Temporally Accurate Events Detection Through Ball
Possessor Recognition in Soccer
Marc Peral
1 a
, Guillem Capellera
2 b
, Antonio Rubio
2 c
, Luis Ferraz
2 d
,
Francesc Moreno-Noguer
1 e
and Antonio Agudo
1 f
1
Institut de Rob
`
otica i Inform
`
atica Industrial CSIC-UPC, Barcelona, Spain
2
Kognia Sports Intelligence, Barcelona, Spain
Keywords:
Events Detection, Action Spotting, Sports Analytics.
Abstract:
Recognizing specific actions in soccer games has become an increasingly important research topic. One key
area focuses on accurately identifying when passes and receptions occur, as these are frequent actions in
games and critical for analysts reviewing match strategies. However, most current methods do not pinpoint
when these actions happen precisely enough and often fail to show which player is making the move. Our
new method uses video footage to detect passes and receptions and identifies which player is involved in each
action by following possession of the ball at each moment. We create video clips, or tubes, for every player
on the field, determine who has the ball, and use this information to recognize when these key actions take
place. Our results show that our system is better than the latest models in spotting passes and can identify
most events with an accuracy down to 0.6 seconds.
1 INTRODUCTION
Soccer has grown in popularity in recent years, as
the increase in the revenue of top clubs reflects (De-
loitte, 2023). This growth comes hand-in-hand with
the multiplication of data acquisition in terms of play-
ers and ball positional information (Capellera et al.,
2024a; Capellera et al., 2024b), video footage of
games and gathering of events statistics. The enor-
mous amount of collected data calls for ways to ex-
ploit its potential (Goes et al., 2021). Soccer clubs
have a team of analysts that study the behavior of
their team and the habits of their next opponent to
design a strategy for the upcoming games. To do
so, they devote a vast number of hours rewatching
recorded games. Understanding soccer semantics lets
them plan future tactics, but they have to stare at game
footage and there is not much time left for pondering
as schedules are tight. The industry has identified this
pain point in the sector and some companies emerged
a
https://orcid.org/0000-0001-6521-6476
b
https://orcid.org/0009-0006-7266-078X
c
https://orcid.org/0000-0002-6771-8645
d
https://orcid.org/0000-0001-7851-9193
e
https://orcid.org/0000-0002-8640-684X
f
https://orcid.org/0000-0001-6845-4998
to automate the process of spotting actions.
The latest technological improvements are achiev-
ing a pruning of the limitations of soccer analysis
in terms of time and subjectivity (Cossich et al.,
2023). A strong area of this automation is the de-
tection of events. Recent research (Giancola and
Ghanem, 2021; Zhou et al., 2021; Denize et al., 2024)
tries to spot high-level actions providing a big pic-
ture of the match to follow the flow of the game,
although that is not enough for analysts who need
to conduct a deep dive examination. To perform a
proper match review, they focus their attention on the
key events that give more details about team tactics,
such as passes and receptions. These are undoubt-
edly the most common events in a soccer game and
we call them touch events because they happen ev-
ery time a player touches the ball. That is why this
work focuses on finding which is the player in the
field that has the ball at every frame, if any, to af-
terwards retrieve who is executing those touch events
and when those occur. Taking into account these
touch events, we find approaches that use informa-
tion from previous events to define the next up (Yeung
et al., 2023; Simpson et al., 2022), methods employ-
ing trajectory data (Vidal-Codina et al., 2022; Kim
et al., 2023; Sanford et al., 2020), and others that work
with video recordings of matches (Sorano et al., 2021;
Peral, M., Capellera, G., Rubio, A., Ferraz, L., Moreno-Noguer, F. and Agudo, A.
Temporally Accurate Events Detection Through Ball Possessor Recognition in Soccer.
DOI: 10.5220/0013317700003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages
221-231
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
221

Figure 1: Events detection pipeline. Our approach is composed of three blocks: 1) video tube computation where feature
extraction is also performed to get logits per player and per frame. 2) Possession is chosen between the N players, if any,
resulting into a one-hot possession vector per frame. 3) The detection of events shows how a pass P and a reception R are
detected in this chunk of T frames.
Sanford et al., 2020; Baikulov, 2023; Philipp Singer,
2022). Some models correct manually annotated data
and achieve better synchronization (Biermann et al.,
2023). As a consequence, we observe that most of
the proposed approaches do not fulfill the temporal
accuracy requirements, and the ones that get closer
to it are missing some crucial details like identifying
which player from the ones in the pitch is the passer
or the receiver.
The approach we propose is the first to achieve
the detection of touch events and the identification of
which of the players in the field is conducting those
in a reasonable temporal precision for soccer analysts.
As illustrated in Figure 1, our method crops a video
tube (Yu and Yuan, 2015) for each player in the frame
and finds which one has ball possession. A tube cor-
responds to the visual information embedded inside
consecutive bounding boxes for the same player in
the video space. From per player possession infor-
mation, we spot when those key events take place,
additionally yielding which player in the field per-
formed them. We test our model in a large dataset of
matches from top European leagues. Despite no other
methods providing all the indispensable information,
we still compare to state-of-the-art methods that par-
tially meet our requirements and prove that ours out-
performs the latest. We bring a valuable contribution
to the state of the art as we enable the detection of
passes and receptions in a way that is useful for soc-
cer analysts.
2 RELATED WORK
Previous work approaches events detection trying to
assign a start and end time, but the state of the art
evolved to consider events occurring in a specific
frame avoiding ambiguity and subjectivity. As stated
in (Giancola et al., 2018), events are defined as in-
stantaneous in the soccer rulesbook. For example, a
goal happens at the moment the ball crosses the goal
line between the goalposts and the crossbar. For this
reason, we consider events as occurrences in a certain
frame.
Table 1 depicts how recent methods range in a
wide variety of inputs utilized to detect various types
of event. In this section we go over those approaches
elucidating their strengths and weaknesses. Despite
all this, one can notice that only our model is capa-
ble of spotting events precisely in time, even when
missing information, as well as identifying the player
performing those.
Data Observability: A common obstacle in current
soccer applications is the lack of data. Algorithms
that use previous events (Yeung et al., 2023; Simp-
son et al., 2022) need every previous action that hap-
pened and their position in the field to work. But the
frequent missing data problem appears when the in-
formation of a player is absent. It may be caused
by a failure in the tracking systems or because the
player falls out of the camera view (Guti
´
errez-P
´
erez
and Agudo, 2024a) for some seconds. Most vision-
based algorithms (Giancola and Ghanem, 2021; Zhou
et al., 2021; Denize et al., 2024; Sorano et al., 2021;
Sanford et al., 2020; Baikulov, 2023; Philipp Singer,
2022) found a way to disregard this lack of informa-
tion, even methods that use video tubes apply some
type of padding or ignoring strategy. But the ones
based on the players’ trajectories (Vidal-Codina et al.,
2022; Kim et al., 2023) are not robust to this lack of
data as they need full visibility of the arrangement of
the players on the pitch.
Sparse vs. Touch Events: Sparse events refer to soc-
cer actions scattered in time, like foul, corner, penalty,
goal, yellow or red card. These can help to give a
brief understanding of the course of the game but are
not enough for a soccer analyst to examine the match
in depth, check if formation decisions outperformed,
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
222

Table 1: Modalities of events detection methods. Inputs are depicted as: - previous events class and their position
in field, - players and ball trajectories, - broadcast video, - tactical camera video, - method uses video tubes.
Missing data column illustrates: - can work when missing players information either visual or positional, - needs dense
information of all players or events. Player identification is represented as: - outputs which player is performing the
event, - outputs in which region of the field the event happens but not the player, - does not output player information.
Paper Event type Input Missing data Player identification ↓ Window [s]
NetVLAD++ (Giancola and Ghanem, 2021) Sparse 5-60
Zhou et al. (Zhou et al., 2021) Sparse 5-60
Comedian (Denize et al., 2024) Sparse 1-5
NMSTPP (Yeung et al., 2023) Touch -
Seq2Event (Simpson et al., 2022) Touch -
Vidal-Codina et al. (Vidal-Codina et al., 2022) Touch 20
BallRadar (Kim et al., 2023) Touch 2
PassNet (Sorano et al., 2021) Touch 1-4
Sanford et al. (Sanford et al., 2020) Touch / 1.7
Baikulov (Baikulov, 2023) Touch 1
Singer et al. (Philipp Singer, 2022) Touch 0.15-0.7
Ours Touch 0.6
or recognize which players stood out. Sparse events
provide the context to interpret the wider information
that events like passes and receptions unravel. These
other events that provide sufficient details to go over
soccer matches are touch events.
Touch events occur when a player touches the
ball. These are passes and receptions that may later
be subdivided into types of passes such as crosses or
throw-ins. To glimpse the difference between sparse
and touch events notice that the average of events per
match in Premier League season 2020-2021 is 2.7 for
goals and 2.9 for yellow cards, while receptions and
passes appear 696 and 940 times, respectively, per
match.
SoccerNet-v2 dataset (Deliege et al., 2021) is one
of the largest available soccer events datasets, with
more than 500 match recordings from TV broadcast
and 17 types of events annotated. Although many
strategies (Giancola and Ghanem, 2021; Zhou et al.,
2021; Denize et al., 2024) had been designed to de-
tect actions, their task focuses only on sparse events
that fall short of what soccer analysts need. Soccer-
Net (Deliege et al., 2021) creators identified the need
to spot more fine-grained actions and released a new
challenge called Ball Action Spotting. As these new
actions are closer in time, one cannot use broadcast
video with changes in scenes, zooms, and replays.
For this reason, a new dataset was published (Deli
`
ege
et al., 2023) with continuous footage that keeps most
of the players on shot, which is called a tactical cam-
era. Our approach works on the mentioned tactical
camera videos to detect those touch events that pro-
vide the crucial data that soccer experts need.
Prediction with Past Events Sequences: Previous
research (Yeung et al., 2023; Simpson et al., 2022)
uses information from previous events to try to predict
the next up, which could be a good initial approxi-
mation of the problem. However, their results may
be distorted by the use of limited information. Their
methods are biased toward predicting shots because
of the high influence of the event positional informa-
tion they use on the type of events in the processed
sequence.
Detection with the Location of Players and the
Ball: Using trajectories data has a clear drawback,
the low availability of datasets. This lack of public
data relies on the high cost of its gathering, as it re-
quires expensive hardware and meticulous computa-
tion of homographies.
The majority of methods that employ trajectories
use a set of rules that rely on meticulous spatial in-
formation, and some of them need 3D coordinates. If
one wanted to obtain the requested data from a video
recording, one would need to perform a camera cali-
bration (Guti
´
errez-P
´
erez and Agudo, 2024b) delimit-
ing the field using the visible lines and then compute
a set of homographies per frame that would proba-
bly produce not accurate enough location informa-
tion. Today, the required level of accuracy is collected
by placing GPS sensors on players and the ball or
multiple cameras around the pitch, between 16 and 20
in every stadium (Liga, 2020), and doing so reaches
exorbitant prices.
We were not able to compare to methods employ-
ing coordinates from players, ball and/or events be-
cause their datasets not only are inaccessible but do
not have match recordings either. Apart from the high
cost for accessing the data, the inconvenience of this
practice is that they do not have enough time pre-
cision. State-of-the-art methods (Kim et al., 2023;
Vidal-Codina et al., 2022) spot actions using posi-
tional information in the field by first determining the
ball possessor, but their temporal accuracy is of 2 and
20 seconds, respectively. We know from soccer ex-
Temporally Accurate Events Detection Through Ball Possessor Recognition in Soccer
223

perts that less than a second of precision is needed for
touch events. Our model does not need real coordi-
nates from players, ball, or previous events and still
achieves finer temporal precision.
Spotting with Full Frame from Video Record-
ing: Recent works have switched to exploiting video
footage because images constitute a richer source of
information. (Sanford et al., 2020) conducted an ab-
lation for both trajectories and image-based solutions.
Although their acceptance window takes 1.7 seconds,
they show how most of their successful predictions
fall inside a closer range of 0.5 seconds. Despite these
models being able to detect in which frame events
happen with a reasonable temporal precision, the ob-
stacle for soccer analysts to use the information pro-
vided is that they only provide temporal information
about detected events, but give no clue of which is the
player performing those.
The winner of the SoccerNet Ball Action Spot-
ting challenge (Baikulov, 2023) used a transfer learn-
ing approach fine-tuned with a sampling strategy to
combat class imbalance. Both this and the win-
ner of the Bundesliga Data Shootout (Philipp Singer,
2022), were using grayscale neighbor frames stacked
in triplets as the color channels of an image be-
fore extracting the features with a 2DCNN. There
is no possibility of testing our method on the previ-
ously mentioned challenges data because they do not
contain bounding boxes for the players. Neverthe-
less, we still compare with the state of the art as we
tested (Baikulov, 2023) in our dataset matches, prov-
ing our superiority, as will be shown later.
Locating with Video Tubes: As (Yu and Yuan, 2015)
state, video tube proposals work well for dynamic ac-
tion recognition with moving cameras. (Honda et al.,
2022) use both trajectories information and video
tubes to predict who the receiver of a pass is. How-
ever, their dataset only considers successful passes in
situations with all players visible.
A Bundesliga Data Shootout contender (Ya-
mamoto, 2022) proved the adequacy of focusing on
the region where the ball is to spot events. They used a
transfer learning approach pretraining the feature ex-
traction with a ball detection task that upgraded their
final results. Nevertheless, as the other participants
in the challenge, they provide only temporal infor-
mation of predicted events. The same obstacle arises
with (Sanford et al., 2020). They prove employing
players bounding boxes improves over their baseline,
but they aggregate features from the tubes and do not
retrieve which player is preforming each action.
In (Sorano et al., 2021) an object detection mod-
ule that finds the bounding boxes for the players that
are closer to the ball is used, affirming that it makes a
significant contribution to their detections. They used
the information obtained from this detector and com-
bined it with features extracted from the whole video
frame. Hence, their visual information is deficient as
they shrink the image resolution and downsample the
framerate to 5 Hz. Although they are in a similar
task of detecting passes, we cannot access their pri-
vate dataset of four matches. Nonetheless, we test
their method in our dataset and prove their method
cannot find passes as ours does.
Overall, none of the aforementioned approaches is
capable of detecting touch events and the associated
player with accurate temporal precision. Our method
accomplishes this by (1) operating with visual infor-
mation from video footage, (2) focusing on ball pos-
session making use of video tubes and (3) applying a
set of rules to spot passes and receptions with a tem-
poral accuracy.
3 OUR APPROACH
In this section, we explain how our method finds the
events and their performers. Figure 1 shows the three-
stage pipeline we propose to sort out the problem,
with video data and players bounding boxes as input.
In the first block (in blue), a video tube for each visi-
ble player is cropped, extracting features and provid-
ing a score of the player being in possession of the
ball. In the second one (in green), all scores in a par-
ticular frame are combined to find the player in pos-
session of the ball, if any. Finally, in the last block (in
red), a set of rules is exploited in order to determine
precisely when touch events happen and which player
is performing them. Next, we introduce in depth each
of the blocks in our proposal.
3.1 Video Tubes Extraction
This first stage makes use of spatio-temporal video
tubes, which pay off in moving camera situations for
dynamic action identification (Yu and Yuan, 2015).
To generate the video tubes, we consider a collar of
T
f
frames, i.e., we collect one of the input bounding
boxes, cropping a region H × W × Ch in a specific
frame together with the corresponding one in the pre-
vious and next T
f
frames, obtaining a video tube of
2T
f
+1 frames, where H, W and Ch indicate the color
image resolution. To normalize the data, for each box
we enlarge the smallest rectangle side to make it a
square, then add 20% extra margin to include more vi-
sual context, and resize them to 128×128 pixels (see
Figure 2), obtaining the same size for all the players.
A clear way to identify a possessor is by finding
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
224

Figure 2: Video tubes extraction. Left: In every input
image, a bounding box extraction (in red) per player is ap-
plied. Two instances are displayed in the figure. Middle:
Zooming the selected player by means of a bounding box
in red. To normalize data, the square that results from en-
larging smallest rectangle side and the square after adding
extra margin are displayed in cyan and black, respectively.
Right: Final region to be cropped.
the ball in their bounding box region. We know that
our algorithm does that because some of the false pos-
itives we detect include players without the ball but
moving their legs forward when they have completely
white boots. After visually analyzing that effect, we
found that three frames is the time the ball takes to
get from the edge of the video tube to the player or
to leave the cropped region when a pass is performed.
Without loss of generality, the parameter T
f
is then set
to 3 frames because there should be enough frames to
appreciate the action that is being performed. This
observation is consistent with (Honda et al., 2022),
where they demonstrate that longer sequences lead to
a decrease in accuracy due to excess context informa-
tion.
3.2 Possession Likelihood from Video
Tubes
In this stage, we propose to obtain the possession like-
lihood from the video tubes extracted above. To this
end, we extract a feature vector independently for ev-
ery image of a player video tube using ResNet50 (He
et al., 2016), fine-tuned for this task and with a Tem-
poral Shift Module (TSM) (Sudhakaran et al., 2020).
We included this module because, as the authors man-
ifested, it shifts channels in time to highlight the nar-
row differences between consecutive frames. As we
use dense visual information without downsampling
the frame rate, the dissimilitude from a frame to the
next one is subtle. We take the penultimate layer out-
put of the ResNet (He et al., 2016) and concatenate
the features from all tube frames. With a fully con-
nected layer, we reduce the time dimension, resulting
in a 256-feature vector.
After finding the embedding for a video tube, an-
other dense layer breaks it down to two values, being
the possessor or not. From this classification problem
we derive the likelihood of a video tube following a
player in possession of the ball, which is necessary
to determine the actual possessor from all the play-
ers in the field and subsequently find the touch events
they may be performing. In training, we use Addi-
tive Angular Margin (AAM) loss over common Soft-
max. This was implemented first in deep face recog-
nition tasks (Deng et al., 2019) to enforce inter-class
diversity and to solve intra-class appearance varia-
tions, which also apply to our problem, as in the same
class we may find players with different positions or
clothes. The AAM loss function for the i-th sample
belonging to the y-th class can be written as:
L
f
= −log
e
s·cos
(
θ
y
i
+m
)
e
s·cos
(
θ
y
i
+m
)
+
∑
N
j=1, j̸=y
i
e
s·cosθ
j
, (1)
where x
i
∈ R
d
is the deep feature of the i-th sam-
ple and W ∈ R
d×N
indicates the weights, with d the
embedding feature dimension and N the number of
classes. θ
j
stands for the angle between the feature
x
i
and the weight W
j
. We set the value of the scaling
variable s to 1, where s is the radius of the hyper-
sphere in which the learned embedding features are
distributed. The parameter m, set to 0.5, represents
the margin penalty that will enhance the aforemen-
tioned intra-class and inter-class relations.
3.3 Per-Frame Possessor Identification
With possession likelihoods from isolated video
tubes, we move to a more genuine situation with all
players in the frame trying to point out which one
is in control of the ball, if any. At this stage, we
add a head to the network that chooses between the
only two feasible options, a single player is in posses-
sion of the ball or none of them. For a given frame
t, the new head takes as input the likelihoods of all
the players in a window frame from t − T
p
to t + T
p
,
i.e., the dimensionality of the input is (2T
p
+ 1) × N
dimensional, where T
p
is the collar of frames set to
provide temporal context and N the number of play-
ers. Checking the slope that possession likelihoods
show when changing from possession to not posses-
sion or vice versa, we observe that it takes five frames
to completely toggle, so we adopt two frames as the
value for our parameter T
p
. This is directly related
to the previously specified T
f
value (see Section 3.1),
and varying it would also vary the slope in possession
likelihoods and, therefore, the preferable value for T
p
.
When a player is not visible in any of the frames, we
add zero padding for their likelihood in that frame.
We smoothed these inputs using a Gaussian filter to
reduce the effect of noisy spurious detections.
After that, the possessor identification head uses
Conv-TasNet (Luo and Mesgarani, 2019), a convo-
lutional solution created for speech separation. This
Temporally Accurate Events Detection Through Ball Possessor Recognition in Soccer
225

Figure 3: Events estimation. The figure shows our model
output at the three stages and the corresponding ground
truth for a 160-frame chunk. As it can be seen, our esti-
mation (orange symbols) is very close to the reference one
(red symbols).
network uses 1D convolutions in the time domain
to separate the speakers from an audio input, a task
that uncovers clear similitude with ours at this stage,
where we want to find who is the possessor of the ball
among all players in the pitch. Still taking advantage
of the resemblance to speech recognition tasks, we
add a Time-Delay Neural Network (TDNN) (Waibel
et al., 2013) that will let our model scan the past and
future of possession scenarios in a time-shift invariant
way. At this point, we do an average pooling to end
up with only a value per player. Lastly, there are two
fully connected layers that get us to the output of this
stage. The first expands the embedding size to capture
the relations between all players. The second shrinks
back to a single hot vector of size N+1, that is, each
player that could be the possessor plus the negative
class if none of them are. Thanks to that, we have a
guess for who the possessor is in every frame, if any.
At this stage, we use a cross-entropy loss between the
prediction and the ground truth class when training.
3.4 Touch Events Detection
Touch events get their name from the requirement that
a player is in contact with the ball when performing
them. Knowing the possessor of the ball in every
frame, if any, we designed a logic that tags a recep-
tion whenever a player gains possession of the ball
and a pass when they are no longer the possessor. In
this logic, there are two parameters, T
s
and T
e
, that
help filtering out some false positives that may appear.
When a player touches the ball just once, annotators
tag a single-pass event in the ground truth instead of
annotating both a reception and a pass. To match this,
we make use of T
s
, set to seven frames, which skips
receptions for first-touch passes. The value of T
e
de-
termines the minimum required frames a possession
sequence must last to be considered. The majority of
false positives appear when the ball passes in front
of a player who is behind it some meters away. This
is one of the main challenges for models that use vi-
sual information only, but setting the value of T
e
to
three we discard a large part of them. In Figure 3 one
can observe the output of our approach at each of the
three stages for a chunk of 160 frames and how the fi-
nal predicted events result close in time to the ground
truth.
4 EXPERIMENTAL RESULTS
Our data comes from 25Hz tactical camera videos
of 36 matches (28 training, 4 validation, and 4 test)
from Spanish LaLiga (1st and 2nd division) and Ital-
ian SerieA season 2022-2023 with player bounding
boxes that we use as our input, and the touch events
in frames precision that become our ground truth.
As our full pipeline is composed of 3 blocks, we
propose to evaluate everyone of them, incorporating
some comparisons w.r.t. competing approaches when
possible. Particularly, we consider three tasks: (1) ob-
taining possession likelihoods of isolated video tubes,
(2) identifying the possessor, if any, from all players
present in a frame and (3) spotting the touch events
with a super-tight temporal precision.
4.1 Possession Likelihood from Video
Tubes
To create a first dataset of isolated video tubes, we
select every frame in which a touch event occurs and
crop a tube for each player. Non-visible players in
all the context frames or overlapping with the positive
box are discarded. For every positive tube we have
approximately 20 negatives, so the dataset is very im-
balanced, with 75,683 positives and 1,516,586 nega-
tives.
In this first task, we evaluate the model ability
to discern whether an isolated video tube follows
on a player in possession of the ball. To overcome
the imbalance, during training we force a 50% ra-
tio of positives in every batch, with the other half
randomly selected negatives. We drop the remain-
ing negative samples at the end of every epoch. To
prove that using video tubes instead of still images
makes sense, we train a Timeless model that only uses
features from the middle frame of the tubes. Next,
we set a Baseline without shift modules to corrob-
orate the convenience of TSM (Lin et al., 2019) or
GSM (Gate-Shift Module) (Sudhakaran et al., 2020).
As shown in Table 2, not considering the temporal
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
226

Table 2: Possession likelihood from video tubes analy-
sis. The table includes precision, recall and Area Under Re-
ceiver Operating Characteristic (AUROC) curve metrics. In
all cases, two possibilities by using ResNet18 and ResNet50
are provided.
Model ResNet ↑ Precision ↑ Recall ↑ AUROC
Timeless
18 73.12 92.80 99.10
50 77.58 92.87 99.14
Baseline
18 82.84 97.90 99.78
50 85.85 97.59 99.78
Baseline
+ TSM
18 80.73 97.67 99.76
50 86.29 98.01 99.81
Baseline
+ GSM
18 84.41 97.77 99.79
50 88.12 97.85 99.81
Table 3: Possessor identification evaluation. The table
reports purity and coverage on possessor sequences as well
as Accuracy per frame showing the use of Gaussian filter.
As it can be observed, our solution provides better solution
than the baseline.
Model Gaussian filter ↑ Purity ↑ Coverage ↑ Accuracy
Max +
Threshold
No 49.87 66.96 70.12
Yes 52.65 67.50 70.60
Ours w/o
CTN+TDNN
No 44.08 59.83 60.90
Yes 44.19 59.74 60.86
Ours with
CTN+TDNN
No 57.03 66.87 70.95
Yes 56.61 67.79 71.88
context achieves a minor performance according to
all metrics, while ResNet50 (He et al., 2016) always
outperforms ResNet18 (He et al., 2016). Moreover,
adding the shift modules to the feature extraction is
advantageous, as the solution always beats that pro-
vided by the baseline. TSM (Lin et al., 2019) is cho-
sen because, for our use case, a higher recall is pre-
ferred, as we will see later.
4.2 Per-Frame Possessor Identification
After the classification task, the model can retrieve the
likelihood of an isolated video tube following a player
in control of the ball. We now have to distinguish be-
tween the video tubes of all players in the frame and
find who is the actual ball possessor, if any. To evalu-
ate this task, we create a new dataset where the ground
truth possessor is derived from the annotated touch
events. A player is assigned as the possessor from the
first touch event until the last one they perform con-
secutively. With this we have a background or nega-
tive class found between the last event performed by
a player and the first event performed by the next one.
We reduce the number of matches because now, un-
like for the previous task, we can use all the frames
when the ball is in play, that is, when the game is not
stopped. The new dataset consists of 5 new matches:
3 for training, 1 for validation, and 1 for testing.
In the evaluation of this task, apart from the usual
per-frame accuracy, we introduce purity and cover-
age, metrics usually employed for segment-wise com-
parison in audio speaker change detection. The use of
standard metrics like precision and recall requires the
definition of a tolerance parameter that every author
fixes to set the maximum distance to be matched be-
tween boundaries and still will not show the dissim-
ilarity between segments. For that reason, we adopt
these metrics to suit this segment-wise comparison.
According to (Bredin, 2017), given R the set of ref-
erence possessor segments for our task and H the set
of hypothesized segments, coverage is defined as:
coverage(R , H ) =
∑
r∈R
max
h∈H
|
r ∩ h
|
∑
r∈R
|
r
|
, (2)
where
|
r
|
is the duration of segment r and r ∩ h
is the intersection of segments r and h. Purity is
computed analogously by interchanging R and H
in Equation (2). An over-segmented hypothesis with
too many possession changes implies high purity but
low coverage because possessor predictions cover a
low percentage of the ground truth. In contrast, an
under-segmented hypothesis implies a high coverage
but low purity.
We test the fully connected network with and
without ConvTasNet and TDNN (Waibel et al., 2013),
while also toggling the use of a Gaussian filter. We ab-
late them against a straightforward 2-step solution. It
first checks whether any player has a likelihood over
a given threshold, set to 0.5. If not, it chooses the
negative class; otherwise, it selects the player with
the maximum likelihood. Table 3 shows how this
straightforward solution is better than a simple neu-
ral network, but the incorporation of the audio sepa-
ration strategies outperforms in this analogous task of
finding the possessor. We also observe how the reg-
ularization from the Gaussian filter slightly upgrades
the outcome.
4.3 Touch Events Detection
As it was discussed in Section 2, datasets and source
code for event detection tasks are usually not avail-
able, as they are normally linked to private compa-
nies. A dataset (Deli
`
ege et al., 2023) for touch events
was available, but it did not contain information about
player bounding boxes in the image and, therefore,
we were unable to use it. We compare to the top-
performing model in that dataset, but contrast to ours
in a new set of matches that none of the methods
has already seen before. This new test set comprises
6 matches from LaLiga 2023-2024 with various sta-
dium sizes and weather conditions. To do this com-
Temporally Accurate Events Detection Through Ball Possessor Recognition in Soccer
227

Table 4: Events detection evaluation and comparison. The table reports the results separated by pass and reception.
Precision, Recall and F1 are shown for both 0.6- and 1-second windows. Mean Average Precision with 1 second acceptance
windows (mAP@1) is the metric used in the state of the art.
Model Event
0.6s 1s
↑ Prec ↑ Rec ↑ F1 ↑ Prec ↑ Rec ↑ F1 ↑ mAP@1
Baikulov (Baikulov, 2023)
Recep - - - - - - -
Pass 77.15 52.36 62.38 81.22 56.44 66.60 29.09
Ours
Recep 54.23 57.53 55.83 64.96 67.12 66.02 38.36
Pass 70.12 64.67 67.28 77.62 71.33 74.34 44.15
Figure 4: Four chunks from a test set match that show how our model fits the ground truth better than PassNet (Sorano
et al., 2021), even when using their criteria of non-instantaneous passes.
parison, we separate detections of passes from re-
ceptions, because (Baikulov, 2023) is trained to spot
passes and drives. We can contrast our pass predic-
tions to theirs, but receptions do not map to drives as
they are semantically different.
Table 4 shows how they obtain a better precision
and we achieve a higher recall in the 0.6- and 1s-
acceptance windows. According to soccer experts,
when treating passes as instantaneous events that hap-
pen in a specific frame, a model with high recall that
lets them filter out false positives is preferable. A
model with high precision and lower recall would
force them to go again through the match to find the
false negatives missed by the model, achieving no
reduction in the time they spent analyzing a match.
Apart from that, we also reach a higher F1 and larger
mean Average Precision for 1 second acceptance win-
dows (mAP@1), the latter being the metric they use
for their evaluation.
We also show a comparison with PassNet (Sorano
et al., 2021). Apart from only looking for passes and
not retrieving which is the player passing or receiving
the ball, the main difference between their model and
ours is that they consider passes to have a start frame
and an end one. Not accounting for passes as instanta-
neous events makes their metrics focus on how many
frames in the segments considered positives can the
model predict as frames from a pass. But they will
never be able to find how many instances of a real
pass occurred in a chunk of a game.
Their dataset is not available, but the code for their
method is, so we train it in 9 of our dataset matches
(225% times the data in their original train set) and
compare using the same set of 6 test matches that we
used in the previous comparison. We try to train their
algorithm on instantaneous events to use our metrics
for comparison, but PassNet (Sorano et al., 2021) only
learns to put every frame to negative due to the imbal-
ance in frames. Therefore, we train it by considering
their criteria of each pass having a start and end frame
for each pass, and with this we are able to replicate
very similar results to the ones they show in their pa-
per using their metrics. Having PassNet (Sorano et al.,
2021) trained, we try to use our metrics by defining
the start frame of each pass as the frame in which the
event happens, but they barely get to a 19% precision
and 21% recall, so comparison here is nonsense.
For being able to compare we have to redefine our
network output considering the frames after a pass de-
tection as positives and the ones after a reception as
negatives, building this way predictions that follow
their non-instantaneous definition of a pass. With this
twist, we can compare their domain using their met-
rics. As seen in Table 5, we outperform PassNet (So-
rano et al., 2021) in all their metrics except the num-
ber of positive frames detected as positive, that is, the
recall of frames contained in passes. Figure 4 shows
how their model has a high recall on passes because
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
228

Table 5: Pass detection comparison. The table exposes the Precision, Recall and F1 metrics for pass detection, by considering
passes as segments with a start frame and an end one. Therefore each frame is part of a pass (Pass) or a negative frame (No
Pass).
Model
Pass No Pass
↑ Prec ↑ Rec ↑ F1 ↑ Prec ↑ Rec ↑ F1
PassNet (Sorano et al., 2021) 43.13 72.24 54.01 70.57 41.22 52.04
Ours 69.95 52.74 60.14 74.46 86.10 79.86
it just learned to predict positive in high-action re-
gions. For that reason, they have a low precision in
passes and a low recall on the negative class. From
their output, it would be impossible to detect when the
passes happen or even just the number of instances of
passes that appear. Our model, despite being trained
on instantaneous passes, is capable of capturing the
dynamics that real passes show in the ground truth.
4.4 Extended Model Output
Figure 3 shows a simple example of two players mak-
ing two successive passes to clarify the model out-
put at every stage. We extend the model outputs for
a longer 550 frames (22s) chunk in Figure 5. The
chunk starts with a pass (specifically, a throw-in) per-
formed by the player P3 and follows a possession of
the players in the same team. Some players are clearly
not in possession of the ball during this chunk, so
by removing those we focus only on relevant play-
ers. We specifically selected a chunk where some of
our model errors are represented. We can observe in
Figure 5(b) how we wrongly detected a reception be-
fore a pass event where the ground truth was a single
first-touch pass. In Figure 5(a) there is a false positive
that was not filtered out, but in Figure 5(c) a spurious
detection for the same player was correctly discarded.
A false negative appears in Figure 5(d), where a pass
was missed because our model considered the detec-
tion too spiky.
5 CONCLUSIONS
In this paper, we have proposed a method that pre-
cisely detects touch events in soccer from videos. Un-
like most of the recent work, our approach is capable
of retrieving which player from all players in the pitch
is the one performing each action while delivering at
the needed temporal accuracy. Despite low data avail-
ability, we compare to two state-of-the-art models that
partially fit our task requirements and prove our supe-
rior robustness in the detections even when using their
metrics. In the future, we plan to follow the intricate
duty of making an end-to-end version of our whole
Figure 5: Model outputs and limitations. The graph
shows the model outputs focused on relevant players for a
chunk of 550 frames. A false positive can be observed in
(a), a correctly filtered spurious detection in (c) and a false
negative missed in (d). In (b) a first touch pass was detected
as a multiple contact reception-pass. See legend in Figure 3.
model, as at the moment it is trained at every stage
independently.
ACKNOWLEDGMENT
This work has been supported by the project
GRAVATAR PID2023-151184OB-I00 funded by
MCIU/AEI/10.13039/501100011033 and by ERDF,
UE and by the Government of Catalonia under 2020
DI 00106.
REFERENCES
Baikulov, R. (2023). Winning solution for soccernet ball
action spotting challenge 2023. https://github.com/l
Romul/ball-action-spotting.
Biermann, H., Komitova, R., Raabe, D., M
¨
uller-Budack, E.,
Ewerth, R., and Memmert, D. (2023). Synchroniza-
tion of passes in event and spatiotemporal soccer data.
Scientific Reports, 13(1):15878.
Bredin, H. (2017). Tristounet: triplet loss for speaker turn
embedding. In ICASSP, pages 5430–5434.
Capellera, G., Ferraz, L., Rubio, A., Agudo, A., and
Moreno-Noguer, F. (2024a). Footbots: A transformer-
Temporally Accurate Events Detection Through Ball Possessor Recognition in Soccer
229

based architecture for motion prediction in soccer. In
ICIP, pages 2313–2319.
Capellera, G., Ferraz, L., Rubio, A., Agudo, A., and
Moreno-Noguer, F. (2024b). Transportmer: A holis-
tic approach to trajectory understanding in multi-agent
sports. In ACCV.
Cioppa, A., Giancola, S., Somers, V., Magera, F., Zhou,
X., Mkhallati, H., Deli
`
ege, A., Held, J., Hinojosa, C.,
Mansourian, A. M., et al. (2023). Soccernet 2023
challenges results. arXiv preprint arXiv:2309.06006.
Cossich, V. R., Carlgren, D., Holash, R. J., and Katz, L.
(2023). Technological breakthroughs in sport: Current
practice and future potential of artificial intelligence,
virtual reality, augmented reality, and modern data vi-
sualization in performance analysis. Applied Sciences,
13(23):12965.
Deliege, A., Cioppa, A., Giancola, S., Seikavandi, M. J.,
Dueholm, J. V., Nasrollahi, K., Ghanem, B., Moes-
lund, T. B., and Van Droogenbroeck, M. (2021).
Soccernet-v2: A dataset and benchmarks for holistic
understanding of broadcast soccer videos. In CVPR,
pages 4508–4519.
Deli
`
ege, A., Cioppa, A., Giancola, S., Seikavandi, M. J.,
Dueholm, J. V., Nasrollahi, K., Ghanem, B., Moes-
lund, T. B., and Droogenbroeck, M. V. (2023). Ball
action data and labels for soccernet ball action spot-
ting challenge. https://www.soccer-net.org/data#h.yk
gf675j127d.
Deloitte (2023). Annual review of football finance 2023.
Deng, J., Guo, J., Xue, N., and Zafeiriou, S. (2019). Ar-
cface: Additive angular margin loss for deep face
recognition. In CVPR, pages 4690–4699.
Denize, J., Liashuha, M., Rabarisoa, J., Orcesi, A., and
H
´
erault, R. (2024). Comedian: Self-supervised learn-
ing and knowledge distillation for action spotting us-
ing transformers. In WACV, pages 530–540.
Giancola, S., Amine, M., Dghaily, T., and Ghanem, B.
(2018). Soccernet: A scalable dataset for action spot-
ting in soccer videos. In CVPRW, pages 1711–1721.
Giancola, S. and Ghanem, B. (2021). Temporally-aware
feature pooling for action spotting in soccer broad-
casts. In CVPR, pages 4490–4499.
Goes, F., Meerhoff, L., Bueno, M., Rodrigues, D., Moura,
F., Brink, M., Elferink-Gemser, M., Knobbe, A.,
Cunha, S., Torres, R., et al. (2021). Unlocking the
potential of big data to support tactical performance
analysis in professional soccer: A systematic review.
European Journal of Sport Science, 21(4):481–496.
Guti
´
errez-P
´
erez, M. and Agudo, A. (2024a). No bells just
whistles: Sports field registration by leveraging geo-
metric properties. In CVPRW.
Guti
´
errez-P
´
erez, M. and Agudo, A. (2024b). Pnlcalib:
Sports field registration via points and lines optimiza-
tion. SSRN.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In CVPR, pages
770–778.
Honda, Y., Kawakami, R., Yoshihashi, R., Kato, K., and
Naemura, T. (2022). Pass receiver prediction in soccer
using video and players’ trajectories. In CVPR, pages
3503–3512.
Kim, H., Choi, H.-J., Kim, C. J., Yoon, J., and Ko, S.-
K. (2023). Ball trajectory inference from multi-agent
sports contexts using set transformer and hierarchical
bi-lstm. In ACM SIGKDD, pages 4296–4307.
Liga, D. F. (2020). Positional tracking takes a big leap for-
ward as latest generation is installed at bundesliga and
bundesliga 2 stadiums. https://www.dfl.de/en/innova
tion/positional-tracking-takes-a-big-leap-forward-a
s-latest-generation-is-installed-at-bundesliga-and-b
undesliga-2-stadiums/.
Lin, J., Gan, C., and Han, S. (2019). Tsm: Temporal shift
module for efficient video understanding. In ICCV,
pages 7083–7093.
Luo, Y. and Mesgarani, N. (2019). Conv-tasnet: Surpassing
ideal time–frequency magnitude masking for speech
separation. T AUDIO SPE, 27(8):1256–1266.
Philipp Singer, Yauhen Babakhin, P. P. (2022). Winning
solution for bundesliga data shootout. https://www.ka
ggle.com/competitions/dfl-bundesliga-data-shootout/
discussion/359932.
Sanford, R., Gorji, S., Hafemann, L. G., Pourbabaee, B.,
and Javan, M. (2020). Group activity detection from
trajectory and video data in soccer. In CVPRW, pages
898–899.
Simpson, I., Beal, R. J., Locke, D., and Norman, T. J.
(2022). Seq2event: Learning the language of soccer
using transformer-based match event prediction. In
ACM SIGKDD, pages 3898–3908.
Sorano, D., Carrara, F., Cintia, P., Falchi, F., and Pap-
palardo, L. (2021). Automatic pass annotation from
soccer video streams based on object detection and
lstm. In ECML, pages 475–490.
Sudhakaran, S., Escalera, S., and Lanz, O. (2020). Gate-
shift networks for video action recognition. In CVPR,
pages 1102–1111.
Vidal-Codina, F., Evans, N., El Fakir, B., and Billingham,
J. (2022). Automatic event detection in football using
tracking data. Sports Engineering, 25(1):18.
Waibel, A., Hanazawa, T., Hinton, G., Shikano, K., and
Lang, K. J. (2013). Phoneme recognition using time-
delay neural networks. In Backpropagation, pages
35–61. Psychology Press.
Yamamoto, D. (2022). Third place solution for bundesliga
data shootout. https://www.kaggle.com/competitions/
dfl-bundesliga-data-shootout/discussion/360236.
Yeung, C. C., Sit, T., and Fujii, K. (2023). Transformer-
based neural marked spatio temporal point process
model for football match events analysis. arXiv
preprint arXiv:2302.09276.
Yu, G. and Yuan, J. (2015). Fast action proposals for human
action detection and search. In CVPR, pages 1302–
1311.
Zhou, X., Kang, L., Cheng, Z., He, B., and Xin, J. (2021).
Feature combination meets attention: Baidu soccer
embeddings and transformer based temporal detec-
tion. arXiv preprint arXiv:2106.14447.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
230

Figure 6: Four extra chunks from a test set match that show how our model fits the ground truth better than PassNet (Sorano
et al., 2021), even when using their criteria of non-instantaneous passes.
Table 6: Baikulov’s method (Baikulov, 2023) folds.
Events detection results for each of Baikulov’s method folds
isolated and the arithmetic mean.
Fold
0.6s 1s
↑ Prec ↑ Rec ↑ F1 ↑ Prec ↑ Rec ↑ F1 ↑ mAP@1
0 71.40 64.71 67.89 76.50 70.01 73.11 33.37
1 86.32 37.37 52.16 90.08 39.98 55.38 20.53
2 79.65 47.66 59.64 83.74 51.62 63.86 27.06
3 78.15 54.70 64.36 82.16 58.98 68.67 31.66
4 79.01 54.42 64.45 82.03 57.88 67.87 29.54
5 74.19 53.93 62.46 78.32 58.30 66.84 30.38
6 71.36 53.74 61.31 75.73 58.34 65.91 31.06
Mean 77.15 52.36 62.38 81.22 56.44 66.60 29.09
APPENDIX
Video Tubes Crops Resizing: The size of bound-
ing boxes in our dataset of matches recordings at
1920×1080px range from 15×20px for the furthest
players that appear smaller in the image to 70×105px
for the closest ones. We resize to 128×128 because it
is the size of the regions to be cropped after adding the
20% extra margin to the biggest bounding boxes we
find in our dataset, with this we avoid downsampling
and therefore losing information in largest boxes.
Comparing with the State of the Art: To compare
with (Baikulov, 2023), the winner of the SoccerNet
Ball Action Spotting challenge (Cioppa et al., 2023),
we just had to run their code in our matches footage
downsampled from 1080p to 720p. Their approach
employs a 7-fold cross-validation in the first pre-
training step that leads to a set of 7 models. Therefore,
we computed their best method performance as they
did, computing the detection results (only for passes)
for their 7 different-fold models and doing the arith-
metic mean of their outputs. Table 6 exposes how
each fold independently detects passes in our 6-match
test set for event detection, and the aforementioned
arithmetic mean that becomes the final output.
The comparison is done with the metrics we use
for events detection (precision, recall and F1 at 0.6s
acceptance window) but we also consider their metric
mean Average Precision with an acceptance window
of 1 second (mAP@1). This metric was introduced
by the SoccerNet first dataset (Giancola et al., 2018)
and defines a prediction as a true positive if it lands in-
side an acceptance window of δ seconds. Then, vary-
ing the tolerance, they compute a Precision-Recall
curve for each value of δ and finally average along all
the classes. Unlike in the challenge they won where
the acceptance windows range from 1 to 5 seconds,
(Baikulov, 2023) exposes that actions are densely al-
located and should be predicted more accurately us-
ing only 1 second windows, which goes along with
our focus on temporally accurate detection.
In Figure 4 we also compare with PassNet (So-
rano et al., 2021) and show how our model fits the
ground truth better even when using their pass crite-
ria. We add some zoomed in chunks in Figure 6 for
extra comparison.
Temporally Accurate Events Detection Through Ball Possessor Recognition in Soccer
231