Stroke Comparison between Professional Tennis Players and

Amateur Players using Advanced Computer Vision

Lisa Baily1, Nghia Truong2, Jonathan Lai3 and Phong Nguyen2

1The American School in Japan, 1-1-1 Nomizu, Chofu-shi, Tokyo, Japan
2Tokyo Techies, Shinbashi 2-16-1, Minato, Tokyo, Japan
3Tokyo Coding Club, Nishi Azabu 3-24-16, Minato, Tokyo, Japan

Keywords: Pose Estimation, Pose Tracking, Machine Learning, Computer Vision, Euclidean Distance, Tennis, Analysis.

Abstract: In this paper, we created a method to find how professional and amateur tennis serves differ from each other.
We collected videos from online and from our own recordings and turned those videos into frames. From
those frames, we manually selected ones appropriate for our study and ran those through a pose estimation
system, which turned those frames into simple stick figures of the players including all the x and y coordinates
of the player. By normalizing all data, we were able to calculate the Euclidean distance between two compared
players’ joints and analyze their consistency in their serves. Our results from our t-tests showed that there was
a significant difference between the amateur’s consistency and the pro’s consistency and body parts like both
shoulders showed a significant difference.

1 INTRODUCTION images from videos or photos and provide numerical

evaluations. From those outputs, we can analyze data
Tennis is a popular competitive and leisure sport that more specifically and efficiently and derive
is played in a one-on-one or two-on-two format. The compelling results. Applications of computer vision
sport is largely composed of various “strokes” to keep in the field of sports include but are not limited to
the ball in play, such as the forehand and backhand analysis and evaluation of tennis players (Mukai,
strokes during a rally and a serve to start the game. Of Asano, Hara, 2011), highlight detection (Ren, Jose,
those strokes, the serve plays a critical role, as it has 2009) and support decision making (Owens, Harris,
been shown to be one of the two most important shots Stennett, 2003).
along with the return in determining wins We propose using computer vision to analyze
(O'Donoghue and Brown, 2008). It is also a shot with tennis shots, and potentially provide amateur players
high variance, with variability in power, ball speed, with the level of specificity and data necessary to help
accuracy, ball impact location and angular velocities them improve. Although tennis includes many types
(Whiteside, et al. 2014, Martin, et al. 2016,). Given of strokes, we chose to focus on one of the most
the serve’s significance and variance, amateur players important: the serve (O'Donoghue and Brown, 2008).
often observe professional players who compete at Although the serve does not require much movement,
international tournaments like Wimbledon and the as the shot is hit in one stationary location, the way
US Open to emulate the form of those top players and the serves are hit varies between players, thus making
improve their own serve. However, simply watching it difficult to improve just by watching professionals'
them play is not nearly sufficient if the goal is to play. With a computer vision algorithm, recognizing
understand the real differences between an amateur what is different and how it is different from
and a professional. professionals to amateurs will become clearer.
Today, computer vision is a rapidly growing We first split the collected videos into frames and
technology within the broader fields of computer then used an accurate pose estimation system to
science and artificial intelligence (Arai and Kapoor simplify the frames into a stick representation of the
2019; Shavit and Ferens 2019). It is both fairly new player. After normalizing all data into the same size
and has a wide range of applications. It can take in and making it comparable, we were able to analyze
the similarities and differences between professional

44
Baily, L., Truong, N., Lai, J. and Nguyen, P.
Stroke Comparison between Professional Tennis Players and Amateur Players using Advanced Computer Vision.
DOI: 10.5220/0010145800440052
In Proceedings of the 8th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2020), pages 44-52
ISBN: 978-989-758-481-7
Copyright c 2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
 Stroke Comparison between Professional Tennis Players and Amateur Players using Advanced Computer Vision

and amateur players, leading to the conclusion that observed in the “descending windup or acceleration
not only were the patterns between the professionals phases”. While this does identify certain components
and amateurs different, but that specific body part of the serve that hold significant weight, our research
positioning showed a significant divergence. hopes to add data and detail to those components in
order to better understand the angles and stroke
lengths that separate the professional player from the
2 RELATED WORK amateur.
Baily and Nguyen (2018) developed a method to
A survey of what has been already published in this classify different tennis strokes based on an armband
area revealed a range of existing publications that that measures data from its accelerometer, gyroscope,
agreed on the importance of analyzing the serve in quaternion, and EMG. The authors use a supervised
greater detail along with other strokes, but chose to learning model, a Support Vector Machine (SVM), to
focus on different components. determine the correct tennis shot based solely data
In Whiteside et al. (2014), the researchers focused from the armband.
on the tossing component of the serve and how
important the consistency of it is to the resulting serve.
From their research, they were able to recognize that 3 PROPOSED METHOD
while professionals were not consistent in the
horizontal placement of the ball, they were In this section we describe our proposed method we
consistently tossing the ball at the same height. This used to analyze differences in player serves. We first
paper's main topic was about the serve but it differs collected sample serve videos from both amateurs and
from our paper, as they focused mainly on the toss of professionals from the Internet and our own
the ball, rather than focusing on the whole serving recordings. We manually looked through each video
motion. and identified the sets of frames that capture the serve
Chow et al. (2007) focused on how the activation motion. A pose estimation algorithm is used to
of the muscles varied before and after the impact in reconstruct the poses of each player appearing in
the tennis volley, as many players are concerned those frames, and the result is put through a pose
about the after effect, potentially leading to severe tracking system to label each person with an integer
injuries on the wrist. They collected data by placing identifier. We then manually labelled the result with
electrodes on the players’ bodies. This data the player name, ID number, and whether they are
collection was conducted with several controls, such left-handed or not. The labelled pose data is then
as the tennis string and racket type. From the EMG normalized to account for the difference in body size,
data, they were able to conclude that the oversize position in image and left-handedness. Finally, we
tennis balls “do not significantly increase upper calculated the Euclidean distance between the same
extremity muscle activation compared to regular size limbs in all pairs of serve clips collected and made
balls during a tennis volley”. While this paper observations based on statistical analysis. This is
focused primarily on volleys and not the serve, the visually represented in Figure 1 below.
level of detail it went into showing how even
miniscule changes in one’s form can lead to
drastically different physiological impacts in the
long run reinforced how important our research is
when it comes to a stroke that covers a much wider
range of motion than volleys.
This importance is corroborated by Chow et al.
(2009) which looked into how different types of
serves affect the players' conditions. They included 3
types of serves - flat, topspin, and slice, and examined
how those shots activate the middle and lower trunk
muscles. For each subject, their two highest rated
EMG and kinematic data, which are coordinate data Figure 1: Our data pipeline. Black arrows denote manual
steps, and blue arrows denote steps done using computer
extracted from their raw videos, were used to analyze
programs.
the differences. Even though there were no significant
effects for the serve type on muscle activation, they
found that on average, the largest EMG levels were

45
icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support

Essential to standardizing our results was the and assigned an identifier to each human, then
algorithm used for pose estimation, which has been manually reviewed the results and recorded the IDs
one of the major challenges in computer vision since of the main players as well as whether they’re left-
its introduction. In pose estimation, an algorithm handed or not. The tracking system used is Pose Flow
attempts to determine the positions and the poses of (Xiu, et al, 2018), which is available as an open
the humans in a given digital image and helps to source project on GitHub (Machine Vision and
ensure that the data collected is comparable. In this Intelligence Group, 2018). In this system, the pose
case, a human pose is defined as a set of points estimation result is fed to an optimization framework
describing the important body joints. For our problem, to build the association of cross-frame poses and form
we used the pose estimation algorithm proposed by pose flows, then to a pose flow non-maximum
Fang, et al. (2017). The framework, named Alpha suppression to robustly reduce redundant pose flows
Pose, first detects all human locations in an image. and re-link temporal disjoint ones. The result of this
Each location is treated as a single-person image and step is a database of poses in MS COCO format with
fed to a Symmetric Spatial Transformer Network player name, tracking ID, video link and handedness.
(Jaderberg et al., 2015) to find the region of interest,
continued to a Single Person Pose Estimator (Newell, 3.1 Data Processing
et al., 2016) to estimate the pose in local image and
finally through a Spatial De-Transformer Network to Serve videos of 4 professionals and 3 amateurs were
remap the estimated human pose back to the original used to conduct this research. 3 out of the 4
coordinate. The estimated poses are then refined professionals’ data were collected via the internet and
through the use of parametric Pose Non-Max the rest of the videos were collected on our own. In
Suppression (Fang et al., 2018) to obtain the final the data collection, we used videos including 4~13
human poses. We used the Alpha Pose authors’ serves per player and as a control, all of the videos
official implementation available on GitHub were captured from the back angle of the player. With
(Machine Vision and Intelligence Group, 2017), the videos, we turned them all into frames, thus
which outputs human poses in the Microsoft COCO making the data manipulation easier. All of the videos
(Lin et al., 2015) format1. were at 30 frames per second. We manually cut the
One of the common concerns in pose estimation frames into smaller sections, with only one full stroke
is that in a 2D image, very often some of the important per section. To keep the frame number per cut equal,
body joints are not visible. Alpha Pose addresses this we set a constant of 72 frames. This resulted in each
by representing a joint using 3 numbers: x-coordinate, player having 4~13 serve videos, each consisting of
y-coordinate and a confidence score. The third 72 frames, and the number for professionals and
number ranges from 0 to 1, with lower values amateurs were roughly equivalent, which makes the
assigned to less visible joints. Even when a joint is comparison more accurate. To further simplify and
completely invisible, unless it lies outside of the make the analysis accurate, we selected 21 frames
image, the model does a good job predicting its from those 72 frames, including the contact point of
position and assigning a confidence score. Our videos the serve and 10 frames before and after. We selected
were chosen so that the main player is always those specific frames because the time at which a
completely visible in most of the frames, so missing player takes before and after their contact point of the
data wasn’t a big concern. Also, for the sake of ball during a serve is different and only selecting
simplicity, we didn’t use the confidence score in our frames around the contact point reduces variation
analysis. between players during analysis.
The pose estimation step is repeated for all frames In Figure 2, the image highlighted in yellow is the
we wanted to analyze. Note that this analysis is done “contact point” frame, which is the point at which the
in 2 dimensions and not 3, and because we are player makes contact with the ball at the maximum
analyzing each frame, we compare sets of static poses height. By adding on 10 frames before and after, the
of the players, not their overall motion. Since there images capture the serve motion around the ball hit of
can be multiple people in a frame, we needed to the serve for a total of 21 frames.
accurately identify the main player in all frames. We
did this by running the pose estimation results
through a pose tracking system, which analyzed the
connectivity of the poses between consecutive frames

1
https://cocodataset.org/#format-data

46
 Stroke Comparison between Professional Tennis Players and Amateur Players using Advanced Computer Vision

Figure 5: Comparison of two players (one left-handed

which is the white stick figure and the other right-handed
with the red stick figure) initially without any scaling or
shifting.

We created separate scales for the x axis and the y

Figure 2: One of the professional’s 21 frames, the contact
axis. To find the right scales for the x coordinates, we
point frame and 10 frames before and after.
looked through all of the poses’ x coordinates of the
left and right shoulder and found the distance between
We then ran the Alpha Pose system on all those
them. We repeated this process for the y coordinates,
frames we manually collected and the output includes
the left and right hip, and we selected the greatest
a stick figure of the players with 17 important points
values of both the x and y to create the scale. These
on the player’s body.
scaling factors were then normalized to a set width
In Figure 3 and 4, we display the output of the and height. After finding the scaling factors we
Alpha Pose detection so that one can see the lines and applied it to all frames and finally shifted the poses,
key points drawn on the player’s body, representing in order for them to overlap with each other. With the
the simple outline of a human body in one frame. scaling and shifting, the poses now are comparable,
as shown in Figure 6.

Figure 3: A female professional player before and after

Alpha Pose detection.
Figure 6: The same players from Figure 5 but scaled and
shifted.

To further improve the comparison, we also

flipped left-handed players so that their data can be
analyzed as well with the right-handed players, which
is displayed in Figure 7 below.

Figure 4: A male professional player before and after Alpha

Pose detection.

Even though all of the videos were taken from the

back of the player, the distance between the player
and the camera varied throughout different videos so
normalizing the scales of the players became essential. Figure 7: Final output with scaling, shifting, and flipping
It is clear in Figure 5 that because the scaling is not (for left-handed players only).
applied, the poses do not overlap or match well to
each other.

47
icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support

3.2 Comparison players, only one player was left-handed. Similar to

the data collection method for the first player, we
As shown in Figure 8, we first collected the data, then ignored the first couple of serves and took the next 5
manually selected the important frames and put those to 10 serves, making sure that we collected their real
images through pose estimation. serving style. The point of this research is to compare
pros to pros, amateurs to amateurs, and amateurs to
pros to see whether the consistency amongst those
data sets are significantly different.

5 RESULTS
In this section, we will discuss the results collected
from our data. We first looked at 2 boxplots, side by
Figure 8: Flow diagram of the comparison process. side, of the sums of the Euclidean distances between
limbs for amateurs and pros.
Then with the normalization completed, we The results in Figure 9 clearly show that the
analyzed the data by taking the Euclidean distance distribution for the amateurs was more spread out
between each of the 17 points on the two players for when compared to the pros implying a greater
all of the frames. We calculated the Euclidean variance in the data. The median, as well as the
distance between the same joints of two players by interquartile range of the data, for amateurs are
using the equation . Each greater than that for the pros. Knowing that there are
player has 17 key points detected from the pose clear distinctions between the distributions of the pros
estimation and for each of the key points, the same and amateurs, we looked more closely to where
point on the other player’s pose estimation was exactly those differences arise by creating histograms
compared, using the equation above. We repeated this specific for each player.
process for all 17 points and summed up the distances
for us to compare.
To further analyze the differences between
players, we used t-tests to compare the distributions
of the data sets. The t-test data are specifically for the
player’s differences with themselves at their contact
point. Because we were aware that the variances
between each of the players were different, we used a
Welch’s t-test, which can be used on datasets with
varying standard deviations or heteroscedasticity.
Also, we used this type of test because the number of
Figure 9: Boxplot of the distributions of the sums of the
samples were different for each player. Euclidean distances between limbs for the amateur and pro
category.

4 EXPERIMENTAL SETTING In Figure 10, Figure 11 and Figure 12, the x axis
represents a normalized Euclidean distance between
To start off, we gathered videos from several angles each of the players, and the y axis represents the
of one player hitting overhead serves. Those videos frequency of those distances occurring. Figure 10
were 10 to 30 minutes of a player practicing the serve. compares Amateurs to other Amateurs, Figure 11
The first couple of serves, around 4 to 5, were ignored compares Professionals to Professionals while Figure
as they showed significant differences with the 12 compares Professionals to Amateurs. There is a
following serves and were likely warm-ups, so we clear distinction between the distributions of
collected 5 to 10 strokes of each player after their professionals and amateurs. The professionals’
warm-ups. To get a wider variety of players, we histograms are more tightly distributed and mostly
collected data from the Internet where there are plenty skewed to the right, meaning the differences between
of professional players’ practice videos. In total we their serves were not very large. However, the
gathered 4 professionals, 3 amateurs, and within those histograms of the amateur players have larger ranges

48
 Stroke Comparison between Professional Tennis Players and Amateur Players using Advanced Computer Vision

and their distributions are not as skewed compared to Because the histograms only provide qualitative
the professionals. This shows how amateur players data, we then used Welch's unequal variance t-test, a
were not consistently making similar movements, type of statistical analysis to determine whether there
thus shifting the distribution towards larger values. In is a significant difference between the means of two
Figure 12, it shows a histogram with pros and groups. This test showed a similar result when testing
amateurs being compared to each other. Compared to for significant differences between professional and
Figure 10 and 11, there are no distinct features that amateur players.
stand out when comparing pros to amateurs. We conducted a Welch’s t-test between the
professionals’ sums of distances and the amateurs’
sums of distances and the resulting p-value was
0.0036. From this, we were able to conclude that there
is, in fact, a significant difference between the means
of the two groups, the professionals’ sum and the
amateurs’ sum.
To further analyze where exactly those
Figure 10: Histograms of the distributions of the sums of differences were, we conducted several t-tests, shown
Euclidean distances between limbs comparing Amateurs to in Table 1, each for the key points on the player’s
Amateurs. body, and found that, while neither of the right wrist
nor left hip were significant, there were significant
differences in the rest of the body points analyzed (all
p-values less than 5%). The p-values for the shoulder
comparisons were most significant. With this, it is
evident that one of the most consistent differences
between amateurs and professionals is in the
shoulders.
In Figures 13, 14, 15, where we plot the
distribution of differences in left shoulder locations
across different player types, it is clear that the
differences between the distributions for the
professional and amateur players are significant. For
instance, Figure 14 shows that Professionals
compared to other different Professionals have a
significantly right skewed distribution while the
Amateurs compared to other different Amateurs
(Figure 13) or Amateurs compared to Professionals
(Figure 15) have a significantly less right skewed
distribution and in some cases are almost
symmetrically distributed.

Figure 11: Histograms of the distributions of the sums of

Euclidean distances between limbs comparing Pro to Pro.

Figure 13: Histograms of the distributions of the sums of

Euclidean distances between the left shoulder comparing
Amateurs to Amateurs (only left shoulder).

Figure 12: Histograms of the distributions of the sums of

Euclidean distances between limbs comparing Pro to
Amateur.

49
icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support

Table 1: All of the collected p-value results for different

types of distributions.

Compared Distributions P-values

Pro Sum to Amateur Sum 0.0036

Pro Right-Sum to Amateur Right- 0.0532

Sum

Pro Left-Sum to Amateur Left-Sum 0.00463

Pro Upper-Sum to Amateur Upper- 0.021998

Sum

Pro Left-Elbow to Amateur Left- 0.02279

Elbow

Pro Right-Elbow to Amateur Right- 0.003554

Elbow
Figure 14: Histograms of the distributions of the sums of
Euclidean distances between the left shoulder comparing Pro Right-Shoulder to Amateur 3.729 10
Pro to Pro (only left shoulder). Right-Shoulder

Pro Left-Shoulder to Amateur Left- 1.21 10

Shoulder

Pro Right-Wrist to Amateur Right- 0.9789

Wrist

Pro Left-Wrist to Amateur Left-Wrist 0.0346

Figure 15: Histograms of the distributions of the sums of Pro Right-Hip to Amateur Right-Hip 2.324 10
Euclidean distances between the left shoulder comparing
Amateur to Pro (only left shoulder). Pro Left-Hip to Amateur Left-Hip 0.0857

We conducted another test to see if there are clear Pro-to-Pro to Amateur-to-Amateur 2.735
distinctions between the distributions of differences 10
of professional player serves compared to other
professional players and the differences of amateur Pro-to-Pro to Pro-to-Amateur 3.083
player serves compared to other amateur players. In 10
other words, we are comparing the difference in the
pro distribution versus the amateur distribution. Amateur-to-Amateur to Pro-to- 0.3798
Amateur
From this we were able to conclude that those two
groups are, in fact, significantly different from each
other, with respect to intra-group differences, with a
p-value of 2.735 10 . In contrast, there was no 6 DISCUSSION
significant difference in amateur distribution to the
distribution of pro vs amateur differences.
In this section, we will discuss some possible
Although we only focused on some of the p-value
explanations and implications of our results and will
results, the numbers in Table 1 shows all of our results
evaluate the strengths and weaknesses of our research.
and although some values are not significant, others
To start off, not only have we confirmed the
show a significant value, like the pro-to-pro to pro-to-
obvious result that professional body movements
amateur.
during serves are significantly different to amateurs
in terms of consistency. We have also shown that
professionals are more consistent among each other

50
 Stroke Comparison between Professional Tennis Players and Amateur Players using Advanced Computer Vision

as a group then amateurs. Our main result however is dimensions, we are only looking at the x and y
our ability to narrow down the differences to each coordinates. However, because we are focusing on
limb area and do so with only a simple single analyzing players from only a single camera angle, 3-
recording of the player without the need for special dimensional analysis poses significant challenges that
set ups. Indeed, nearly half of our analyzed player require dedicated testing with a multiple camera setup
videos came from publicly available videos. to adequately address. Finally, we conducted our
Among our limb differences, while most limb research with only 7 athletes, which included 3
areas showed significant differences from pros to amateur and 4 professional players, and that is
amateurs, the right wrist and left hip were not considerably a low number of data points. In our
significantly different, in fact the right wrist was future work, the research can be further developed by
significantly similar. Given that we analyzed serves collecting more data for different players to ensure
frames around the ball contact point, this implies most more diversity in our collection.
players, professional and amateurs alike, can manage
to position their racket to an optimal contact point
with the ball, even if the rest of their body and 7 CONCLUSIONS
footwork is dissimilar or suboptimal. Although, the
left hip and leg is where most players are often taught In this paper, we collected videos of both amateur and
to keep their weight during a serve, the p-values seem
professional tennis players, and through the use of
to indicate there isn’t a significant difference in how pose estimation and tracking, we were able to
pros and amateurs position this limb even if there simplify frame images from videos into stick figures.
might be some small variations. This may imply that
With the given data, we analyzed the differences
most players, even amateurs, reach a good level of between players’ key points on their body, such as
consistency with this limb.
their shoulders and elbows. This led us to understand
Our findings are definitely informative to tennis
better how the consistency between pros and
players. This gives players points they can focus on amateurs differ and where the biggest differences lie.
improving and points where they may not need to
For example, in our P-value table, we found
spend as much effort, rather than watching significant differences in both shoulders while the
professionals and not knowing where to pay attention. right wrist showed little difference between
It allows amateur players to have an objective
professionals and amateurs. In future works, we look
understanding in their performance consistency, to further identify differences between professionals
compared to other professionals and other amateur and amateurs looking at differences in limb position
players. This data can be helpful to tennis coaches, as
and also body dynamics. Through our t-tests, we
it gives them a focus point in their lessons. Our data were able to conclude that the distributions of overall
is applicable to a wide range of players in a wide Euclidean distance between limbs as well as specific
range of situations because of our normalization
limbs such as the left shoulder, right shoulder, and
methods we applied on all stroke data and the right hip, for professionals and amateurs were
minimal requirements for the analysis videos, limited significantly different.
to only their shooting angle, without need for special
preparation.
However, the drawbacks are that we had to
manually select the 21 frames (1 contact point frame, REFERENCES
10 frames before and after), which we would ideally
like to automate. Additionally, because we looked Arai, K., Kapoor, S. 2019. Advances in Computer Vision,
into each video by frames, this means that we only Proceedings of the 2019 Computer Vision Conference
(CVC), Volume 1. Springer.
considered a series of static poses, not a time
Baily, L., Nguyen, P., 2018. Tennis Stroke Classification
evolution and that is one limitation our research has. using Myo Armband. The 1st International Young
The static poses are adequate enough for the research Researchers Conference, 2018.
but it also means that the overall flow of the strokes Chow, J., Knudson, D., Tillman M., and Andrew, D., 2007.
are disregarded, meaning we could have been Pre and post impact muscle activation in the tennis
overlooking important parts regarding the overall volley: effects of ball speed, ball size and side of the
movements of the player’s strokes. Another weak body. British Journal of Sports Medicine.
point of our research is that our analysis was only in Chow, J., Park, S., Tillman, M. 2009. Lower trunk
2 dimensions, not 3 dimensions. This is a limitation kinematics and muscle activity during different types of
tennis serves. BMC Sports Sci Med Rehabil 1, 24
as even though the player’s movements are in 3

51
icSPORTS 2020 - 8th International Conference on Sport Sciences Research and Technology Support

Fang, H.-S., Xie, S., Tai, Y.-W., Lu, C., 2017. RMPE:
Regional multi-person pose estimation, in International
Conference on Computer Vision (ICCV).
Jaderberg, M., Simonyan, K., Zisserman, A., Kavukcuoglu,
K., 2015. Spatial transformer networks. In Conference
on Neural Information Processing Systems (NIPS),
pages 2017–2025.
Lin, T.-Y., Maire, M., Belongie, S., Bourdev, L., Girshick,
R., Hays, J., Perona, P., Ramanan, D., Zitnick, C. L.,
Dollár, P., 2015. Microsoft COCO: Common objects in
context, in International Conference on Computer
Vision (ICCV).
Machine Vision and Intelligence Group at Shanghai Jiao
Tong University, 2017. AlphaPose. GitHub repository.
https://github.com/MVIG-SJTU/AlphaPose
Machine Vision and Intelligence Group at Shanghai Jiao
Tong University, 2018. PoseFlow. GitHub repository.
https://github.com/YuliangXiu/PoseFlow
Martin C, Bideau B, Delamarche P, Kulpa R, 2016.
Influence of a Prolonged Tennis Match Play on Serve
Biomechanics. PLoS ONE 11(8): e0159979.
Mukai, R., Asano, T. and Hara, H., 2011. Analysis and
Evaluation of Tennis Plays by Computer Vision, 2011
International Conference on Mechatronics and
Automation (ICMA), pages 784–788
Newell, A., Yang, K., and Deng, J., 2016. Stacked
hourglass networks for human pose estimation. In arXiv
preprint arXiv:1603.06937
O'Donoghue, P., Brown, E., 2008. The Importance of
Service in Grand Slam Singles Tennis. International
Journal of Performance Analysis in Sport. 8. 70-78.
Owens, N., Harris, C., Stennett, C., 2003. Hawk-eye tennis
system, International Conference on Visual
Information Engineering.
Ren, R., Jose J. M., 2009. General highlight detection in
sport videos, ACM Multimedia Modeling 2009, pages
27-38
Shavit, Y., Ferens, R., 2019. Introduction to Camera Pose
Estimation with Deep Learning. In arXiv preprint
arXiv:1907.05272.
Whiteside, D., Giblin, G., Reid, M., 2014. Redefining
Spatial Consistency in the Ball Toss of the Professional
Female Tennis Serve. 32 International Conference of
Biomechanics in Sports.
Xiu, Y., Li, J., Wang, H., Fang, Y., Lu, C., 2018. Pose Flow:
Efficient online pose tracking. In arXiv preprint
arXiv:1802.00977.

52