Available via license: CC BY 4.0
Content may be subject to copyright.
Puck localization and multi-task event recognition in broadcast hockey videos
Kanav Vats Mehrnaz Fani David A. Clausi John Zelek
University of Waterloo
Waterloo, Ontario, Canada
{k2vats, mfani, dclausi,jzelek}@uwaterloo.ca
Abstract
Puck localization is an important problem in ice hockey
video analytics useful for analyzing the game, determining
play location, and assessing puck possession. The problem
is challenging due to the small size of the puck, excessive
motion blur due to high puck velocity and occlusions due to
players and boards. In this paper, we introduce and imple-
ment a network for puck localization in broadcast hockey
video. The network leverages expert NHL play-by-play
annotations and uses temporal context to locate the puck.
Player locations are incorporated into the network through
an attention mechanism by encoding player positions with
a Gaussian-based spatial heatmap drawn at player posi-
tions. Since event occurrence on the rink and puck location
are related, we also perform event recognition by augment-
ing the puck localization network with an event recognition
head and training the network through multi-task learning.
Experimental results demonstrate that the network is able
to localize the puck with an AUC of 73.1% on the test set.
The puck location can be inferred in 720p broadcast videos
at 5frames per second. It is also demonstrated that multi-
task learning with puck location improves event recognition
accuracy.
1. Introduction
Ball tracking in sports is of immense importance to
coaches, analysts, athletes and fans. The location of the
ball is directly related with the location of the play and can
also be used in tasks such as player and team possession
analysis. Hence, a computer vision based ball track-
ing/localization system can be of high utility. Although
there has been significant effort for soccer ball tracking
[1,7,22,24], hockey puck tracking is more challenging
due to a puck’s small size, velocity, and regular occlusion
due to players and opaque boards.
Many authors either only track the ball in screen coor-
dinates [9,15,26] or track ball on the field by treating it
Figure 1: Subset of 1500 puck locations in the dataset. The
puck locations on the ice rink are highly correlated with the
event label. Faceoffs(red) are located at the faceoff circles,
shots(blue) are located in the offensive zones and dump
in/outs (yellow) are presents in the neutral zone.
as a two-stage process: (1) tracking the ball in the screen
coordinates (2) registering the screen coordinates to the
field coordinates using automated homography [8,18]
after performing tracking. A big issue in ball tracking is
the requirement of a large amount of frame-by-frame ball
annotations for training which can be very difficult and
time consuming to obtain [12].
In this paper, we introduce a successful network for
localizing hockey puck on the ice rink. The model directly
estimates the puck location on the ice rink (instead of
the afore-mentioned two-stage approach). Rather than
estimating puck location from static images, the model
estimates the puck location from video using the temporal
context and leverages player location information with
heatmaps using an attention mechanism (Fig. 2). Instead
of annotating data on a frame-by-frame basis, we utilize
the existing NHL data available on a play-by-play basis
annotated by expert annotators. Experimental results
demonstrate that the network is able to locate the puck with
an AUC of 73.1% on the test set. The network is able to
localize the puck during player and board occlusions. At
test-time, the network is able to perform inference using a
sliding window approach in previously unseen untrimmed
broadcast hockey video at 5frame per second (fps).
1
arXiv:2105.10563v1 [cs.CV] 21 May 2021
Figure 2: The overall network architecture. Green represents model layers while pink represents intermediate features. The
network consists of four components: (1) Video Branch, (2) Player Branch, (3) Attention, and (4) Output. The Video Branch
extracts spatio-temporal features from raw hockey video. The Player Branch extracts play location information from player
Gaussian heatmaps. The Attention component fuses the player location and spatio-temporal video information. The Output
component produces the puck location output from the features obtained from the attention component.
Player and puck location information is related with
event occurring on the rink (Fig. 1). Other research
leverages player and ball trajectories for event recognition
using a separate tracking/localization system [11,17]. We
attach an event recognition head to the puck localization
model to leverage the puck location information for event
recognition and train the whole network using multi-task
learning. Experimental results demonstrate that event
recognition accuracy can be improved using puck location
information as an additional signal.
2. Background
2.1. Ball tracking using traditional computer vision
In soccer, a common approach to the ball tracking
problem is a two-stage approach [7,24]: (1) ball tracking
in screen coordinates and (2) sports field registration
via homography. Yamada et al. [24] perform camera
calibration by matching straight and curved lines in the
soccer field coordinates to the model. Candidates for the
ball are identified by looking for white patches and tracking
is performed with the help of a 3D motion model. Ishii
et al. [7] use a two synchronized camera system to track
the soccer ball in 3D coordinates with ball detection done
through template matching and tracking is done with the
help of a 3D Kalman filter. Ariki et al. [1] use a combi-
nation of global and local search for soccer ball tracking,
with the global search consisting of template matching
and local approach consisting of a particle filter. Yu et al.
[25] propose a trajectory based algorithm for ball tracking
in tennis where instead of determining whether an object
candidate is the ball, trajectory candidates are classified
into ball trajectories. Wang et al. [22] propose a unique
conditional random field (CRF) based algorithm to exploit
the contextual relationship between the players and ball for
ball tracking. Yakut et al. [23] used background subtraction
to track hockey puck in zoomed in broadcast videos for
short time intervals. The puck tracking performance
deteriorated with high motion blur, fast camera motion and
occlusions.
2.2. Ball tracking using deep learning
Recently, deep neural networks (DNNs) have found ap-
plication in sports ball tracking. Zhang et al. [26] track
golf ball in high resolution, slow-motion videos using a
patch based object detector and discrete Kalman filter. Ko-
morowski et al. [9] use a fully convolutional network uti-
lizing multiscale features to predict soccer ball confidence
maps. Reno et al. [15] use a convolutional neural network
(CNN) with image patches as input to detect the presence
of tennis balls. Our work is related to Voeikov et al. [21]
where they introduce a multi-task approach for tracking a
table-tennis ball using a cascade of detectors using frame-
level ball location annotations.
Puck tracking in hockey is relatively unexplored due to
the high level of difficulty involved. Pidaparthy et al. [12]
propose using a CNN to regress the puck’s pixel coordinates
from single high-resolution frames collected via a static
camera for the purpose of automated hockey videography.
The method involved an extensive annotation pipeline for
model training. Instead of inferring the ball location from
images and frame level annotations, we use a CNN to pre-
2
dict the puck location on the ice rink directly from short
videos with approximate annotations without using any ex-
ternal homography model.
2.3. Event recognition in sports
In the literature, video understanding in sports is often
framed as event spotting, aimed at associating events with
anchored time stamps [5,10], player level action recog-
nition [4,20] and event recognition which involves di-
rectly classifying a video into one of the known categories.
[13,16]. Event recognition is an important task in vision-
based sports video analytics. Tora et al. [16] recognize
hockey event from video by gathering player level contex-
tual interaction with the help of an LSTM. Others make use
of pre-computed player and ball trajectories for recognizing
events [11,17]. Mehrsa et al. [11] use player trajectories
obtained from a player tracking system in order to utilize
them for event recognition as well as team-classification in
ice hockey. Sanford et al. [17] use player and ball trajecto-
ries obtained from a tracking system for detecting events in
soccer. Instead of using player trajectories, we use puck lo-
cation information to recognize hockey events using multi-
task learning.
3. Methodology
3.1. Dataset
The dataset consists 8,987 broadcast NHL videos of two
second duration with a resolution of 1280 ×720 pixels and
a framerate of 30 fps with the approximate puck location
on the ice rink annotated. The annotations are rough and
approximate such that the puck location corresponds to the
whole two second video clip rather than a particular frame.
The videos are split into 80% samples for training and 10%
samples each for validation and testing. Fig 1shows the
distribution of a subset of puck location data. The videos
are also annotated with an event label which can be either
Faceoff, Advance (dump in/out), Play ( player moving the
puck with an intended recipient e.g., pass, stickhandle ) or
Shot. The distribution of event labels is shown in Fig. 3.
3.2. Puck localization
The overall network architecture consists of four
components: Video branch, Player branch, Attention and
Output. The architecture is illustrated in Fig. 2. The next
four subsections explain the components in detail.
3.2.1 Video branch
The purpose of the video branch is to obtain rele-
vant spatio-temporal information to estimate puck
location. The video branch takes as input 16 frames
Table 1: Network architecture of player location backbone.
k,s and p denote kernel dimension, stride and padding re-
spectively. Chi,C hoand bdenote the number of chan-
nels going into and out of a block and batch size re-
spectively.Additionally each layer contained a residual-skip
connection with a 1×1convolution.
Input: Player heatmap b×256 ×256
Layer 1
Conv2D
Chi= 1, C ho= 2
(k = 3×3,s=2,p=1)
Batch Norm 2D
ReLU
Layer 2
Conv2D
Chi= 2, C ho= 4
(k = 2×2,s=2,p=0)
Batch Norm 2D
ReLU
Layer 3
Conv2D
Chi= 4, C ho= 8
(k = 2×2,s=2,p=0)
Batch Norm 2D
ReLU
Output b×32 ×32 ×8
Table 2: Network architecture of Regblocks 1 and 2 for out-
put pw∈R200. k,s and p denote kernel dimension, stride
and padding respectively. C hi,Choand bdenote the num-
ber of channels going into and out of a block and batch size
respectively. Additionally each layer contained a residual-
skip connection with a 1×1×1convolution.
Input: F0b×4×32 ×32 ×256
Reg Block 1
Conv3D
Chi= 256, C ho= 200
(k = 2×2×2,s=2×2×2,p=0)
Batch Norm 3D
ReLU
Reg Block 2
Conv3D
Chi= 200, C ho= 200
(k = 2×2×2,s=2×2×2,p=0)
Batch Norm 3D
ReLU
Global average pooling
Sigmoid activation
Output b×200
{fi∈R256×256×3, i ∈ {1..16}} sampled from a short
video clip Vof two second duration. The frames are
passed through a backbone network consisting of four
layers of R(2+1)D network [19] to obtain features
Fv∈R4×32×32×256 to be used for further processing. The
R(2+1)D network consists of (2+1)D blocks which splits
spatio-temporal convolutions into spatial 2D convolutions
3
Figure 3: Distribution of event labels in the dataset. The
dataset is imbalanced with Play event having the most oc-
currence.
followed by a temporal 1D convolution.
3.2.2 Player branch
The location of puck on the ice rink is correlated with
the location of the players since the puck is expected to
be be present where the player ”density” is high. We
make the assumption that the location of players remains
approximately the same in a short two second video clip.
In order to encode the spatial player location, we take the
middle frame fmof the video Vand pass it through a
FasterRCNN [14] network to detect players. After player
detection, we draw a Gaussian with a standard deviation
of σpat the centre of the player bounding boxes to obtain
the player location heatmap H. An advantage of using this
representation is that the player location variability in the
video clip can be expressed through the Gaussian variance.
The player location heatmap His passed through a player
location backbone network to output player location
features Fp∈R32×32×8. The exact configuration of the
player location backbone is shown in Table 1. The player
location features Fpare passed to the attention block for
further processing.
3.2.3 Attention
The purpose of attention is to make the network incorpo-
rate player locations by considering the relationship be-
tween video features Fvand player location features Fp.
The player location features Fpand video features Fvare
concatenated along the the channel axis by repeating the
player location features along the temporal axis. The con-
catenated features Fcat ∈R4×32×32×264 are then passed
through a variation of the squeeze and excitation [2,6] net-
work consisting of a 3×3convolution, non-linear excita-
tion and 1×1convolution. The 3×3squeeze operation
learns the spatial relationships between player locations on
the rink and video features. The squeeze operation outputs
features F0
cat ∈R4×32×32×132. The squeeze operation is
followed by non linear activation and 1×1convolution to
obtain features Fa∈R4×32×32×256. The 1×1convolution
learns the channel wise relationships between the feature
maps in F0
cat. Finally, the output of the attention block is
the hadamard product of the attention features Faand the
video features Fvfollowed by a skip connection.
Fo=Fa⊗Fv+Fv(1)
3.2.4 Output
The features Foobtained from the attention component
are finally passed through two RegBlocks to output the
probability of puck location on the ice rink. Global average
pooling is done at the end of the two RegBlocks to squash
the intermediate output to one dimensional vectors. This is
done independently for rink width and height dimensions
through two separate branches. The overall network out-
puts two vectors, pw∈R200 and ph∈R85, in accordance
with the dimension of the NHL rink. The exact details of
RegBlocks 1 and 2 are shown in Table 2. Regblocks 3 and
4 have a similar architecture, the only difference is that
instead of a R200 vector pw, a R85 vector phis output by
changing the output channels to 85.
3.2.5 Training details
We use the cross entropy loss to train the network. In or-
der to create the ground truth, we use a one dimensional
Gaussian with mean at the ground truth puck location and
a standard deviation σfor both directions. The Gaussian
variance encodes the variability in ball location in the short
video clip (Fig. 5) . The total loss Lpuck is the sum of the
loss in horizontal axis Lwand vertical axis Lh, which is
given by:
Lpuck =Lw+Lh(2)
Lpuck =−1
200
200
X
i=1
wgt log pw−1
85
85
X
j=1
hgt log ph(3)
Where wgt ∈R200 and hgt ∈R85 denote the ground truth
probabilities and pw∈R200 and ph∈R85 denote the pre-
dicted probabilities.
For data augmentation, each frame is sampled from a uni-
form distribution U(0,60) so that the network sees different
frames of the same video when the video sampled different
times. The data augmentation technique is used is all exper-
iments unless stated otherwise. We use the Adam optimizer
with an initial learning rate of .0001 such that the learning
rate is reduced by a factor of 1
5at iteration number 5000.
The batch size is 15.
4
Figure 4: (a) Accuracy (φ) vs threshold (t) curve. (b) The
best performing model gets an overall AUC of 73.1% on
test set.
3.3. Multi-task event recognition
The event occurring on the rink in hockey is highly cor-
related with the puck location. For example, faceoff occurs
on the faceoff circles, shots are expected to occur in the of-
fensive zones etc. In order to leverage the shared informa-
tion between puck location and event recognition, we learn
the event and puck location in hockey video clip using a
single network through multi-task learning. This is done by
appending a third event recognition head at the end of fea-
tures Forepresenting the probability of the predicted event
pe∈R4. Let Chi,Choand kdenote the number of chan-
nels going into and out of a kernel and kernel size respec-
tively. The event recognition head consists of a 3D convolu-
tion layer with Chi= 256,C ho= 256 with k= 2 ×3×3
and stride = 2 followed by 3D batch normalization , fol-
lowed by another 3D convolution Chi= 256,C ho= 512
with k= 2 ×3×3and stride = 2, adaptive pooling and
fully connected layer. The total loss is the linear combina-
tion of equation 2and the event loss Lewhich is a cross
entropy loss between the ground truth and predicted event
probability. Following Cipolla et al. [3], the overall loss for
the muti-task network is given by:
Lmulti =1
σ2
1
Lw+1
σ2
2
Lh+1
σ2
3
Le+log(σ1)+log(σ2)+log(σ3)
(4)
The rest of the training details and data augmentation are
the same as in section 3.2.5.
4. Results
4.1. Puck localization
4.1.1 Accuracy metric
A test video is considered to be correctly predicted at a tol-
erance tfeet if the distance between the ground truth puck
Figure 5: Construction of ground truth for a training sample
with puck located at w= 44 ft and h= 5 ft. (a) Ground
truth distribution vector wgt ∈R200 (b) Ground truth dis-
tribution vector hgt ∈R85
location zand predicted puck location zpis less than tfeet.
That is ||z−zp||2< t. Let φ(t)denote the percentage of ex-
amples in the test set with correctly predicted position puck
position at a tolerance of t. We define the accuracy metric
as the area under the curve (AUC) φ(t)at tolerance of t= 5
feet to t= 50 feet.
4.1.2 Trimmed video clips
The network attains an AUC of 73.1% on the test dataset il-
lustrated in Fig. 4(b). The AUC in the horizontal direction
is 81.4% and AUC in vertical direction is 87.8%. From Fig.
4(a), at a low tolerance of t= 12 f t, the accuracy in ver-
tical(Y) direction is 76% and the accuracy in horizontal(X)
direction is 63%. At a tolerance of t= 20 ft, the accuracy
in both directions is greater than 80% .
Fig. 6show the zone wise accuracy. A test example is
classified correctly if the predicted and ground truth puck
location lies in the same zone. From Fig. 6(a), the network
gets an accuracy of ∼80% percent in the upper and lower
halves of the offensive and defensive zones. From Fig. 6
(b), after further splitting the ice rink in nine zones, the
network achieves an accuracy of more than 70% in five
zones. The network also has failure cases. From Fig. 6(b),
it can be seen that accuracy is low (less than 60% ) in the
bottom halves of the defensive and offensive zones. This is
due to the puck being occluded by the rink boards.
4.1.3 Untrimmed broadcast video
We also test the network on untrimmed broadcast videos
using a sliding window of length land stride s. The
window length lis the time duration covered by the sliding
window and stride sis the time difference between two
consecutive application of the sliding window. Due to the
difficulty of annotating puck location frame-by-frame in
5
Figure 6: Zone-wise accuracy. The figure represents the
hockey rink with the text in each zone represents the per-
centage of test examples predicted correctly in that zone.
The position of the camera is at the bottom. In (b), the ac-
curacy is low in the lower halves of the defensive and offen-
sive zones since the puck gets occluded by the rink board.
720pvideos, we do not possess the frame-by-frame ground
truth puck location. Therefore, we perform a qualitative
analysis in this section. The videos used for testing are
previously unseen video not present in the dataset used for
training and testing the network.
To determine the optimal values of stride svalidation
is performed on a 10 second clip. Some frames from the
validation 10 second clip are shown in Fig. 7. Whenever
visible, the location of the puck is highlighted using a red
circle. Fig 8shows the trajectories obtained. The network is
able to approximately localize the puck in untrimmed video
within acceptable visual errors, even though the network is
trained on trimmed video clips where puck location is an-
notated approximately. The puck is not visible during many
frames of the video, but the network is still able to guess
the puck location. This is because the network takes into
account the temporal context and player location. Since the
network is originally trained on 2second clips, the window
length lis fixed to 2s. Fig 8, shows that as the stride sis
decreased, the puck location estimates become noisy. Since
between two passes, the puck motion is linear, we do not de-
crease stride below 0.5sas it leads to very noisy estimates
(Fig. 9). The optimal stride s= 1sgives the most accu-
rate result. A lower stride results in noisy results and higher
strides produces very simple predictions.
The network is tested on another 10 second video with
l= 2sand s= 1sshown in Fig 10. The predicted puck
trajectory is shown in Fig 10. The puck is occluded by the
rink board during a part of the video (shown in images 5
and 6). The network is able to localize the puck even when
it is not visible due to board occlusions.The inference time
of the network on a single GTX 1080Ti GPU with 12GB
memory is 5fps.
Table 3: Comparison of AUC with different values of σ
with a three layer backbone network. Network with σ= 30
shows the best performance
σAUC AUC(X) AUC(Y)
20 62.5 71.3 85.07
25 68.5 77.9 85.6
30 69.0 78.5 85.5
35 68.9 78.8 85.4
Table 4: Comparison of AUC with different number of lay-
ers of the backbone R(2+1)D network. A four layer back-
bone shows the best performance.
Layers AUC AUC(X) AUC(Y)
2 56.3 73.2 74.1
3 69.0 78.5 85.5
472.5 81.3 87.3
5 72.4 81.0 87.3
4.2. Ablation studies
We perform an ablation study on the number of layers
in the backbone network, puck ground truth standard devia-
tion, presence/absence of player branch consisting of player
locations and data augmentation .
4.2.1 Puck ground truth standard deviation
The best value of standard deviation σof puck location
ground truth 1D Gaussian is determined by varying σfrom
20 to 35 in multiples of five. From Table 3, the number of
layers in the backbone is fixed to three while player loca-
tion based attention is not used. Maximum AUC of 69%
is attained with σ= 30 feet. A lower value of σmakes
the ground truth Gaussian more rigid/peaked which makes
learning difficult. A value of sigma greater than 30 low-
ers accuracy since a higher σmakes the ground truth more
spread out which reduces accuracy on lower tolerance val-
ues.
4.2.2 Layers in backbone
We determine the optimal number of layers in the R(2+1)D
backbone network by extracting the video branch features
from different layers without using the player location
based attention. The puck ground truth standard deviation
is set to the optimal value of 30. From Table 4, the
maximum AUC of 72.5% is achieved by using 4 layers
of R(2+1)D network. Further increasing the number of
backbone layers to 5 causes a decrease of 0.1in AUC due
to overfitting.
6
Figure 7: Some frames from the 10 second validation video clip. Whenever visible, the location of the puck is highlighted
using the red circle. The initial portion of the clip is challenging since the puck is not visible in the initial part of the clip.
Figure 8: Puck trajectory on the ice rink for the validation
video. The trajectory becomes noisy with s= 0.5sand
lower.
Figure 9: Puck trajectory for the validation video with a
very low stride of 0.125 seconds. The trajectory is ex-
tremely noisy and hence is not a good estimate.
4.2.3 Player location based attention
We add the player branch and the attention mechanism to
the network with 4 backbone layers and σ= 30. Three val-
ues of player location standard deviation σp={15,20,25}
are tested. From Table 5, adding the player location based
attention mechanism brought an improvement in the overall
AUC by 0.6% with σp= 15. Further increasing σpcauses
the player location heatmap to become more spread out ob-
fuscating player location information.
Table 5: Comparison of AUC values with/without player
branch. The player branch with σp= 15 shows the best
performance.
Player detection σpAUC AUC(X) AUC(Y)
No - 72.5 81.3 87.3
Yes 15 73.1 81.4 87.8
Yes 20 72.8 81.5 87.3
Yes 25 72.2 80.4 87.9
Table 6: Comparison of AUC values with uniform and ran-
dom sampling
Sampling method AUC AUC(X) AUC(Y)
Constant interval 70.3 79.4 86.4
Random 73.1 81.4 87.8
4.2.4 Data augmentation
We compare the data augmentation technique done using
randomly sampling frames from a uniform distribution (ex-
plained in Section 3.2.5) to sampling frames at a constant
interval. From Table 6, removing random sampling de-
creases the overall AUC by 3.2% which demonstrates the
advantage of the data augmentation technique used.
4.3. Multi-task event recognition
The network performing only event recognition task
with zero weights assigned to the puck location loss is
treated as a comparison baseline. We compare the macro
averaged precision, recall and F1 score values correspond-
ing to the four events for the multi-task learning setting and
the baseline.
From Table 7, the multi-task setting performs better
compared to the baseline where puck location is not used
as an additional signal which demonstrates that learning the
two tasks together is beneficial for event recognition. This
is because multi-task learning with puck location provides
7
Figure 10: Some frames from the test 10 second clip. Whenever visible, the location of the puck is highlighted using the red
circle
(a) (b)
Figure 11: The predicted puck trajectory for the test video with window length two seconds (l= 2s) and stride one second
(s= 1s) . The network is able to localize the puck even when it is not visible due to board occlusions.
Table 7: Precision, Recall and F1 score values for the net-
work corresponding to the multi-task and baseline settings.
The multi-task setting shows better performance.
Precision Recall F1 score
Muti task
Play 81.8 87.2 84.4
Shot 56.4 60.6 58.4
Advance 63.2 31.3 41.9
Faceoff 76.3 90.0 82.6
Macro Avg. 69.4 67.3 66.8
Baseline
Play 81.0 88.6 84.6
Shot 63.5 56.0 59.5
Advance 55.4 31.3 40.0
Faceoff 75.9 82.0 78.8
Macro Avg. 69.0 64.5 65.8
contextual location information which greatly improves F1
score of events such as Faceoff (82.6multi-task vs 78.8
baseline) which always occur in specific rink locations. The
Advance event has the lowest F1 score value of 41.9. This
is because it often gets confused with Play and Shot events.
5. Conclusion
A model has been designed and developed to localize
puck and recognize events in broadcast hockey video. The
model makes use of temporal information and player lo-
cations to localize puck. We append an event recognition
head to the puck localization model and train the whole
network using multi-task learning. We also perform abla-
tion studies on the model parameters and data augmentation
used. We attain an AUC of 73.1% on the test set and qual-
itatively localize the puck in untrimmed broadcast videos.
We also report an ice rink region based average accuracy
of 80.2% with the ice rink split into five zones and 67.3%
with the rink split into nine regions. Experimental results
also demonstrates that the puck location signal aids event
recognition with the multi-task learning setting improving
the macro-average event recognition F1-score by one per-
cent. Future work will focus on using high resolution im-
ages/videos and frame-wise puck location annotations to
improve performance.
Acknowledgments. This work was supported by Stath-
letes through the Mitacs Accelerate Program and Natu-
ral Sciences and Engineering Research Council of Canada
(NSERC). We also acknowledge Compute Canada for hard-
ware support
References
[1] Y. Ariki, Tetsuya Takiguchi, and Kazuki Yano. Digital cam-
era work for soccer video production with event recogni-
8
tion and accurate ball tracking by switching search method.
In 2008 IEEE International Conference on Multimedia and
Expo, pages 889–892, 2008. 1,2
[2] A. Bhuiyan, Y. Liu, P. Siva, M. Javan, I. B. Ayed, and
E. Granger. Pose guided gated fusion for person re-
identification. In 2020 IEEE Winter Conference on Applica-
tions of Computer Vision (WACV), pages 2664–2673, 2020.
4
[3] R. Cipolla, Y. Gal, and A. Kendall. Multi-task learning using
uncertainty to weigh losses for scene geometry and seman-
tics. In 2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 7482–7491, 2018. 5
[4] M. Fani, H. Neher, D. A. Clausi, A. Wong, and J. Zelek.
Hockey action recognition via integrated stacked hourglass
network. In 2017 IEEE Conference on Computer Vision
and Pattern Recognition Workshops (CVPRW), pages 85–93,
2017. 3
[5] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and
Bernard Ghanem. Soccernet: A scalable dataset for action
spotting in soccer videos. In The IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR) Workshops,
June 2018. 3
[6] J. Hu, L. Shen, and G. Sun. Squeeze-and-excitation net-
works. In 2018 IEEE/CVF Conference on Computer Vision
and Pattern Recognition, pages 7132–7141, 2018. 4
[7] Norihiro Ishii, Itaru Kitahara, Yoshinari Kameda, and Yuichi
Ohta. 3d tracking of a soccer ball using two synchronized
cameras. In Horace H.-S. Ip, Oscar C. Au, Howard Leung,
Ming-Ting Sun, Wei-Ying Ma, and Shi-Min Hu, editors, Ad-
vances in Multimedia Information Processing – PCM 2007,
pages 196–205, Berlin, Heidelberg, 2007. Springer Berlin
Heidelberg. 1,2
[8] Wei Jiang, Juan Camilo Gamboa Higuera, Baptiste Angles,
Weiwei Sun, Mehrsan Javan, and Kwang Moo Yi. Optimiz-
ing through learned errors for accurate sports field registra-
tion. In Proceedings of the IEEE/CVF Winter Conference on
Applications of Computer Vision (WACV), March 2020. 1
[9] Jacek Komorowski, Grzegorz Kurzejamski, and G. Sar-
was. Deepball: Deep neural-network ball detector. ArXiv,
abs/1902.07304, 2019. 1,2
[10] William McNally, Kanav Vats, Tyler Pinto, Chris Dulhanty,
John McPhee, and Alexander Wong. Golfdb: A video
database for golf swing sequencing. In The IEEE Confer-
ence on Computer Vision and Pattern Recognition (CVPR)
Workshops, June 2019. 3
[11] Nazanin Mehrasa, Yatao Zhong, F. Tung, L. Bornn, and G.
Mori. Learning person trajectory representations for team
activity analysis. ArXiv, abs/1706.00893, 2017. 2,3
[12] H. Pidaparthy and J. Elder. Keep your eye on the puck: Auto-
matic hockey videography. In 2019 IEEE Winter Conference
on Applications of Computer Vision (WACV), pages 1636–
1644, Jan 2019. 1,2
[13] AJ Piergiovanni and Michael S. Ryoo. Early detection of
injuries in mlb pitchers from video. In Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern
Recognition (CVPR) Workshops, June 2019. 3
[14] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun.
Faster r-cnn: Towards real-time object detection with region
proposal networks. In Proceedings of the 28th International
Conference on Neural Information Processing Systems - Vol-
ume 1, NIPS’15, page 91–99, Cambridge, MA, USA, 2015.
MIT Press. 4
[15] V. Ren`
o, N. Mosca, R. Marani, M. Nitti, T. D’Orazio, and E.
Stella. Convolutional neural networks based ball detection in
tennis games. In 2018 IEEE/CVF Conference on Computer
Vision and Pattern Recognition Workshops (CVPRW), pages
1839–18396, 2018. 1,2
[16] Moumita Roy Tora, Jianhui Chen, and James J. Little. Clas-
sification of puck possession events in ice hockey. In Pro-
ceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) Workshops, July 2017. 3
[17] Ryan Sanford, Siavash Gorji, Luiz G. Hafemann, Bahareh
Pourbabaee, and Mehrsan Javan. Group activity detection
from trajectory and video data in soccer. In Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern
Recognition (CVPR) Workshops, June 2020. 2,3
[18] R. A. Sharma, B. Bhat, V. Gandhi, and C. V. Jawahar. Auto-
mated top view registration of broadcast football videos. In
2018 IEEE Winter Conference on Applications of Computer
Vision (WACV), pages 305–313, 2018. 1
[19] D. Tran, H. Wang, L. Torresani, J. Ray, Y. LeCun, and M.
Paluri. A closer look at spatiotemporal convolutions for ac-
tion recognition. In 2018 IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition, pages 6450–6459,
2018. 3
[20] K. Vats, H. Neher, D. A. Clausi, and J. Zelek. Two-stream
action recognition in ice hockey using player pose sequences
and optical flows. In 2019 16th Conference on Computer and
Robot Vision (CRV), pages 181–188, 2019. 3
[21] Roman Voeikov, N. Falaleev, and Ruslan Baikulov. Ttnet:
Real-time temporal and spatial video analysis of table tennis.
2020 IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition Workshops (CVPRW), pages 3866–3874,
2020. 2
[22] Xinchao Wang, Vitaly Ablavsky, Horesh Ben Shitrit, and
Pascal Fua. Take your eyes off the ball: Improving ball-
tracking by focusing on team play. Computer Vision and
Image Understanding, 119, 01 2013. 1,2
[23] Mehmet Yakut and Nasser Kehtarnavaz. Ice-hockey puck
detection and tracking for video highlighting. Signal, Image
and Video Processing, 10, 03 2015. 2
[24] A. Yamada, Y. Shirai, and J. Miura. Tracking players and
a ball in video image sequence and estimating camera pa-
rameters for 3d interpretation of soccer games. In Object
recognition supported by user interaction for service robots,
volume 1, pages 303–306 vol.1, 2002. 1,2
[25] X. Yu, C. . Sim, J. R. Wang, and L. F. Cheong. A trajectory-
based ball detection and tracking algorithm in broadcast ten-
nis video. In 2004 International Conference on Image Pro-
cessing, 2004. ICIP ’04., volume 2, pages 1049–1052 Vol.2,
2004. 2
[26] X. Zhang, T. Zhang, Y. Yang, Z. Wang, and G. Wang. Real-
time golf ball detection and tracking based on convolutional
neural networks. In 2020 IEEE International Conference on
Systems, Man, and Cybernetics (SMC), pages 2808–2813,
2020. 1,2
9
