Play Evaluation Based on Predicting the Outcome of Back-Row Attacks

in Volleyball

Hikaru Yoshihara

, Ning Ding

and Keisuke Fujii

1 a

Nagoya University, Japan

Nagoya Institute of Technology, Japan

Keywords:

Machine Learning, Volleyball, Tracking Data.

Abstract:

In volleyball, statistical analysis based on data aggregation at the team or match levels has developed, and its

use for player performance evaluation and tactical analysis has expanded. However, there has been limited

discussion on the quantitative evaluation of how individual plays affect rally outcomes. To address this issue,

a model that predicts rally outcomes under speciﬁc conditions using player location data is useful. This study

aims to evaluate plays based on a prediction model, focusing on the ﬁrst transition following a back-row

attack. We extracted 103 target scenes from game footage recorded from behind the end line and manually

created tracking data for six players per team. Using this dataset, we trained an XGBoost model to predict

the future probability of scoring and the probability of blocking by two or more opponents in each game state

(receive, toss, attack). To quantify play evaluation, we propose the Valuating Volleyball States by Estimating

Probabilities (V2SEP), which expresses play evaluation values in each state based on the prediction model,

weighting them according to the percentage of points scored when a player is blocked. To verify the validity

of the prediction model used in V2SEP, we assessed F1 scores and SHAP values for each state. The results

indicate that the predictions were reasonably accurate and reﬂected not only the contributions of directly

involved players but also those of other players affecting scoring and block induction. Furthermore, the play

evaluation metrics demonstrate expected trends whereas some scenes show the limitations, suggesting that

V2SEP may be useful for play evaluation in volleyball.

1 INTRODUCTION

Volleyball is a six-player team sport in which players

rally the ball across a net while trying to prevent it

from touching the ground. The game is played on

a rectangular court measuring 18 m × 9 m. Each

team is allowed a maximum of three touches before

returning the ball to the opponent’s court. The basic

sequence of play consists of a receive, a set, and an

attack.

According to volleyball rules, back-row players

are not allowed to attack near the net. If they attack,

they must jump from behind the attack line. An at-

tack executed under this restriction is called a back-

row attack (or back-row attack). One advantage of a

back-row attack is that involving a back-row player in

the offense can create a numerical advantage against

the opponent’s block. Additionally, combination at-

tacks from both front-row and back-row players in-

https://orcid.org/0000-0001-5487-4297

creases the variety of offensive plays, such as exe-

cuting a delayed attack. Furthermore, pressure ex-

erted by the back-row attack can contribute to a higher

success rate for front-row attacks. However, because

back-row attacks are performed farther from the net,

they have a higher likelihood of errors, such as hit-

ting the net or sending the ball out of bounds. There-

fore, players must carefully adjust their approach and

attack angles. Moreover, not only the attacker but

also the positioning and movements of surrounding

players play a crucial role in reducing the number of

blockers and enhancing the effectiveness of the attack.

In this study, we focus on back-row attacks and aim

to quantitatively evaluate sequences of play by pre-

dicting attack outcomes based on player positions and

movements.

In recent years, data analysis in volleyball has ad-

vanced, expanding its applications in player perfor-

mance evaluation and tactical analysis. One exam-

ple is a study that analyzed the number, duration,

and height of tosses in a single men’s World Champi-

Yoshihara, H., Ding, N. and Fujii, K.

Play Evaluation Based on Predicting the Outcome of Back-Row Attacks in Volleyball.

DOI: 10.5220/0013666100003988

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2025), pages 29-37

ISBN: 978-989-758-771-9; ISSN: 2184-3201

onship match (Hashihara et al., 2009). Additionally,

by tabulating the tosses and spikes that occur during

transitions (switching between offense and defense in

response to an opponent’s attack), it was revealed that

the ability to execute ﬁrst transitions signiﬁcantly im-

pacts match results (Yonezawa, 2003). In particu-

lar, ﬁrst transitions in response to combination attacks

were found to have a strong inﬂuence on match results

(Yonezawa, 2004).

Machine learning has also been applied to vol-

leyball analysis. For example, studies have used

both rule-based and black-box models (Lalwani et al.,

2022), as well as Bayesian networks with hidden

Markov processes (Ge and Song, 2024). The former

utilizes team-level factors such as win percentage, av-

erage points per match, and team ranking, while the

latter incorporates intra-match factors such as points

scored, successful serves, and successful blocks to

predict match results and analyze the inﬂuence of

these features on predictions. While team-based and

match-based data aggregation analyses are progress-

ing, discussions on the quantitative evaluation of each

play’s impact on rally outcomes remain insufﬁcient.

Simple statistical analysis is challenging because rally

outcomes depend on numerous factors, including the

attacker’s position, their approach, the setter’s accu-

racy, coverage by surrounding players, and the oppo-

nent’s defense. In addition, although player location

data on the court coordinates have been publicly re-

leased and utilized for analysis in other sports (Fujii,

2025) such as soccer (Somers et al., 2024), basketball

(Scott et al., 2024), and badminton (Ding et al., 2024),

no such datasets are publicly available for volleyball.

To assess the impact of plays on rally outcomes

while considering multiple factors simultaneously, a

predictive model based on tracking data is needed.

For instance, in soccer, the Valuing Actions by Es-

timating Probabilities (VAEP) framework quantita-

tively evaluates the contribution of actions such as

passing and dribbling to goal-scoring opportunities

(Decroos et al., 2019). Additionally, a method called

Valuing Defense by Estimating Probabilities (VDEP),

derived from VAEP, has been introduced to assess de-

fensive effectiveness (Toda et al., 2022; Umemoto

et al., 2022), contributing to the development of a

framework for analyzing both offensive and defensive

plays. Although such quantitative evaluation meth-

ods for individual plays have been proposed in other

sports and have contributed to improving competitive

performance, similar research in volleyball remains in

its early stages.

In this study, we propose Valuating Volleyball

States by Estimating Probabilities (V2SEP), a frame-

work for evaluating plays based on predictive mod-

els that utilizes tracking data of six players per team,

extracted from match videos (Figure 1). Our anal-

ysis focuses speciﬁcally on the ﬁrst transition fol-

lowing a back-row attack. The key contributions

of this study are as follows: The main contribu-

tions of this study are: (1) We propose a quan-

titative framework to evaluate the impact of indi-

vidual volleyball plays on rally outcomes. (2) We

present the ﬁrst publicly available volleyball player

location and event dataset for performance analy-

sis. (3) Experimental results indicated that the pre-

diction models were reasonably accurate and capture

both direct and indirect player contributions on scor-

ing and block, and demonstrate the overall validity

of the proposed method while showing some limi-

tations. The rest of this paper is structured as fol-

lows. Section 2 reviews related work. Section 3 intro-

duces our dataset and Section 4 describes our V2SEP

method. Section 5 presents experimental results, and

Section 6 concludes the paper. Datasets are avail-

lable at https://github.com/keisuke198619/V2SEP-

volleyball.

Figure 1: Overview of this study. We tracked the players

and obtained their coordinates on the court by using a ho-

mography transformation. Then, we used this data to pre-

dict the result of the attack and evaluate the play.

2 RELATED WORK

As studies analyzing data on a match-by-match ba-

sis, a study has analyzed the relationship between the

number of blockers and the type of attack against the

opponent’s attack during the ﬁrst transition, as well

as the match outcome, reconﬁrming the signiﬁcant ef-

fect of spiking on match results (Minowa et al., 2016).

There is also a study on the defensive side that an-

alyzed factors such as the position before the block,

the number of blockers, and block height, and that

different factors were weighted in different phases of

the game (Matsui et al., 2011). Another study has re-

ported measurements of the time required for block-

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

ing at different attack positions, and it showed that

reaction time is delayed for attacks from the left and

right due to the larger movement steps required (Ya-

mada et al., 2012).

A previous study on back-row attacks found that

about 90% of combination attacks included a back-

row attack (Yoshida et al., 2018). It was also noted

that the occurrence rate and scoring rate of back-row

attacks have increased, and that the increase in the

scoring rate of back-row attacks has also inﬂuenced

the scoring rate of other attacks (Yoshida et al., 2016;

Nakanishi and Ohkubo, 2021). Additionally, it was

reported that the choice of back-row attack position

depends on the position of the toss (Adin-Marian and

Marilena, 2014).

To analyze the movement of individual players,

for example, from a biomechanics perspective, there

is a research that analyzes the speed and jump time

during a spike, and the angle of the torso and arms at

each timing such as during the jump, attack, and land-

ing (Awang Irawan et al., 2023). Another research

compared spiking with one foot and both feet, ana-

lyzing their respective advantages in terms of center

of gravity velocity and horizontal speed (Huang et al.,

1999). Furthermore, the study examining the rela-

tionship between players’ physical characteristics and

performance revealed strong correlations, particularly

with height, muscle mass, and bone mass (Sanjayku-

mar et al., 2024). Machine learning has also been ap-

plied to analyze individual player contributions. For

example, one study used a Bayesian hierarchical lo-

gistic model to estimate data from the World Cham-

pionships (W. Fellingham, 2022).

As described above, previous studies have mainly

focused on collecting and analyzing data on a match-

by-match basis or examining methods for analyzing

individual player performance. However, detailed

analysis on a rally-by-rally basis using tracking data,

such as player movement and placement, is still un-

explored ﬁeld. This study takes a different approach

from previous research by aiming to use tracking data

to predict attack outcomes and evaluate play on a

rally-by-rally basis.

3 METHODS

3.1 Datasets

In this section, we explain the dataset in this study,

preprocessing, and computing variables. We empha-

size that our dataset is the ﬁrst publicly available vol-

leyball player location and event dataset for perfor-

mance analysis.

3.1.1 Videos

In this study, we selected and analyzed 13 match

videos uploaded to the YouTube channel “Volleyball

Watchdog”. We chose these videos because they were

all ﬁlmed with a ﬁxed camera positioned behind the

end line, providing a relatively clear view of all six

players on each team. All videos had a frame rate

of 30 fps. The beginning and ending frames of se-

quences in which a serve, reception, set, and back-row

attack occurred were recorded in a spreadsheet. Based

on these records, 103 video clips were extracted using

the free software Aviutl.

3.1.2 Video Annotation

We performed automatic tracking using basketball-

SORT (Hu et al., 2024) on the split videos.

Basketball-SORT is an object detection approach spe-

cialized for tracking basketball plays, designed to re-

solve complex occlusion issues involving multiple ob-

jects by utilizing players’ trajectories and appearance

features. The tracking results contained errors such as

multiple IDs assigned to a single player, ID swaps be-

tween different players, and incorrectly sized bound-

ing boxes (Figure 2). We corrected these issues man-

ually using Labelbox, an online annotation software.

Figure 2: Examples of errors in tracking. (a) is an example

of the ID being swapped. (b) is an example of a player

with multiple IDs. (c) is an example of an incorrectly sized

bounding box.

3.1.3 Acquisition of Player Coordinates

Based on the corrected bounding box data, we applied

a homography transformation to obtain player coordi-

nates on a 9 m × 9 m court. We used the center of the

bottom edge of each bounding box as the reference

point for player coordinates. Additionally, during the

homography transformation, the coordinates of the

Play Evaluation Based on Predicting the Outcome of Back-Row Attacks in Volleyball

four corners of the court were manually recorded in

advance and used as reference points (Figure 3). The

transformed coordinate data had issues such as in-

terruptions in tracking data when players moved off-

screen and ﬂuctuations in the y-coordinates due to

jumping. To address tracking interruptions, we ap-

plied linear interpolation using the last known coordi-

nates before a player exited the screen and the ﬁrst de-

tected coordinates upon their return. To correct ﬂuc-

tuations in the y-coordinates, We checked frames im-

mediately before takeoff and after landing manually,

and performed linear interpolation using the recorded

coordinates at these points. After these corrections,

we smoothed the coordinate data to mitigate abrupt

velocity changes caused by manual annotation adjust-

ments. A moving average with a window width of 5

was applied for smoothing.

Figure 3: By performing a homography transformation

based on the four red points shown on the left, we obtained

the coordinates on the coat as shown on the right.

3.1.4 Event Annotation

For each of the videos, we recorded the frame num-

ber at which each event occurred, the ID of the player

who executed it, the time elapsed between events, the

attacker’s jump time, and the attack result. The three

types of events considered here are receive, toss, and

attack. The player ID was assigned based on the cre-

ated coordinate data.

3.2 Proposed Method: V2SEP

Here we show a quantitative framework to evaluate

the impact of individual volleyball plays on rally out-

comes.

3.2.1 Preprocessing

In this study, we computed evaluation indices based

on the VDEP concept proposed in previous study

(Toda et al., 2022). As a preliminary step, we cre-

ated a dataset for predicting attack outcomes by calcu-

lating the movement speed and direction of all play-

ers at the moment of an event, using their position

data and event information. To compute these val-

ues, we ﬁrst extracted the ﬁve frames before and after

the event frame. The movement speed was calculated

as the average speed over these ten frames, while the

movement direction was determined by calculating

the direction from the player’s coordinates in the ﬁrst

frame to those in the last frame. For the ﬁve players

other than the one performing the event, we calculated

both their distance from the event-performing player

and the direction to that player, then sorted them

by distance. Finally, for each event, we compiled

the event-performing player’s coordinates, movement

speed, and direction at the time of the event, as well

as the other players’ coordinates, movement speed,

direction, and their distance and direction relative to

the event-performing player. We stored the compiled

data in a CSV ﬁle.

3.2.2 V2SEP

We propose V2SEP to evaluate volleyball play.

Speciﬁcally, we apply the VDEP methodology to vol-

leyball and modify it to evaluate the process of attack.

In this study, we extracted the scenes from the op-

ponent’s serve, receive, toss, and attack. Hereafter,

the game states in which the attack outcome is pre-

dicted are called the receive, toss, and attack states,

respectively, and are given by the states s

= [s

receive

toss

, s

attack

]. We deﬁne the future probability of scor-

ing P

point

) and the probability of an opponent block

by two or more players P

blocked

) (hereafter called

the block probability) in each state. The attacking

team is preferred to act so that P

point

) is higher

or P

blocked

) is lower. Therefore, we propose the

following formula to calculate the evaluated value of

play.

V 2SEP

) = P

point

) − C ∗ P

blocked

) (1)

C is a parameter that weights the scoring probability

and blocking probability. In the collected data, there

were 41 videos where two or more players blocked

during an attack, and in 15 of those cases, the at-

taching team lost the point due to a block. Thus,

≃ 0.366, so we set C = 0.366.

3.2.3 Probability Prediction Method

For the classiﬁer to predict probabilities, we used eX-

treme Gradient Boosting (XGBoost), following pre-

vious studies. Gradient boosting methods are known

to perform well on a variety of learning problems in-

volving heterogeneous features, noisy data, and com-

plex dependencies. For the features used to predict

the probability in a state, we assumed that not only the

features in that state but also all the previous features

were used. Speciﬁcally, the features of the receiving

and tossing states were used to predict the probability

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

of the tossing state, and the features of the receiving,

tossing, and attacking states were used to predict the

probability of the attacking state. For the attack re-

sult, we deﬁned point and blocked as labels to record

whether a point was scored as a result of a back-row

attack and whether two or more opponents jumped for

a block against the attack, respectively. If the attack

results in a score, point = 1 is assigned, and if the at-

tack fails to score, point = 0 is assigned. Similarly,

blocked = 1 was assigned when two or more oppo-

nents jumped for a block. In the data created for this

study, out of 103 videos, there were 52 videos with

point = 1 and 51 videos with point = 0. Addition-

ally, 42 videos had blocked = 1, and 61 videos had

blocked = 0.

3.2.4 Validation

Since the method proposed in this study assumes that

the predictions made by the classiﬁer are valid, it is

necessary to ﬁrst verify that the classiﬁer is making

valid predictions. In addition, by visualizing the de-

gree of inﬂuence in the prediction, it will be possible

to assess which features have the most inﬂuence and

to what extent. Therefore, we calculated F1 scores

and SHAP values for the score and block predictions

in each state, respectively. Next, two veriﬁcations are

performed to determine whether the calculated eval-

uation values work as expected. In the ﬁrst veriﬁca-

tion, we divided the data into four groups based on

the result labels point and blocked, and calculated the

average evaluation value for each group. In the sec-

ond veriﬁcation, we compared the calculated evalua-

tion values with the actual footage to qualitatively de-

termine whether the evaluation values are legitimate.

Scenes that are prone to receiving unjustiﬁed evalua-

tion values are then investigated to identify areas for

improvement in the proposed method.

4 EXPERIMENTS

4.1 Validation of the Models

The proposed method (V2SEP) in this study assumes

that the attack result predictions are valid. Therefore,

we ﬁrst calculated the F1 score for both the score and

block predictions. The results are shown in Table 1.

In both predictions, the F1 score gradually increased,

which was consistent with our intuition that the at-

tack result becomes more predictable as the game pro-

gresses.

Next, feature importance was visualized and qual-

itatively analyzed by calculating SHAP values. SHAP

Table 1: Evaluation Results for point and blocked.

receive toss attack

point 0.501 0.531 0.553

blocked 0.493 0.633 0.647

values quantify how each feature contributes to a

paticular prediction, showing how each variable in-

ﬂuences the model’s output. Figure 4 and Figure 5

shows the top-ranked feature importance in the at-

tack state. For the estimation of scoring probability

in the attack state, the features of players who per-

formed back-row attacks tended to have higher ab-

solute SHAP values than those of players who made

tosses. On the other hand, for the estimation of block

probability, the features of players who made tosses

tended to have larger absolute SHAP values. In both

predictions for the attack state, 14 or 15 of the top 20

features with the highest absolute importance values

belonged to players who were not directly involved

in the event, conﬁrming that the movements of such

players have a signiﬁcant impact on the prediction.

Figure 4: Top 20 absolute SHAP values of features in attack

events for score predictions. Orange indicates the feature

values for the player who attacked, green for the player who

tossed, yellow for the player who received, and blue for the

other players.

4.2 Validation of Evaluation Values

We conducted two validations to determine whether

the evaluation values calculated based on V2SEP

functioned as expected. First, we divided the data

Play Evaluation Based on Predicting the Outcome of Back-Row Attacks in Volleyball

Figure 5: Top 20 absolute SHAP values of features in attack

events for block predictions. Orange indicates the feature

values for the player who attacked, green for the player who

tossed, yellow for the player who received, and blue for the

other players.

into four groups based on the attack result labels,

point and blocked, and calculated the average eval-

uation value for each group. The results are shown

in Table 2. When comparing the evaluation values in

the receiving state, although the group with point =

1, blocked = 0 had a lower value, there was no signiﬁ-

cant difference between the groups. However, the dif-

ferences became more pronounced as the game pro-

gressed toward the attack state. Furthermore, when

comparing the two groups where point = 1, the group

with blocked = 0 had a higher average evaluation

value. Similarly, for the two groups where point = 0,

the group with blocked = 0 also had a higher average

evaluation value.

Table 2: The average evaluation value for each group.

receive toss attack

point = 1, blocked = 0 0.332 0.352 0.362

point = 1, blocked = 1 0.357 0.307 0.333

point = 0, blocked = 0 0.365 0.317 0.316

point = 0, blocked = 1 0.352 0.305 0.281

Next, we visually and qualitatively analyzed

whether each video had a valid evaluation value.

Speciﬁcally, we compared the calculated evaluation

values with the actual footage, selected scenes where

the evaluation was clearly justiﬁed or unjustiﬁed,

and examined trends in scenes that were particularly

prone to unjustiﬁed evaluation values. For this analy-

sis, we used the following criteria: whether there was

any disruption in movement during receiving or toss-

ing, whether the setter (who tosses the ball directly

leading to an attack) was forced to move due to a dis-

rupted receive, and how many players participated in

the attack.

First, an example of a scene that received a cor-

rectly high evaluation is shown in Figure 6. In this

case, the evaluation values were 0.449 for the receiv-

ing state, 0.517 for the tossing state, and 0.579 for

the attacking state. When checking the frames be-

fore and after the tossing state, the movement of the

player making the toss appeared relatively smooth,

with three front-row players and one back-row player

participating in the attack. Furthermore, from the

tossing state to the attacking state, the opponent’s

blockers were drawn toward the two front-row play-

ers, leaving the back-row attacker unblocked.

Figure 7 shows an example of a scene that re-

ceived a correctly low evaluation. In this case, the

evaluation values were 0.483 for the receiving state,

0.086 for the tossing state, and -0.025 for the attack-

ing state. While there was no signiﬁcant disruption in

movement during the receiving state, the unsuccessful

receive forced the tosser to make a toss while moving

backward from the edge of the court. The toss then

went to a back-row player, but due to its predictable

trajectory and the delay between the toss and the at-

tack, the attacker was blocked by three opponents.

Figure 8 shows an example of a scene with an un-

justiﬁed evaluation. In this situation, the 6th player

was originally supposed to make the toss, but due to a

disrupted receive, another player was forced to do so.

Additionally, tossing from a position far from the net

limits the attack patterns and increases the likelihood

of a block, meaning this scene should have received a

lower evaluation.

5 DISCUSSIONS

This section discusses the validity of the prediction

model and the proposed method (V2SEP) based on

the analysis results from the previous section. Table 1

shows that the F1 score tends to improve as the game

progresses and the number of features used for predic-

tion increases. In particular, the F1 scores for predic-

tions in the attack state were 0.553 for score predic-

tion and 0.647 for block prediction. Considering the

difﬁculty of predicting outcomes in team sports, these

can be regarded as relatively accurate predictions.

The analysis of SHAP values in Figure 4 and

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

Figure 6: Examples of scenes with correctly high ratings. From left to right: receiving state, tossing state, and attacking state.

Figure 7: Examples of scenes with correctly low ratings. From left to right: receiving state, tossing state, and attacking state.

Figure 8: Examples of scenes with incorrect evaluation values. From left to right: receiving state, tossing state, and attacking

state.

Figure 5 suggests that the attacker’s movement has

a greater inﬂuence on whether a point is scored,

while the setter’s movement has a greater inﬂuence

on whether the opponent blocks. Furthermore, among

the top 20 features ranked by SHAP value, informa-

tion from players not directly involved in the event

also impacted the prediction. This indicates that the

proposed method can incorporate the movements of

players who are not directly involved in the event into

its evaluation. However, the SHAP values for features

related to players not directly involved in the event

tended to be larger than those for features related to

setters and attackers, suggesting that the inﬂuence of

non-involved players may be overly emphasized.

Table 2 also conﬁrms that as the game progresses,

differences emerge in the average evaluation values

across groups. This aligns with the analysis of Ta-

ble 1, which indicated that prediction accuracy im-

proves as the game progresses. In the attack state,

when comparing groups with the same point value,

those with blocked = 0 tend to have a higher average

evaluation value. Similarly, when comparing groups

with the same blocked value, those with point = 1

tend to have a higher average evaluation value. These

ﬁndings suggest that the evaluation values follow the

expected trend to some extent.

In the qualitative evaluation based on matching

with actual video footage, higher evaluation values

were often assigned to scenes in which the player suc-

cessfully guided the opponent’s block (Figure 6) and

scenes where multiple players were involved in the

attack (Figure 7), conﬁrming a certain level of valid-

ity. On the other hand, there were cases where evalua-

tion values were not assigned correctly. For example,

higher-than-expected evaluation values were given in

scenes where the number of players able to participate

in the attack decreased due to a loss of balance after

receiving (see ﬁgure) and in scenes where the setter

was unable to make the toss (Figure 8). The former

issue is likely due to the fact that changes in player co-

ordinates were small for both stable and unbalanced

receptions, making it difﬁcult for the model to distin-

guish between them. Additionally, since there were

only a limited number of scenes where a player lost

balance during receiving, the dataset appears insufﬁ-

cient for considering the impact of reception stabil-

ity on attack outcomes. The latter issue likely arises

from the fact that the dataset used in this study does

Play Evaluation Based on Predicting the Outcome of Back-Row Attacks in Volleyball

not distinguish players by speciﬁc positions, such as

front/rear or left/center/right, and does not explicitly

identify the setter’s movements.

Aside from these issues, this study limited the

target scenes and created the dataset by extracting

footage from matches of multiple teams, making

team-by-team comparisons difﬁcult. To enable such

comparisons, it would be necessary to construct a

dataset by focusing on speciﬁc teams and selecting

target scenes accordingly. Furthermore, since this

study focused solely on back-row attacks against op-

ponent serves, its general applicability remains undis-

cussed. To conﬁrm whether the proposed method can

be applied to other situations, its validity must be ver-

iﬁed by expanding the range of target scenes.

6 CONCLUSIONS

In this study, we evaluated plays in each state based

on a prediction model for the outcomes of back-row

attacks in volleyball. To assess play performance,

we proposed V2SEP, which utilizes the probabil-

ity of scoring—estimated based on features in each

state—as well as block prediction. We then veri-

ﬁed the validity of the calculated evaluation values.

Given that volleyball is a team sport where outcome

prediction is inherently difﬁcult, the prediction model

was found to be reasonably accurate. Although some

scenes were still not evaluated appropriately, the cal-

culated evaluation values demonstrated a certain de-

gree of validity and generally followed the expected

trends.

Future challenges include distinguishing players

by speciﬁc positions when inputting features for pre-

diction and verifying the general applicability of the

proposed method by extending the target scene to

cover the period from the opponent’s serve until

the ball drops. Additionally, to generate prediction

data more efﬁciently, it is necessary to develop a

volleyball-speciﬁc tracking method or automate the

modiﬁcation of tracking data.

REFERENCES

Adin-Marian, C. and Marilena, C. (2014). Preferential and

performance-related lateralization in volleyball as far

as attack hit execution is concerned. Procedia - Social

and Behavioral Sciences, 127:58–62.

Awang Irawan, F., Widya Permana, D. F., Nurrahmad, L.,

Anam, K., Hadi, Romadhoni, S., and Syaraﬁna Ghas-

sani, D. (2023). A motion analysis of volleyball open

spike: Kinematics and performance. International

Journal of Human Movement and Sports Sciences,

11(1):134–142.

Decroos, T., Haaren, J. V., Bransen, L., and Davis, J.

(2019). Actions speak louder than goals: Valuing

player actions in soccer. In Proceedings of the 25th

ACM SIGKDD International Conference on Knowl-

edge Discovery & Data Mining, pages 1851–1861.

Ding, N., Takeda, K., Jin, W., Bei, Y., and Fujii, K.

(2024). Estimation of control area in badminton dou-

bles with pose information from top and back view

drone videos. Multimedia Tools and Applications,

83(8):24777–24793.

Fujii, K. (2025). Machine Learning in Sports: Open Ap-

proach for Next Play Analytics. Springer.

Ge, Y. and Song, N. (2024). Exploring the relationship be-

tween key factors and victory and defeat in volleyball

matches using bayesian network modeling. Journal of

Electrical Systems, 20(6):2241–2250.

Hashihara, Y., Yoshida, Y., and Yoshida, M. (2009). The

playing system of international-class men’s volleyball

teams. Journal of volleyball sciences, 11(1):12–18.

Hu, Q., Scott, A., Yeung, C., and Fujii, K. (2024).

Basketball-sort: An association method for complex

multi-object occlusion problems in basketball multi-

object tracking. Multimedia Tools and Applications,

83(38):86281–86297.

Huang, C., Liu, G.-C., and Sheu, T.-Y. (1999). Kinematic

analysis of the volleyball back row jump spike. 17

International Symposium on Biomechanics in Sports,

pages 49–52.

Lalwani, A., Saraiya, A., Singh, A., Jain, A., and Dash, T.

(2022). Machine learning in sports: A case study on

using explainable models for predicting outcomes of

volleyball matches. In Proceedings of the 2nd Inter-

national Conference on Sports Engineering.

Matsui, T., Yajima, T., and Miyakozawa, T. (2011). The

composition for the accomplishment process of block-

ing action for producing effective it in volleyball

game: With focus on serve phase, rally phase and at-

tack tempos. Journal of volleyball sciences, 13(1):30–

37.

Minowa, K., Imamaru, K., and Matsumoto, Y. (2016). Team

performance evaluation and outcome of the game by

the ﬁrst transition in volleyball games. Journal of vol-

leyball sciences, 18(1):19–26.

Nakanishi, Y. and Ohkubo, S. (2021). Research on the at-

tack tactics in volleyball. Journal of volleyball sci-

ences, 23(1):31–38.

Sanjaykumar, S., Natarajan, S., Yoga Lakshmi, P.,

Kalmykova, Y., Lobo, o., Pavlovic, R., and Setiawan,

E. (2024). Machine learning analysis for predicting

performance in female volleyball players in india: Im-

plications for talent identiﬁcation and player develop-

ment strategies. Journal of Human Sport and Exer-

cise, 20(1):207–215.

Scott, A., Uchida, I., Ding, N., Umemoto, R., Bunker, R.,

Kobayashi, R., Koyama, T., Onishi, M., Kameda, Y.,

and Fujii, K. (2024). Teamtrack: A dataset for multi-

sport multi-object tracking in full-pitch videos. In

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

Proceedings of the IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition, pages 3357–

3366.

Somers, V., Joos, V., Cioppa, A., Giancola, S.,

Ghasemzadeh, S. A., Magera, F., Standaert, B., Man-

sourian, A. M., Zhou, X., Kasaei, S., et al. (2024).

Soccernet game state reconstruction: End-to-end ath-

lete tracking and identiﬁcation on a minimap. In Pro-

ceedings of the IEEE/CVF Conference on Computer

Vision and Pattern Recognition, pages 3293–3305.

Toda, K., Teranishi, M., Kushiro, K., and Fujii, K. (2022).

Evaluation of soccer team defense based on prediction

models of ball recovery and being attacked: A pilot

study. In Published in PLOS ONE.

Umemoto, R., Tsutsui, K., and Fujii, K. (2022). Location

analysis of players in uefa euro 2020 and 2022 using

generalized valuation of defense by estimating proba-

bilities. arXiv preprint arXiv:2212.00021.

W. Fellingham, G. (2022). Evaluating the performance of

elite level volleyball players. Journal of Quantitative

Analysis in Sports, 18(1):15–34.

Yamada, Y., Fukutomi, K., Kanda, S., Kaneko, M., Goto,

H., Ueda, Y., Eto, N., Takanashi, Y., Kawagishi, Y.,

and Ishigaki, H. (2012). Measurement of the time

structure of the block in volleyball: in the light of

the offence area and height. Journal of volleyball sci-

ences, 14(1):7–11.

Yonezawa, T. (2003). A study on the transition of the vol-

leyball game. Fukuoka University Review of Sports

and Health Science, 33(1):27–34.

Yonezawa, T. (2004). A study on the ﬁrst transition of the

volleyball game : The performance which inﬂuenced

on the ﬁrst transition. Fukuoka University Review of

Sports and Health Science, 35(1):1–9.

Yoshida, Y., Hamagushi, J., and Yamada, K. (2016). The

research in the ways they attack patterns on woman’s

volleyball. BULLETIN OF Sports Research Center,

HOSEI UNIVERSITY, 34:5–10.

Yoshida, Y., Hamagushi, J., and Yamada, K. (2018). The

research in the ways they attack patterns on men’s vol-

leyball. BULLETIN OF Sports Research Center, HO-

SEI UNIVERSITY, 38:93–99.

Play Evaluation Based on Predicting the Outcome of Back-Row Attacks in Volleyball