextual
feature.
We
depart
from
previous
formation
tracking
work
by
alroles
change
are
relatively
infrequent,
sparse,
905884 This
522,
2002.
5
Target
Tracking.
904
902
operate.
isRW
asupervised
future
fertile
area
of
research
as
semi892
908
[19
911
876
897
[2]
P.P. Carr,
Y.
Sheikh,
and
I.Matthews.
Matthews.
Monocular
Object
Deor
unsupervised
learning
good
bador
plays
RW
w tends
wto
888
roles
change
are
relatively
or
temporally
sparse,
891
899
890
[18]
mm
Driv
with
application
toand
multi-target
tracking
and
mo
Carr,
Y.
Sheikh,
and
I.infrequent,
Monocular
Object
Deric
can
give
objective
measure
of
how
atemporally
team
899 909
915
910
Multi-Agent
Action
from
Visual
Evidence.
In
AAAI,
1999.
2ST
mm
and
based [2]
on
this
assumption
we
show
that
this
can
enhance
908
We
depart
from
previous
formation
tracking
work
by
alLCB
872
SF
Forecasting.
[7]
R.[15]
Li
a5
[3]
I.3D
J.an
Cox
and
M.
L.
Miller.
finding
ranked
assignments
1968.
2,
player
behavior
in
continuous
team
sports.
To
model
the
889
891
lowing
non-rigid
deformations,
which
is
more
indicative
of
ing,
4(4),
2012
local
player
behaviors
within
alearning
team
formation,
we
use
the
906
W
terms
of
multi-agent
tracking,
we
implicitly
ass
Prozone.
www.p
[2]
P.
Carr,
Y.
Sheikh,
and
I.On
Matthews.
Monocular
Object
De909
and
based
on
this
assumption
we
show
that
this
can
enhance
893
In
this
paper,
we
proposed
aformation
method
of
tracking
sports
870
tection
using
Geometric
Primitives.
In
ECCV,
2012.
6687,
M
C
877
can
occur,
as
well
efficient
and
effective
retrieval
which
We
depart
from
previous
tracking
work
by
al892
905
supervised
or
unsupervised
and
bad
plays
903
[19]
C.
912
898
m
ST
tection
using
3Das
Geometric
Primitives.
In
ECCV,
2012.
6[14]
900
and
based
on
this
assumption
we
show
that
this
can
enhance
885 player
889
ST
spondence.
Transactions
on
Aerospace
and
Elec
891900
911
[1
[17]
L.
H
916
operate.
is
agood
future
fertile
area
of
research
as
semitracking
in
situations
where
players
cross-over
and
[5]J.This
S.
Khan
and
M.
Shah.
Detecting
Group
Activities
Using
909
lowing
non-rigid
deformations,
which
is
more
indicative
of
aand
Spati
with
application
to
multi-target
tracking
and
motion
corre873
RLB
PF
[1]
S.
Belongie,
Malik,
and
J.situations
Puzicha.
Shape
matching
and
890role910
local
player
behaviors
within
ateam
team
formation,
we
use
the
892
player
behavior
in
continuous
sports.
To
model
the
tection
using
3D
Geometric
Primitives.
In[7]
ECCV,
2012.
6W.
Li
and
R.
907
[15]
F.R.
Ramos
H.
[13]
S.
Pellegrini,
A
910
roles
change
are
relatively
infrequent,
or
temporal
ran
player
tracking
inas
where
players
cross-over
and
894
893using
878
Y.
Zhang,
will
greatly
help
coaches
and
commentators
alike
due
to2[20]
the
player
as
our
representation
and
we
show
afinding
novel
and
[3]
I.
J.
and
M.
Miller.
On
finding
ranked
assignments
871
players
the
team
formation
our
contextual
feature.
W
[3]
I.R
J. Cox
Cox
and
M.
L.L.
Miller.
On
assignments
lowing
non-rigid
deformations,
which
is
more
indicative
of
913
901
899
can
occur,
as
well
as
efficient
and
effective
retrieval
which
906
904
32(1):486–489,
1995.
player
tracking
in
situations
where
players
cross-over
and
886 References
912
890
PG
917
Rigidity
oftems,
Formation.
In
ACM
Multimedia,
2005.
2plays
892
910
2004
player
behavior
in
continuous
team
sports.
To
model
theing
are
spondence.
Transactions
onranked
Aerospace
and
Electronic
Sys-J.within
[1
[8]
Liu,
P
874
supervised
or
unsupervised
learning
good
bad
Coccluded.
RB
891901 911
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
893application
local
player
behaviors
within
atracking
team
formation,
we
use
the
tive
robots
player
role
as
our
representation
and
we
show
aand
novel
and
[3]
I.to
J.
Cox
andon
M.
L.
Miller.
On
finding
ranked
assignments
911
908
a
Spatio-Tem
[20
Never
Walk
A
894
879
heavy
burden
currently
required
for
such
analysis.
with
multi-target
tracking
and
motion
corre895
are
occluded.
with
application
to
multi-target
and
motion
correand
based
this
assumption
we
show
that
this
can
Learning
for
E
accurate
method
of
assigning
this
higher-level
feature
by
872
902
depart
from
previous
formation
tracking
work
alWand
913R911
914
900
m
malike
m
player
behavior
in
continuous
team
sports.
To
model
the
tems,
32(1):486–489,
1995.
2toBobick.
will
greatly
help
coaches
and
commentators
due
the
907
areSummary
occluded.
905
local
player
behaviors
within
ashow
team
formation,
we
use
the
with
co
887 We
[4]
S.
Intille
A.higher-level
Aw
Framework
for
Re
875
891
8.
and
Future
Work
893
SG
[18]
STA
Autonomous
Age
with
to
multi-target
tracking
and
motion
correplayer
role
as
our
representation
and
we
aby
novel
and
894
LCM
892
can
occur,
as
well
as
efficient
and
effective
retrieval
which
902
[16]
M
522,
2002.
5method
912
912
accurate
of
assigning
this
feature
by
[20]
Y.
909
895Belongie,
spondence.
Transactions
on
and
Electronic
Sysspondence.
Transactions
onAerospace
Aerospace
and
Electronic
Sys880
S.
J.915
Malik,
and
J.application
Puzicha.
Shape
matching
and
GK
Target
Trackin
896
[8]
J.
Liu,
P.
Carr
player
in
situations
where
players
cross
mm
W
914
873
903
enforcing
a[1]
planar
constraint
on
the
input
descriptor.
Infrom
2In
912
player
role
astracking
our
representation
andVisual
we
show
athe
novel
and
[4]
S.for
Intille
and
A.
Bobick.
AF
Framework
for
Recognizing
901
m
Multi-Agent
Action
Evidence.
AAA
lowing
non-rigid
deformations,
which
is
more
indicative
of
local
player
behaviors
within
a
team
formation,
we
use
876
References
spondence.
Transactions
on
Aerospace
and
Electronic
Sys888
heavy
burden
currently
required
such
analysis.
908
906
accurate
method
of
assigning
this
higher-level
feature
by
895
[16]
M.
Rodriguez,
J
893903
Le
CM
8.
Summary
and
Future
Work
RCM
913
892
910
tems,
32(1):486–489,
1995.
2
894
will
greatly
help
coaches
and
commentators
alike
due
to
the
913
enforcing
a
planar
constraint
on
the
input
descriptor.
In
[19]
C.
S
M
SF
896
tems,Y.32(1):486–489,
1995.
2
881 usingare
[2]
P. 904
Carr,
Sheikh,
and
I.
Matthews.
Monocular
Object
DeLB
object
recognition
shape
contexts.
PAMI,
24(4):509–
897
with
context[14]
Prozone.
www.
occluded.
915
References
Multi-Agent
Action
from
Visual
Evidence.
In
AAAI,
1999.
2
[9]
J.
Liu
a
913
accurate
method
of
assigning
this
higher-level
feature
by
In
this
paper,
we
proposed
a
method
of
tracking
sports
tems,
32(1):486–489,
1995.
2
874
terms
of
multi-agent
tracking,
we
implicitly
assume
that
877
916
902
Driven
Crowd
A
m
w
w
enforcing
a
planar
constraint
on
the
input
descriptor.
In
896
[5]
S.
Khan
and
M.
Shah.
Detecting
Group
Activi
894
player
behavior
in
continuous
team
sports.
To
model
the
914
889 914[1]
player
role
as
representation
and
we
show
a assume
novel
and
CM tracking,
909
911
907
GK
[4]
S.
Intille
and
A.
APrimitives.
Framework
for
Recognizing
897
References
terms
ofour
we
implicitly
[2
[17]
rand
RCM
Mmulti-agent
893
882
heavy
burden
currently
required
for
such
analysis.
904
wDetecting
m
[4]
S.
Intille
and
A.
Bobick.
A
Framework
for
Recognizing
895
LCB
S.this
Belongie,
J.we
Malik,
and
J.Bobick.
Puzicha.
Shape
matching
and
PF
tection
3D
Geometric
In
ECCV,
2012.
6that
898
2002.
5using
tial
pat
916
In522,
paper,
proposed
method
tracking
sports
914
enforcing
atracking,
planar
constraint
on
the
descriptor.
In
[15]
F.
and
[5]
S.
Khan
and
M.
Shah.
Group
Activities
Using
905
[4]
S.aformation
Intille
andof
A.
Bobick.
A
Framework
for
Recognizing
878
[17]
L.2Ramos
H.
Rose.
ThH
terms
of
multi-agent
we
implicitly
assume
that
players
using
the
team
as
our
contextual
feature.
897
875
917
895
Rigidity
of
Formation.
In
ACM
2005
903
915
w
roles
change
are
relatively
infrequent,
or
temporally
sparse,
Multi-Agent
Action
from
Visual
Evidence.
AAAI,
1999.
2Multimedia,
912
898
LW
890 local
player
behaviors
within
aofas
team
formation,
we
use
the
mm
Finput
W
accurate
method
of
assigning
this
higher-level
feature
915
roles
change
are
relatively
infrequent,
temporally
sparse,
M
Multi-Agent
Action
from
Evidence.
In
AAAI,
1999.
2 by
[1]
S.
Belongie,
J.Visual
Malik,
and
J.Inor
Puzicha.
Shape
matching
and
883
object
using
shape
contexts.
PAMI,
24(4):509–
908
ing,
LB
W
mfeature.
m
RCB
905
894
LB
899
CM
896
[10]
P.that
Luce
917
C910
Rigidity
Formation.
In
ACM
Multimedia,
2005.
players
using
the
team
formation
our
contextual
915
terms
of
multi-agent
tracking,
we
implicitly
[3]
I.
J.recognition
Cox
and
M.
L.
Miller.
On
finding
ranked
assignments
Action
from
Visual
Evidence.
AAAI,
2
[9]Intive
J.2assume
Liu
and
Y.
[2]
P.
Carr,
Y.
Sheikh,
and
I.Multi-Agent
Matthews.
Monocular
Object
Derobots
wit
2004.
11999.
906
879
roles
change
are
relatively
or
temporally
sparse,
[1]
S.
Belongie,
J.previous
Malik,
and
J.infrequent,
Puzicha.
Shape
matching
and
916
896891 916
References
We
depart
from
formation
tracking
work
by
al[5]
S.a
Khan
and
M.
Shah.
Detecting
Group
Activities
Using
876
899
904
913
RW
and
based
on
this
assumption
we
show
that
this
can
enhance
522,
2002.
5898
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
884
and
based
on
this
assumption
we
show
that
this
can
enhance
role
as
representation
and
we
show
aand
novel
and
[5]
S.
Khan
and
M.
Shah.
Detecting
Group
Activities
Using
enforcing
planar
constraint
on
the
input
descriptor.
mtracking
mGroup
w[18]
RB
W
and
909
Mour
[20]
Y.Y.AZ
900
906 player
LCB
916
roles
change
are
relatively
infrequent,
or
temporally
sparse,
We
depart
from
previous
formation
tracking
work
by
al895911
LCB
897
[5]
S.
Khan
and
M.
Shah.
Detecting
Activities
Using
W
with
application
toW
multi-target
motion
correSTATS
SportsVU
tialIn
patterns.
and
based
on
this
assumption
we
show
that
this
enhance
tection
using
3D
Geometric
Primitives.
In
ECCV,
2012.
6[18]
880
Autonomous
899
907
917
Rigidity
of
Formation.
In5shape
ACM
Multimedia,
2005.
2can
897
object
recognition
using
contexts.
PAMI,
24(4):509–
w
m
900
914
lowing
non-rigid
deformations,
which
is
more
indicative
of
877
905
522,
2002.
ST
[2]
P.
Carr,
Y.
Sheikh,
and
I.
Matthews.
Monocular
Object
De885
917
player
tracking
in
situations
where
players
cross-over
and
Rigidity
of
Formation.
In
ACM
Multimedia,
2005.
2
player
tracking
in
situations
where
players
cross-over
and
B
892 accurate
LCM
wwhich
m
m Shape
[19]
W
method
oftracking
assigning
this
higher-level
by
901 non-rigid
terms
of
multi-agent
tracking,
we
implicitly
assume
that
917
and based
on
this
assumption
wefeature
show
that
this
can
enhance
Rigidity
of
In ACM
Multimedia,
2005.
2 Team
Lear
lowing
deformations,
isFormation.
more
indicative
of
912
910
907
[1]
S.
Belongie,
J.players
Malik,
and
J.w
Puzicha.
M
[19]
C.P.
Sutton
and
A.
player
in
situations
where
cross-over
RCB
900
Transactions
on
Aerospace
and
Electronic
Sys881
896
RLB
898
W
898
908
[10]
Lucey,
A.
RW
901
[16]
M.
Rodriguez,
522,
2002.
5L.
[3]
I.
J.spondence.
Cox
and
finding
ranked
assignments
915
8.
Summary
and
Future
Work
9ma
m
mand
tection
using
3DM.
Geometric
In
ECCV,
2012.
6 To
886
player
behavior
in
continuous
team
model
the
878
906
[2]
P.Miller.
Carr,
Y.On
Sheikh,
and
I.model
Matthews.
Monocular
Object
are
occluded.
[11]PAMI,
I.DeMatth
RCM
R Primitives.
W
mcontexts.
m
are
occluded.
player
tracking
situations
where
players
cross-over
and
893 enforcing
player
behavior
in
continuous
team
sports.
Tosports.
the shape
902
random
fields.
F2
aW
planar
constraint
on
the
input
descriptor.
In
are
occluded.
object
recognition
using
roles
change
are
relatively
infrequent,
or
temporally
sparse,
901
882
911
32(1):486–489,
1995.
2inranked
W
899
902
RB908
897913
909
899
RB
Y.Crowd
Sheik
916
Driven
with
application
to
multi-target
tracking
and
motion
corre[3]879
I.[2]
J.tems,
Cox
and
M.
L.
Miller.
On
finding
assignments
P.
Carr,
Y.
Sheikh,
I. Matthews.
Monocular
Object
DeW
mthe Work
m
887
ited.
tection
using
3D
Geometric
Primitives.
Inuse
ECCV,
2012.
6w In4
9and
8.
Summary
and
Future
local
player
behaviors
within
a2002.
team
formation,
we
the
907
CM
are
occluded.
ing,
2012.
local
player
behaviors
within
aand
team
formation,
we
use
w
m
m4(4),
903
894
902this
522,
5motion
Fimplicitly
W
883
of
tracking,
we
assume
that
9 Rose.
903
and
based
on
assumption
we
show
that
this
enhance
900
909 terms
In
this
paper,
we
proposed
amm
method
of
tracking
sports
914
912
917
with
application
to
multi-target
tracking
and
corre910
Behavi
[4]
S.multi-agent
Intille
and
A.
Bobick.
A
Framework
foracan
Recognizing
898
LCM
900
spondence.
Transactions
onand
Aerospace
and
Electronic
SysLCM
W
M
mMur
888
tection
using
3D
Geometric
Primitives.
In
ECCV,
2012.
6Y.Team
[17]
L.
H.
T
[12]
K.W.
w
m
w
m
[3]
I.
J.
Cox
and
M.
L.
Miller.
On
finding
ranked
assignments
LW
908
player
role
as
our
representation
and
we
show
novel
and
880
[20]
Zhang,
G
player
role
as
our
representation
we
show
a
novel
and
References
w
w
904
[20]
903
884
References
895
904team
901910
[2]
P.this
Carr,
Y.
Sheikh,
andIn
I.sparse,
Monocular
In2contextual
paper,
we
proposed
aMatthews.
method
of
tracking
spo
References
roles
change
are
relatively
infrequent,
orfinding
temporally
player
tracking
in
situations
where
players
cross-over
and
911
players
using
the
formation
as
our
feature.
order
oaO
Multi-Agent
Action
from
Visual
Evidence.
AAAI,
1999.
2
913915
889
M
[11]
I.
Matthews
tems,
32(1):486–489,
1995.
2004.
1
899
with
application
to
multi-target
tracking
and
motion
corre901
RCM
RCM
[3]
I.
J.
Cox
and
M.
L.
Miller.
On
ranked
assignments
m
w
m
Learning
for
Eve
RW
accurate
method
of
assigning
this
higher-level
feature
by
References
W
accurate
method
of assigning
this
higher-level
feature
by In ECCV,
904 909 mm
881
mmm
905
Summ
nd
Fu
uJ.Puzicha.
W
kenhance
885
905
902911
tection
using
3D
Geometric
Primitives.
896 and
players
using
team
as
ourElectronic
contextual
featu
W
687,
19
912
[1]
S.formation
Belongie,
J.on
Malik,
and
Shape
and Internat
890
based
on
this
assumption
we
show
that
this
can
are
occluded.
[1]
S.
Belongie,
Malik,
and
J.the
Shape
matching
and
We
depart
from
previous
tracking
work
by
alspondence.
Transactions
onformation
Aerospace
Sysited.
914
[5]
S.Malik,
and
M.
Shah.
Detecting
Group
Activities
Using
with
application
toJ.
multi-target
tracking
and
correm
mPuzicha.
[18]
STATS
9 matching
wandtracking
mSportsV
m
ST
enforcing
aKhan
planar
constraint
the
input
descriptor.
Inmotion
[4]
S.
and
A.
Bobick.
A
Framework
for
Recognizing
900916
902
CM
CM
905
906
910
886
906
enforcing
a
planar
constraint
on
the
input
descriptor.
In
[1]
S.897
Belongie,
J.Intille
and
J.
Puzicha.
Shape
matching
and
882
903
We
depart
from
previous
formation
work
by
[13]
S.
Pelle
[1]
S.
Belongie,
J.
Malik,
and
J.
Puzicha.
Shape
matching
and
[3]
I.
J.
Cox
and
M.
L.
Miller.
On
finding
ranked
as
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
m
891
913
tems,
32(1):486–489,
1995.
2AAAI,
912
M
MSutton
player
tracking
in
situations
where
players
cross-over
and
lowing
non-rigid
deformations,
which
is
more
indicative
ofthat
spondence.
Transactions
on
Aerospace
and
Electronic
SysRigidity
of
Formation.
In
ACM
Multimedia,
2005.
2m
W
m
mIn
w
terms
of
multi-agent
tracking,
we
implicitly
assume
[12]
K.
Murty.
PG
915
906
Multi-Agent
Action
from
Visual
1999.
2PAMI,
w Am
w
[19]
C.
and
901917
903
887
907
907
LW
m
mapplication
w
m
LW
911
object
using
shape
contexts.
PAMI,
24(4):509–
904
terms
of
multi-agent
tracking,
we
implicitly
assume
that
883
non-rigid
deformations,
which
is
more
indicative
Never
object
recognition
using
shape
contexts.
24(4):509–
Summ
nd
Fu
uorEvidence.
W
kwm
522,
2002.
5lowing
522,
2002.
with
to
multi-target
tracking
and
mo
m
892
898 recognition
914
[4]
S.infrequent,
Intille
andm
A.
Bobick.
A
Framework
Recognizing
m for
913
roles
change
are
relatively
temporally
sparse,
tems,
32(1):486–489,
1995.
2To
w
mGroup
m
SG
907
mm
mteam
order
of
incr
are
occluded.
player
behavior
inKhan
continuous
team
sports.
the
888
random
fields.
m
916
908
References
m
mmodel
m
905
[5]
S.
and
Detecting
Activities
Using
522,
Target
902904
player
behavior
inm
continuous
sports.
To
model
522,
2002.
5908
912
RW
[2]RW
P.M.
Carr,
Y. Sheikh,
Sheikh,
and
I.5I.Matthews.
Monocular
Object
roles
change
are
relatively
infrequent,
or
temporally
sparse,
W
m 2002.
spondence.
Transactions
on
Aerospace
and
Elec
[2]
P.
Carr,
Y.
and
Matthews.
Monocular
Object
De893Shah.
899 and884
915
Multi-Agent
Action
from
Visual
Evidence.
InDeAAAI,
1999.
2
based
onSF
this
assumption
we
show
that
this
can
enhance
908
W
m
m
914
m
m687,
M
M
909
1968.
[4] S.
Intille
and
A.
Bobick.
A
Framework
for
4(4),
2012
local
behaviors
within
a team
formation,
we
use
the
909
906
Prozon
[2]wm
P.3D
Carr,
Y.Multimedia,
Sheikh,
I.2005.
Matthews.
Monocular
Object
Dem
wteam
w6ing,
using
Geometric
Primitives.
In
ECCV,
2012.
local
player
behaviors
within
aRecognizing
formation,
we
use2,
917
mand
Rigidity
of
Formation.
In
ACM
2can
Wand
m
w
m
m
913
903905
tection
using
3D
Geometric
Primitives.
In
ECCV,
2012.
6[14]
tems,
32(1):486–489,
1995.
2enhance
[2]
P.player
Carr,889
Y.
Sheikh,
I.tection
Matthews.
Monocular
Object
DeST
894
and
based
on
this
assumption
we
show
that
this
885
ST
916
900
9
player
tracking
in
situations
where
players
cross-over
and
[5]
S.
Khan
and
M.
Shah.
Detecting
Group
Activities
Using
909
m
m
PF[3] I. J. Cox
m
m
[1]
Belongie,
J.Action
Malik,
and
J.m
Puzicha.
Shape
matching
910 S.Multi-Agent
w
m
m
wshow
890
from
Visual
Evidence.
In AAAI,
1999.
tection
using
3D
Geometric
Primitives.
Inm
ECCV,
6Ram
907915
[15]
F.m
[13]
S.2and
Pellegrini,
910
role
as
our
representation
and
we
a novel
andplayer
M.
L.
Miller.
On
finding
ranked
assignments
m
w2012.
w
[20]
Y.
Zhang,
W.Rea
m
player
role
as
our
representation
and
we
show
a
novel
and
m
m
895
References
914
w
m
w
m
[3]
I.
J.
Cox
and
M.
L.
Miller.
On
finding
ranked
assignments
tection
using
3D
Geometric
Primitives.
In
ECCV,
2012.
6
904906
player
tracking
in
situations
where
players
cross-over
and
[4]
S.
Intille
and
A.
Bobick.
A
Framework
for
886
PG
917
Rigidity
of
Formation.
In
ACMOn
Multimedia,
2 tive m
910
m2005.assignments
are occluded.
901 891
911 object
CKhan
wfinding
w
recognition
using
shape
PAMI,
24(4):509–
rob
[3] Shah.
I.
J. Cox
and
M.contexts.
L.of
Miller.
ranked
with
application
to
multi-target
tracking
and
motion
corre908916
911
accurate
method
assigning
this
higher-level
feature
Never
Walk
[5]
S.
and
M.
Detecting
Group
Activities
Using
m
m
Learning
for
E
896
with
application
to
multi-target
tracking
and
motion
correaccurate
method
of
assigning
this
higher-level
feature
by
915 On
W
mm
massignments
w and
Multi-Agent
Action
from
Visual
Evidence.
In AAA
mand
mm correare
occluded.
[3]
I. 902
J. Cox
and
M.
L.8.
Miller.
finding
ranked
905907
911
m
887
912
Summary
and
Future
Work
SG
Autono
with
toShape
multi-target
tracking
motion
892
spondence.
Transactions
Aerospace
Electronic
Sys522,
2002.
5Formation.
909
9w
enforcing
a on
planar
constraint
descriptor.
912
917 [1]
Rigidity
ofMalik,
In ACM
Multimedia,
2005.
2andinput
S. Belongie,
J.
and
J.application
matching
Target
Tracki
spondence.
Transactions
on
andon
Electronic
Sys897
mthe
w
enforcing
a893
planar
on
the
input
descriptor.
In
916
912
wconstraint
mS.Puzicha.
wAerospace
m
[5]
Khan
M.
Shah.
Detecting
Group
Activi
913 8.
with
toSummary
multi-target
tracking
and
motion
correw
m
m
w
References
spondence.
Transactions
on
Aerospace
andmElectronic
Systems,
32(1):486–489,
1995.
2and
888
906908
[16]
M.
Rod
and
Future
Work
910
903application
SF
m
terms
multi-agent
tracking,
weObject
implicitly
assume
t
913
[2]
P. Carr,
Y. this
Sheikh,
and
I.ofMatthews.
Monocular
Detems,
32(1):486–489,
1995.
2 PAMI,
object
recognition
using
shape
contexts.
24(4):509–
898
[14]
Prozone.
www
w
R
913
m
In
paper,
we
proposed
a
method
of
tracking
sports
tems,
32(1):486–489,
1995.
2
914
917
Rigidity
of
Formation.
In
ACM
Multimedia,
2005
terms
of
multi-agent
tracking,
we
implicitly
assume
that
Driven
[4]
S.
Intille
and
A.
Bobick.
A
Framework
for
Recognizing
894
m
m
889
w
w
mIntille
m
wIn infrequent,
m for Recognizing
911
907909
Mchange
References
roles
areArelatively
or temporally
914
S. Belongie,
J.
Malik,
J.
Puzicha.
Shape
matching
and
PF904 [1]
tection
Geometric
ECCV,
2012.
6mRamos spa
8993Dand
[4]this
S.Multi-Agent
and
A.
Bobick.
Framework
522,
2002.
5using
Inplayers
paper,
we
proposed
aPrimitives.
method
of as
tracking
sports
914
[15]forF.
andR
915
Action
from
Visual
Evidence.
In
AAAI,
2Recognizing
[4]
S.
Intille
and
A. Bobick.
Aour
Framework
[17]
H.
using
the
team
formation
contextual
feature.
roles
change895
are
relatively
infrequent,
or
912
9canL.enha
m
m
m
m
mw1999.
890
wtemporally
msparse,
m
and
based
on
this
assumption
we
show
that
this
m
M
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
908910
M
915
Multi-Agent
Action
from
Visual
Evidence.
In
AAAI,
1999.
2
900
C 905 [2]
players
using
the
team
formation
as
our
contextual
feature.
915
m
m
m
[3]
I.
J.
Cox
and
M.
L.
Miller.
On
finding
ranked
assignments
Multi-Agent
Action
from
Visual
Evidence.
In
AAAI,
1999.
2
916
P.
Carr,
Y.
Sheikh,
and
I.
Matthews.
Monocular
Object
Detive
robots
9m players
2004.
[5] depart
S. Khan
and M.previous
Shah.J. Detecting
Activities
Using
[1]
S. We
Belongie,
Malik,
and
Shape
matching
and
896891
from
formation
by
al- w1m
913
mPuzicha.
w Group
wtracking
player
tracking
inenhance
situations
where
cross-over
a
and
based
assumption
weJ.show
that
this
can
522,
2002.
5 S. Khan
m
M2alMwork
w
w
916
901
909911
[5]
and
M.
Shah.
Detecting
Group
Activities
Using
906 on this
9163D
mMUsing
We
depart
from
previous
formation
tracking
work
by
917 with
Rigidity
of
In ACM
Multimedia,
2005.
[5]
S.
Khan
and
M.
Shah.
Detecting
Group
application
toFormation.
multi-target
tracking
and
motion
corre[18]
STATS
tection
using
Geometric
Primitives.
In
ECCV,
2012.
6 Activities
Autonomous
897
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
m
914tracking
R
lowing
non-rigid
deformations,
which
more
of
are
occluded.
[2]
P.
Y.
Sheikh,
and
I. Matthews.
Monocular
Object
Deplayer
inCarr,
situations
where
players
cross-over
892
917
Rigidity
Formation.
In
ACMisw
Multimedia,
2005.ofindicative
2 2005.
m 902
m
m of
wand
mis
m
917
m
Rigidity
of
Formation.
In
ACM
Multimedia,
2 C. Sutt
lowing
non-rigid
deformations,
which
more
indicative
910912907
[19]
spondence.
Transactions
on
Aerospace
and
Electronic
Sys898 I.tection
[16]
M.
Rodrigue
522,
2002.
5
[3]
J.and
Cox
and
M.
L.
Miller.
On
finding
ranked
assignments
915
8.
Summary
Future
Work
using
3D
Geometric
Primitives.
In
ECCV,
2012.
6
player
behavior
in continuous
teamTosports.
To model
the
M
M
903
are
893 tems,
player
behavior
in continuous
team sports.
model
the
random
m
913
M
911occluded.
32(1):486–489,
1995.
2
916908 899
9
Driven
Crow
application
toMiller.
multi-target
tracking
and
motion
corre[3] with
I.[2]
J. Cox
and M.
L.
On finding
ranked aassignments
P. Carr,
Y.904
Sheikh,
and
I. Matthews.
Monocular
Object
References
local
behaviors
team
formation,
we use the
ing, 4(4
local
player
behaviors
within
awithin
teamwformation,
use
themDe894 we
mIntille
mplayer
msports
m
w
wwe
In909
this 900
paper,
proposed
a method
of
tracking
912914
9
917
withS.application
to
multi-target
and
motion
corre[4]
and
A.
Bobick.
A
Framework
for
Recognizing
spondence.
on tracking
Aerospace
and
Electronic
Systection
using
3D
Geometric
Primitives.
InM
ECCV,
2012.
[17]a6novel
L. [20]
H. and
Rose.
905
M show
RTransactions
player
role
as
our
representation
and
show
Y. Zha
player
role
as
our
representation
and J.we
awe
novel
and
[1]
Belongie,
and
J.AAAI,
Puzicha.
Shape
901895
References
R and
m
wMalik,
players
the
team
formation
as
our
feature.
Action
from
Visual
Evidence.
In
2matching
913915910using
tems,
1995.
2S. contextual
2004.
1Learnin
906
[3]Multi-Agent
I.32(1):486–489,
J.accurate
Cox
M.of
L.assigning
Miller.
On
finding
ranked
assignments
accurate
method
this
higher-level
feature
by1999.
R
method
of
assigning
this
higher-level
feature
byM
object
recognition
using
shape
contexts.
PAMI, 24(4):5
902896
M
M
911 [4]
m
w
w
We
depart
from
previous
formation
tracking
work
by
al907
914916
[5]
S.
Khan
and
M.
Shah.
Detecting
Group
Activities
Using
with
application
to
multi-target
tracking
and
motion
[18]
STATS Sport
M
9 descriptor.
enforcing
aA.planar
constraint
onmatching
the
descriptor.
IncorreS. Intille
and
A
Framework
forand
Recognizing
522,
2002.
5 input
enforcing
a planar
constraint
on
the
input
In
[1] S. Belongie,
and
J. Bobick.
Puzicha.
Shape
903897J. Malik,
M
m
m
M
912non-rigid
908
w
lowing
deformations,
which
is
more
ofElectronic
spondence.
Transactions
on
Aerospace
and
SysRigidity
of
Formation.
InCarr,
ACM
Multimedia,
2005.
22Monocular
terms
of Action
multi-agent
tracking,
we
implicitly
that
915917
Multi-Agent
from
Visual
Evidence.
In
AAAI,
1999.
[19]
C. Sutton
and
[2]
P.
Y.indicative
Sheikh,
and
I.assume
Matthews.
Object
object
recognition
using
shape
contexts.
PAMI,
24(4):509–
904898
terms
of
multi-agent
tracking,
we
implicitly
M sparse,
M assume that
909
913
change
are
relatively
infrequent,
or Geometric
temporally
tems,
32(1):486–489,
1995.
2using
tection
3D
Primitives.
Inrandom
ECCV, 2012.
player
behavior
inroles
continuous
team
sports.
To
model
the
field
916 522,
R
n
905 S.
[5]
Khan
and
M.
Shah.
Detecting
Group
Activities
Using
2002.
5
change
are relatively
infrequent,
sparse,
Mtemporally
m
MM
Mor
899 and roles
910
M
based
on
this
assumption
we
show
that
this
can
enhance
914
[3]
I.
J.
Cox
and
M.
L.
Miller.
On
finding
ranked
assignme
[4]
S.
Intille
and
A.
Bobick.
A
Framework
for
Recognizing
4(4), 201
local
behaviors
a on
team
formation,
we
use
thethat
M
917 P. player
Rigidity and
ofwithin
Formation.
In ACM
Multimedia,
2005.
2 Mthis caning,
[2]
Carr,906
Y.900
Sheikh,
and
I.
Matthews.
Monocular
Object
Debased
this
assumption
we
show
enhance
911 in
9
player
tracking
situations
where
players
cross-over
and
M
with
application
to
multi-target
tracking
and
motion
co
M Visual
mY.
915
Multi-Agent
Action
Evidence.
In AAAI, [20]
1999.
2 Zhang, W
player
role907
as our
representation
andfrom
novel
Mshow
Rare
n 912
M
MM aTransactions
tection
using
3D
Geometric
Primitives.
In
ECCV,
2012.
player
tracking
inwe
situations
where6and
players
and S
M cross-over
M M and Electronic
occluded.
spondence.
M Activities
Mon Aerospace
908901
916 method
[5] S. Khan and
M.
Shah.
Detectingfeature
Group
Using
Learning
accurate
this
higher-level
by
913On
tems,
32(1):486–489,
1995.
2
m for
are
occluded.
[3]
I. J. Cox
and M.ofL.assigning
Miller.
finding
ranked
assignments
9
917 909902 Rigidity of914
Formation.
In
ACM
Multimedia,
2005.
2
M
m
enforcing
a planar toconstraint
ontracking
the
descriptor.
In A Framework
M
m
[4] input
S. and
Intille
and A. correBobick.
for Recogniz
with application
multi-target
motion
910903 References
M
w Visual
M
O
D
915 Y
Multi-Agent
Action from
terms of multi-agent
tracking,
we implicitly
assume
thatM M Evidence. In AAAI, 199
911904
[1] References
S. Belongie,
and
J. Puzicha.
and
DS.
GKhan
mand M.Shape
m matching
916 J. Malik,
[5] or
Shah.
Detecting
Group M
Activities
Us
m
M
roles change
are relatively
temporally
sparse,
M
912905
9
object infrequent,
recognition
using
shape
contexts.
PAMI,
24(4):509–
M
917
Rigidity
of
Formation.
In
ACM
Multimedia,
2005.
2
M
M
O
m
M
M
9
[1] 2002.
S. Belongie,
Malik,
Puzicha. Shape matching
and
M
and based 913
on906
this assumption
we
thisand
canJ.enhance
522,
5 showJ.that
M
M
wobject recogniti m m
mMonocular
M m
914
[2]
P. Carr,
Y.where
Sheikh, players
and I. Matthews.
Object DeM
M
player tracking
in
situations
cross-over
and
907
M Primitives. In ECCV, 2012. 6
m
915
tection using 3DMGeometric
are occluded.
908
m
xx1010
55
44
33
22
11
00
−1−1
−2−2
−3−3
−4−4
−5−5
(a)
(b)
F gure 1 In ou ecen pape 6 we showed ha when eams p ay a home hey have mo e possess on n he o wa d h d wh ch co e a es w h
a s gn can y h ghe numbe o sho s goa s and w ns In h s pape we d g deepe and ook a he o ma ons eams use n home and away
ma ches by ana yz ng an en e season o p aye ack ng da a om P ozone a In he season we ana yzed we ound he same h ng w h home
eams hav ng a mo e possess on n he o wa d h d shown a e he d e ence maps w h eams a ack ng e o gh w h ed s gn y ng mo e
891
possess on b By ana yz ng o ma ons we wan ed o see
h s behav o was caused by eams p ay ng d905
e en o ma ons e g 2 s ke s
906
892
907
893 om
ns ead o 1 op vs m dd e o was he o ma on p ayed d e en y – e g eams p ess ng h ghe up he e d bo
908
894
909
895
910
896
To exp ore f here was any s ra eg c exp ana on n a recen paper we nves ga ed h s prob em by ana yz ng an en re season s
wor h of ba rack ng da a 6 We found ha even hough here was no d fference n shoo ng or pass ng prof c ency home eams
had s gn f can y more sho s and goa s Mos no ab y we found ha home eams had s gn f can y more possess on n he forward
h rd compared o away eams wh ch sugges s ha away eams p ay more conserva ve y han home eams Th s backs up ano her
common y he d be ef ha eams shou d a m o “w n he r home games and draw he r away ones” W h he a d of a who e season
of p ayer and ba rack ng da a from Prozone 3 from a op- er profess ona soccer eague we re nves ga e h s phenomenon by
m
m
f rs see ng f he same home advan age ex s ed across seasons F gure 1 a and f so exp ore whe her References
was causedMbyA AA mm MV A m mMw
911
897
912
898
913
899
914
900
915
901
916
902
917
903
904
905
906
907
908
909
910
911
912
913
914
915
916
916
917909
910
[3] I. J. Cox and M. L. Miller. On finding ranked assignments
with application to multi-target tracking and motion corre-
M
M
911
K J. Puzicha.
M Shape
D m andG
m matching
[1] S. Belongie, J. Malik, and
912
m mPAMI, 24(4):509–
MM m
object recognitionMusing shape contexts.
913
M
522, 2002. 5
914
m
m Monocular
M Object
M
[2] P. Carr, Y. Sheikh, and I. Matthews.
De915
M
m
tection using 3D Geometric Primitives. In ECCV, 2012. 6
916
M
1 eams p ay ng d fferen forma ons a away games e do managers e ec o p ay on y one s r ker away ra her han wo or
2 eams p ay ng he forma on d fferen y e he forma on s he same bu hey us p ayed more defens ve y
917
[3] I. J. Cox and M. L. Miller. On finding ranked assignments
917
m
MM
with application to multi-target tracking and motion corre-
As each game has more han a m on da a po n s e a 22 p ayers a 10 frames-per-second hav ng an exper human abe
forma ons across an en re season s proh b ve and so requ res an au oma c so u on Even hough forma on ana ys s s
performed oday n a qua a ve way by an exper human ana ys e g zona mark ng 7 no me hod ex s s wh ch can au oma ca y
represen and de ec forma ons of a eam n soccer The b gges prob em s dea ng w h he cons an swapp ng of p ayer ro es ha
occurs w h n a forma on as h s n roduces no se n o he s gna mak ng compar sons d ff cu Th s prob em hough a so prov des
a c ue o a so u on as he dynam c swapp ng of p ayer pos ons sugges s ha our me hod a so needs o dynam ca y upda e over
me Leverag ng an Expec a on-Max m za on EM me hod we can de ec and v sua ze forma ons from p ayer rack ng da a
Sec on 3 Us ng h s me hod we d scover many un que charac er s cs ha eams emp oy dur ng a season Sec on 4 Add ona y
we show ha our me hod can be used o dynam ca y v sua ze and summar ze he game n a very mean ngfu and qu ck way
Sec on 5 Before we showcase h s work we f rs re-exp ore he “home advan age”
2 Re-Exploring the Home Advantage
In our prev ous work 6 we ana yzed an en re season of ba - rack ng da a from he Eng sh Prem er League da a from Op a and
we found eams had approx ma e y he same number of passes pass ng accuracy and shoo ng accuracy when p ay ng a home and
away However here were s gn f can y more sho s and goa s for home eams and we found ha h s co nc ded w h home eams
hav ng more possess on n he forward h rd Before proceed ng we wan ed o see f h s same phenomenon occurred across a
season of da a from a d fferen da a source To exp ore forma ons and he way p ayers move dur ng games we used an en re
season of p ayer rack ng and ba even da a from Prozone 3
2014 Research Paper Competition
Presented by:
9
Am 9
m
U
Event
Statistic
Mean for
Home Team
Mean for
Away Team
P-Value
1.61 per match
1.10 per match
<0.001
Goals
1.57 per match
1.21 per match
<0.001
Shots on target
6.48 per match
5.21 per match
<0.001
Shots not on target
8.94 per match
6.97 per match
<0.001
Shooting accuracy
41.7%
42.4%
0.5046
Passes
451 per match
436 per match
0.1483
Occupancy in final third
(of time in play)
14.1% per
match
11.8% per
match
<0.001
Team x-centroid
-1.79 m
-4.90 m
<0.001
Team y-centroid
0.064 m
-0.047 m
0.4397
Points
(win = 3pts, draw = 1pt, loss = 0pts)
(a)
(b)
Figure 2. (a) Table showing that the home advantage exists on the dataset that we worked on for this paper, obtaining similar results to our paper
in [6]. (b) Using the player tracking data, we show the mean team centroid (top) and team spread (bottom) and can see that the home teams had a
mean position which was significantly more forward than when they played away (highlighted by the second last entry in the table).
The dataset consisted of 20 teams who played every team home and away and the data consisted of the location of every player at
10 frames-per-second as well as the ball events. Due to the proprietary nature of the data, we unfortunately can not disclose the
identity of the league or the teams. To anonymize the identity of each team, we randomly assigned each team a unique identifier “AT”. The results of the initial analysis is shown in Figure 2. As can be seen in Figure 2(a), the same pattern exists in the dataset that
we used for this work. Most notably, home teams had more shots and goals, as well as more points and possession in the forward
third (Figure 1(a)), while the shooting proficiency and number of passes was roughly the same. Additionally in Figure 2(b), we show
that home teams are more advanced than away teams in terms of their mean position, while the position across the field is the same.
This is a rather interesting finding as it suggest something is different in the way the teams play: does it mean that teams play with
two strikers at home while they play with an extra midfielder away, or is this a global trend of the formation where all the players
pushed forward? To do this we first have to detect and visualize the formations.
3 Automatically Detecting Formations from Tracking Data
Coordination between players within a team is essential for a team’s success, in terms of providing coverage of the large field area,
and in scoring and defending against their opposition from scoring. To achieve this, the coach/manager designates a formation or
overall structure of play to coordinate the team's movements and each player is assigned a particular responsibility or role within the
formation. For example, in a 4-3-3 formation, the roles may be defined as R = {goal-keeper, left-back, right-back, left center back,
right center back, left center midfield, right center midfield, left wing, right wing, attacking center midfield, striker}, which represent
the players’ positions relative to one another (i.e., the left wing, by definition, is positioned to the left relative to other players) and
their general location on the field. As a formation reflects the relative spatial arrangement of players and not so much with the
absolute position of the player, we normalized for
translation (i.e. removed the centroid of position of
the formation so that they are centered on the halfway line - goal-keeper excluded) in each frame. We
then characterized the formation over a match, by
calculating the mean and covariances of the role
positions about the centroid as shown in Figure 3.
Due to the dynamic nature of soccer, players
frequently swap roles throughout a match, making
the representation and automatic detection of a
formation challenging. For example, using a static
ordering of players throughout a match (e.g.
ordering players by their starting roles or identity),
the frequent role swaps throughout a match cause
(a)
(b)
Figure 3. (a) The positions of the players at every frame across a match relative to
the centroid and ordered by player identity. (b) The formation can be represented
using the mean and covariances of the player positions about the team centroid
(drawn relative to the field-center).
2014 Research Paper Competition
Presented by:
G705 T1, Initial Role Positions
G705 T1, Iteration 15, diff = 0.00m
5
3
4
3
5
8
9 6 10
2
4
G750 T1, Iteration 8, diff = 0.01m
G590 T1, Iteration 18, diff = 0.00m
8
1
1
10
6
2
3
5 8
6 9 7 10
3
4
7
7
9
2
2
G490 T1, Initial Role Positions
1
1
5 9 10
3
6
4
8
1
G590 T1, Initial Role Positions
7
7
1
2
G750 T1, Initial Role Positions
8
5
6
9
3
4
10
6
4
G490 T1, Iteration 11, diff = 0.01m
9
8
7
1
7
5
2
7
2 65
9
8 4
10
3
1
10
2
6
5
3
4
9
10
8
Figure 4. (Top) Shows the mean positions of the players over a half of a game - notice the heavy overlapping that occurs between the midfielders
and the forwards. (Bottom) Using our approach we can accurately assign player roles at each frame which disambiguates role and allows formation
analysis to occur.
great overlap in the player distributions, as shown in the top row of Figure 4, and hence, a static ordering of players does not
accurately represent the structure or formation of the team. To overcome this, we use a dynamic ordering of players by the role that
they occupy at a given instant in time. However, a big challenge in achieving this is that we don’t know what formation a team is
playing in advance, so we can not assign roles. We solve the two tasks of formation discovery and role assignment simultaneously
using an Expectation-Maximization approach on the player tracking data, where we first assign initial player role based on the mean
positions of the players. We then use the Hungarian algorithm [9] to update the role of each player in every frame and calculate the
error between the previous and updated assignments which is similar to [8]. We then continue to iterate this process until
convergence. This process is shown at the bottom of Figure 4.
Given that we have disentangled the player positions and generated a dynamic role for each player at each frame, we can use these
labels to represent the mean formations for each match for every team. The formation of each of the 20 teams (A-T) for every
Figure 5. Mean formation for each match half relative to the centroid (center of each rectangle) overlaid over one another for each team
(A-T). Red represents home games and blue represents away games, with all teams normalized to attack from left to right.
2014 Research Paper Competition
Presented by:
match are visualized in Figure 5, where the red corresponds to their home
performances and blue corresponds to away. As can be seen in this figure, most of
the teams tend to play the same formation with only a slight variation occurring in
some of the positions. For example, only teams B and T seem to have some
variation across the course of a season, while others like teams A, F, P and R only
have a minor change in the midfield (i.e. player 1 holding midfielder vs 2, or player
with one striker vs two). Other than that, most teams tend to be rather staunch in
what they play regardless of whether they are playing at home or away. The most
dominant formation appears to be a 4-4-2, with some teams varying the midfield as
described above. Only one team appeared to dabble with a back three (team T) and
this variation is quite obvious but it was consistent for both home and away Figure 6. To get a closer look at the formation
differences, we conducted analysis on a
performances
zoomed in area of the field.
4 Formation Style Analysis
In the previous section we showed that each team played pretty much the same formation regardless whether they were at home or
away. This begs the question then, how did they play the formation? Were they more attacking, or were they more defensive? To answer these
(a) Mean position of the formation when they were in possession
(b) Mean position of the formation when the opposition were in possession
Figure 7. Plots of the mean position of the team when they were in (a) possession, and (b) when the opposition were in possession. The red
ellipses refer to their behavior when they were at home and the blue ellipses when they played away.
2014 Research Paper Competition
Presented by:
Figure 8. Plots showing the distance ran for each team when they played at home (red) compared to when they played away (blue). (Left) Overall,
across the league we can see that the away teams (blue-line) covered more distance than the home team (red-line). (Middle) While in attack, home
teams tended to run more than away teams, mostly because they are exploiting more area, but the discrepancy between the two is small compared
to the amount of distance away teams had to cover in defense (right). As the amount of possession for both with similar, these findings are
significant.
questions, we looked at the centroid of each formation and looked at each team’s behavior when they were: a) in possession, and b)
when the opponent was in possession (we disregarded the times when the ball was out for a stoppage, which is amazingly around
50% of the time). These plots for the zoomed in area of the field (see Figure 6) are shown in Figure 7. As it is very obvious from
these plots, it can be seen that nearly all teams are more forward at home than they are away, both when they attack and when they
defend. This can help explain how teams have more possession in the forward third at home, as simply having the players in more
advanced positions suggest that they would have more possession in these advanced areas. Similarly, when they are defending
higher up the pitch, it means that they are more likely to regain possession higher up the pitch. The downside to this tactic, is that
they leave more space exposed behind the defensive line which allows teams to be hit on the counter attack. However, in addition
to winning the ball in more advanced positions (which is closer to the opponent’s goal, which will possibly lead to more shots and
more goals seeing the accuracy is roughly the same), the home team will expend less energy. The implications of this are large as the
less energy a team expends, the longer and more sustained attacks a team can lead. In Figure 8, we show the difference in distance
that teams run at home compared to away games (again red refers to home teams, and blue refers to away teams).
5 Match Summarization via Formation Visualization
Statistics given at half-time are quite poor as they only contain sparse statistics which are not symbolic of how the team played or
what the opposition are doing both in terms of formation and how they executing the formation (i.e. attacking or defending deep).
A powerful statistic which broadcasters use during a live-broadcast is to give the possession of both teams over the last 5 minutes
which gives a indication on which team is dominating or not. While this is very insightful, it does not give any information about
where this is happening. Using our formation detection approach, we can dynamically see how the match is played quickly over the
[G700 M350_2]
half by using a sliding window of 5 minutes to see what formation
they are playing and where it is relative to the other team. A filmWEHAM vs SHAMP
strip of our approach is shown in Figure 9.
4−1
0
0.75
1
1.25
3
10
2
8 1
1.5
1.75
2
Shot on target
18
6
9
5
Yellow card
0.5
10
2
210 95
3 6
47
4
7
Yellow card
0.25
3
8
Shot not on target
0
9
5
6
1
Shot not on target
7
4
−0.5
−1
5
9
Shot on target
3
6
Shot not on target
2
10
74
8
Frames 14800−15400
Attack: Red Team
2.25
2
3 10
4
2.5
18 7
4
3
910
5 965
6 2
7 8
1
Shot not on target
1
Frames 10290−10890
Attack: Blue Team
Shot on target
0.5
Shot on target
Shot on target
Neutral
Shot not on target
Frames 2612−3212
Shot on target
Shot not on target
1
2.75
4
10
Figure 9. By representing formations within a match using a sliding window, we can see how the game progresses in terms of formationsx and
location on the field. A 45 min half timeline is shown (circles = goals), and the utility of our formation and role representation is demonstrated
above, with the home team (red), attacking from left to right. We can see that during a neutral portion of the game, the teams are playing what
appears to be a 4-1-4-1 (red), and 4-2-3-1 (blue). Before the blue team scores, we can see a midfielder (role 9) moves forward to aid in the attack.
In the final example, the red team scores, with the whole team positioned close to the goal.
2014 Research Paper Competition
Presented by:
6 Summary and Future Work
In this paper, we presented a novel method which could automatically detect and visualize formations from soccer player tracking
data. Utilizing an entire season’s data from Prozone from a top-tier professional league, we used our automatic approach to see
whether the formation they played had anything to do with explaining why teams are more successful at home rather than away.
Our analysis showed that nearly all teams tend to play the same formation at home as they do away, however, the way they executed
the formation was significantly different. Specifically, we were able to show that at home, teams played significantly higher up the
field compared to when they played away (or conversely, teams sat much deeper at away games). This conservative approach at
away games suggests that coaches aim to win their home games and draw their away games. Additionally, we also show that our
method can visually summarize a game which gives an indication of dominance and tactics. While enabling new discoveries of team
behavior which can enhance analysis, it is also worth mentioning that our automatic formation detection method is the first to be
developed. In future work, we aim to use the team’s formation information as additional contextual features which will enhance
short-term prediction (i.e., see if we can predict the next pass), as well as using it to discover discriminative or unique patterns teams
use in offensive and defensive situations.
References
[1] J. Wilson, “Inverting the Pyramid: The History of Football Tactics”, Orion Books, 2008.
[2] Opta Sports. http://www.optasports.com
[3] Prozone. http://www.prozonesports.com
[4] C. Anderson and D. Sally, “The Numbers Game: Why Everything You Know About Soccer is Wrong”, Penguin Books, 2013.
[5] T. Moskowitz and L. Wertheim. Scorecasting: The Hidden Influences Behind How Sports Are Played and Games Are Won.
Crown Publishing Group, 2011.
[6] P. Lucey, D. Oliver, P. Carr, J. Roth and I. Matthews, “Assessing Strategy using Spatiotemporal Data”, in KDD, 2013.
[7] Zonalmarking. http://www.zonalmarking.net
[8] P. Lucey, A. Bialkowski, P. Carr, S. Morgan, I. Matthews and Y. Sheikh, “Representing and Discovering Adversarial Team
Behaviors using Player Roles”, in CVPR, 2013.
[9] H. W. Kuhn, “The Hungarian Method for the Assignment Problem”, in Naval Research Logistics Quarterly, vol. 2, no 1-2, pp.
83-97, 1955.
2014 Research Paper Competition
Presented by: