An Evaluation of General-Purpose Optical Character Recognizers and

Digit Detectors for Race Bib Number Recognition

Modesto Castrill

on-Santana

1 a

, David Freire-Obreg

1 b

, Daniel Hern

andez-Sosa

1 c

Oliverio J. Santana

1 d

, Francisco Ortega-Zamorano

2 e

, Jos

e Isern-Gonz

alez

1 f

and

Javier Lorenzo-Navarro

1 g

SIANI, Universidad de Las Palmas de Gran Canaria, Spain

Universidad de M

alaga, Spain

Keywords:

Race Bib Number Recognition, OCR, General Object Detection.

Abstract:

Bib numbers are used in mass competitions to identify participants, especially in long-distance races where

runners commonly wear tags to verify that they pass mandatory checkpoints. In this paper, we delve deeper

into the use of existing computer vision techniques for recognizing the digits present in bib numbers. Our anal-

ysis of bib recognition involves evaluating OCRs (Optical Character Recognition) techniques and a YOLOv7

digit detector on two public datasets: RBNR and TGCRBNW. The results reveal that the former scenario is

solvable, while the latter presents extremely in-the-wild challenges. However, the ﬁndings suggest that more

than relying solely on RBN for runner identiﬁcation, other appearance-based cues, e.g., clothing and acces-

sories, may be required due to various circumstances, such as occlusion or incomplete bib recognition. In any

case, all those cues do not necessarily imply that the same person is wearing the RBN across the competition

track, as they are not biometric traits.

1 INTRODUCTION

Since the early 90s, timing systems have been an es-

sential tool for organizers of massive running events

to keep track of runners at speciﬁc locations along the

course track. However, these systems do not control

the real presence of the runner, but rather the presence

of the tag they carry, and strictly, they do not verify

the participant’s identity. This has led to incorrect

classiﬁcation results and insurance policy problems

for organizers. The computer vision community has

started to address the participant identiﬁcation prob-

lem, mostly focusing on solving the Racing Bib Num-

ber (RBN) automatic recognition problem (Ben-Ami

et al., 2012), i.e., recognizing the precise number ex-

hibited in the bib. However, this solution suffers from

the same drawbacks as tag-based systems. Biometrics

https://orcid.org/0000-0002-8673-2725

https://orcid.org/0000-0003-2378-4277

https://orcid.org/0000-0003-3022-7698

https://orcid.org/0000-0001-7511-5783

https://orcid.org/0000-0002-4397-2905

https://orcid.org/0000-0001-5830-7732

https://orcid.org/0000-0002-2834-2067

could be an alternative, but it has yet to be applied

more extensively in this challenging scenario.

In this paper, we do not make use of biometrics.

Instead we explore using state-of-the-art computer vi-

sion techniques to recognize RBNs in running compe-

titions, considering real-life datasets used in the liter-

ature to which we have been granted access. The sce-

nario comprises signiﬁcant challenges, such as varia-

tions in illumination conditions, runner’s image reso-

lution, body pose, and image sharpness, among oth-

ers. RBN recognition is a valid modality in a ﬂexible

multimodal approach for runner recognition, which

should also integrate biometric traits, such as face

and body appearance or gait recognition. However,

all those mentioned modalities present the risk of be-

ing unavailable at a particular moment. The mid-term

expectations offer exciting possibilities for improving

participant identiﬁcation and can be a game-changer

for the running event industry.

In summary, the main contribution of this paper

is evaluating existing RBN recognition benchmarks,

deﬁning a new digit-based instead of RBN-based met-

ric focus, and proposing relevant features for real-

world or in-the-wild benchmarks.

910

Castrillón-Santana, M., Freire-Obregón, D., Hernández-Sosa, D., Santana, O., Ortega-Zamorano, F., Isern-González, J. and Lorenzo-Navarro, J.

An Evaluation of General-Purpose Optical Character Recognizers and Digit Detectors for Race Bib Number Recognition.

DOI: 10.5220/0012562400003654

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

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 910-917

ISBN: 978-989-758-684-2; ISSN: 2184-4313

2 RELATED WORK

In sports, computer vision is an active ﬁeld tackled by

recent surveys (Mendes-Neves et al., 2023) and well-

established workshops such as CVsports and MM-

Sports. Several studies have primarily focused on

analyzing the individual athlete’s movements to ﬁnd

ways of improvement (Li and Zhang, 2019; Thomas

et al., 2017) or studying team sports to collect statisti-

cal data, increasing the team’s performance (Liu and

Bhanu, 2019; Thomas et al., 2017). Despite the active

research in computer vision for sports, our focus is di-

verse, as in this paper we are just interested in iden-

tifying mass runner competition participants by their

RBN. This section summarizes existing proposals for

recognizing runners using computer vision.

2.1 RBN Detection and Recognition

As previously discussed, RBN recognition is the

main approach for participant identiﬁcation in run-

ning competitions, with most previous research pri-

marily focusing on marathon-like scenarios (Kamlesh

et al., 2017). Although such competitions are now or-

ganized everywhere, the amount of publicly available

data is limited. The RBN dataset most widely used as

a benchmark in the literature (Ben-Ami et al., 2012)

includes 217 high-quality images captured by profes-

sional photographers, with 290 annotated RBNs. The

proposed approach detects faces in the images and

estimates the RBN Region of Interest (RoI) to de-

tect text using the Stroke Width Transform (SWT)

and Optical Character Recognition (OCR). Annotated

RBNs correspond to those individuals whose face was

correctly located by a face detector, being eligible to

estimate an RBN RoI. This means that other RBNs

may be present in the dataset images.

The approach described in (Boonsim, 2018) also

adopts the initial face detection step but focuses on

morphological operations applied to the torso area of

the detected runners. In (Shivakumara et al., 2017),

upper body detection is performed and the runner is

segmented using GrabCut to restrict the text detection

and recognition area. This approach is evaluated on

their own dataset of 212 images captured in marathon

competitions. The same research group has more re-

cently described a uniﬁed method for detecting text

from marathon runners and sports players in video,

achieving high accuracy in text extraction in that sce-

nario (Nag et al., 2020). In (de Jes

us and Borges,

2018), the focus is on text detection, skipping the pre-

vious face/body detection, which may undoubtedly be

affected by the runner’s poses, reducing the number of

possible RBNs.

In the past few years, the community has explored

using deep learning for RBN recognition. The ap-

proach in (Wong et al., 2019) uses the convolutional

network YOLO (You Only Look Once) (Bochkovskiy

et al., 2020) for RBN detection and a convolutional

recurrent neural network (CRNN) for text recogni-

tion. A similar approach is described in (Apap and

Seychel, 2019), where a convolutional neural net-

work (CNN) is used for RBN detection, followed by a

CRNN for classiﬁcation. Both approaches report re-

sults on the different literature datasets. The YOLO

architecture is also adopted in (Carty et al., 2021) for

RBN detection, applying a Tesseract OCR for the bib

number recognition, which achieves promising results

on a private dataset. In (Nag et al., 2019), runners

are detected using a Single Shot Multibox Detector

(SSD), and their body parts are extracted to locate

the RBN. Finally, a CRNN is applied for classiﬁca-

tion. In (Kamlesh et al., 2017), a re-identiﬁcation ap-

proach based on RBN recognition is applied, using

TextBoxes (Minghui Liao and Bai, 2018) for RBN de-

tection and a CRNN for classiﬁcation. This approach

is evaluated on their non-public dataset.

RBN digit recognition may be considered a sub-

problem of the overall problem of text recognition

in natural scene images and video, which usually

needs the previous step to spot the pieces of text to

be located and later decode each separate character.

The literature describes two main approaches for text

recognition: 1) the combination of a recurrent neural

network (RNN) on top of a CNN and a connection-

ist temporal classiﬁcation (CRNN+CTC), and 2) the

combination of a CNN and an attention RNN (Luo

et al., 2019). Tesseract, one of the most widely used

OCRs, adopted the ﬁrst approach, i.e., the combi-

nation of RNN on top of CNN and CTC for RBN

recognition, as mentioned earlier. The use of the

two-steps method applied to RBN recognition is pre-

sented in (Ivarsson and Mueller, 2019), reporting the

performance in datasets captured during marathon-

like events. The FOTS (Fast Oriented Text Spotting)

method proposed in (Liu et al., 2018) simultaneously

detects and recognizes text, in contrast to the previous

methods, which perform these tasks separately. This

is done because there is a high correlation between

the features used for detection and recognition.

2.2 Biometrics

Although body information has been used to deter-

mine the Region of Interest (RoI) for RBNs through

face, upper body, or body part location. The use of fa-

cial and body recognition or clothing appearance, to

determine the identity of runners has rarely been em-

An Evaluation of General-Purpose Optical Character Recognizers and Digit Detectors for Race Bib Number Recognition

911

ployed. One of the few references that combine fa-

cial appearance with RBN recognition for improving

identiﬁcation performance is (Wro

nska et al., 2017).

This multimodal strategy may be well-suited for the

task at hand, as both face and RBN occlusions are

frequently present, and the target pose may introduce

difﬁculties for both face and RBN detection.

While the marathon-like scenario, with daylight

conditions, reduces the presence of wild variations

in the problem dataset, the re-identiﬁcation in the

ultra-running scenario proposed in (Penate-Sanchez

et al., 2020) aims to deﬁne a new benchmark for

state-of-the-art re-identiﬁcation approaches based on

body/clothing appearance. The ﬁnal evaluation of

top-ranked re-identiﬁcation approaches suggests rel-

evant challenges in the scenario. More recently, Choi

et al. utilized gait traits, extracting arm swing features

from the silhouette, to overcome issues such as RBN

recognition, face occlusion, and clothing appearance

similarity (Choi et al., 2021).

3 METHODOLOGY

As mentioned earlier, this paper focuses not on bib

detection but on bib number recognition. Therefore,

in the experimental evaluation, we assume that the

bounding box of the bib number has already been ob-

tained using one of the aforementioned detection ap-

proaches. The main emphasis of this study is on text

recognition, precisely digit recognition. In this sense,

the present study allows us to evaluate the ideal bib

number recognition performance, which may be con-

sidered as an upper bound of the reachable perfor-

mance in a real application scenario where ﬁrstly the

bib detection step is required. Previous literature on

bib recognition has utilized various OCR techniques

or deep learning strategies for character classiﬁcation.

General object detectors are, in principle, less spe-

cialized for text recognition tasks than OCRs, as they

are designed as ﬂexible solutions for any object de-

tection task. However, two remarkable features of

bib number text might make general object detectors

suitable for the task: 1) the RBN text to recognize is

composed of digits, i.e., just ten classes or categories

are needed to be trained, and 2) the RBN text is not

manuscript, i.e., not written by hand, and thus there

is great homogeneity in the typography. In this sense,

we are interested in exploring whether a ﬁne-tuned

general object detector, trained with non manuscript

digits, may outperform OCRs in the task of RBN

recognition.

In this sense, we will evaluate as baseline two

freely available OCRs in the experimental evaluation

below, focusing on recognizing only digits whenever

possible. Speciﬁcally, we have selected two open-

source alternatives: Tesseract

, which has been used

for this speciﬁc problem in the literature, and Easy-

OCR

, which integrates CRAFT (Baek et al., 2019)

for character region cropping and CRNN for text

recognition.

Furthermore, to test the hypothesis that an object

detector trained to detect digits improves OCR perfor-

mance, we have adopted the YOLO (You Only Look

Once) architecture, speciﬁcally YOLOv7 (Wang

et al., 2023) using a digit dataset. For this aim,

we have adopted the Street View House Numbers

(SVHN) dataset (Netzer et al., 2011), which is com-

posed of only street numbers captured in varying

real-world conditions sharing similar features with

RBNs, see Figure 1. The ground truth dataset con-

tains bounding boxes per digit in the street number.

This strategy is based on reported experiences by the

community training a digit recognizer with SVHN

with RetinaNet (Lin et al., 2017) and YOLOv4

for

digit spotting within the same scenario.

Even if the chosen scenario of application is dif-

ferent, we have adopted just the SVHN training sub-

set for training, not fusing it with the rest of available

samples in the SVHN data collection. We launched

a 200-epoch training with a learning rate of 0.01 in

a GTX 3080 with a batch size of 32 and an image

width of 640. The default YOLOv7 data augmenta-

tion strategies were adopted, except for deactivating

the possibility of augmenting data by ﬂipping the im-

age, since most digits are not symmetric.

The ﬁne-tuned YOLOv7 detector will provide a

collection of digit-bounding boxes. Assuming only

the bib number area is processed, any detected digit

will be considered belonging to the bib number. We,

therefore, compound the recognized bib number by

ﬁrst sorting the digit bounding boxes along the x-axis;

see Figure 3.

4 EXPERIMENTAL EVALUATION

The experimental evaluation focuses on the two pre-

viously mentioned approaches for recognizing RBNs.

Firstly, we investigate the usage of general-purpose

OCRs, such as Tesseract and EasyOCR. Secondly,

we ﬁne-tune the general object detection architecture

YOLOv7 using SVHN annotated street number sam-

ples to create a dedicated digit detector and classiﬁer.

https://github.com/tesseract-ocr/tesseract

https://github.com/JaidedAI/EasyOCR

https://github.com/penny4860/retinanet-digit-detector

https://github.com/Lwhieldon/BibObjectDetection

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

912

Figure 1: Samples from SVHN dataset (Netzer et al., 2011). Similarly to RBNs, both possibilities, i.e. clearer and darker

fonts, are present in the data collection.

Figure 2: Upper row: samples from RBNR dataset (Ben-Ami et al., 2012). Bottom row: samples from TGCRBNW

dataset (Hern

andez-Carrascosa et al., 2020).

Figure 3: Three digits are detected: 4, 2, and 8. Their re-

spective x-coordinates are sorted to compose the recognized

bib number: 428.

4.1 Datasets

Although various datasets are referred to in the liter-

ature for detecting and/or recognizing race bib num-

bers (RBNs) in running competitions, most are con-

sidered private, and we have not obtained permis-

sion to include them in the experimental evalua-

tion. Therefore, we assess the different approaches

using the only publicly available datasets acces-

sible to us: RBNR (Ben-Ami et al., 2012) and

TGCRBNW (Hern

andez-Carrascosa et al., 2020).

Both datasets consist of images of long-distance

runners, but RBNR comprises photographs of

marathon runners, while TGCRBNW contains frames

extracted from videos of ultra-distance trail runners.

Their main distinction is that TGCRBNW comprises

images captured in low-light conditions. Further-

more, as shown in Figure 2, the RBNR dataset sam-

ples have larger and different fonts. The authors

of TGCRBNW (Hern

andez-Carrascosa et al., 2020)

claim that despite containing only a single font, their

dataset is more extensive and more challenging for

detecting and recognizing race bib numbers. In fact,

the RBNs are annotated even if they are not easily

recognized by humans in the image, as seen in the

bottom row of Figure 2. We will evaluate this claim

in the experiments below.

An Evaluation of General-Purpose Optical Character Recognizers and Digit Detectors for Race Bib Number Recognition

913

Figure 4: Upper row: cropped samples from RBNR dataset (Ben-Ami et al., 2012). Bottom row: cropped samples from

TGCRBNW dataset (Hern

andez-Carrascosa et al., 2020).

4.2 RBN Recognition

As mentioned earlier, the experimental evaluation fo-

cuses on RBN recognition. It skips the RBN detection

step, as both datasets provide ground truth in rectan-

gular bounding boxes around the RBNs. Before lo-

cating and recognizing the text, we decided to enlarge

the bounding boxes by 50% of their dimensions in

width and height. This preprocessing was adopted as,

in some cases, the RBN bounding box has been de-

ﬁned very tightly, and some bib digits do not appear

whole in the cropped area. Examples of the resulting

samples for both datasets are shown in Figure 4.

After enlarging the annotated bounding boxes, we

ﬁrst evaluate the performance of our approach on the

RBNR dataset using the same metrics as the origi-

nal dataset authors deﬁned in (Ben-Ami et al., 2012).

Those metrics are adopted to provide with the preci-

sion the ratio of correctly recognized RBNs out of the

total number of recognized RBNs, and with the recall

the ratio of correctly recognized RBNs out of the total

number of annotated RBNs. The reader must observe

that in their work, Ben-Ami et al. assume that detect-

ing a face triggers the RBN RoI estimation. There-

fore, the metrics do not consider the RBN detection

but the full recognition of the RBN digits. This means

that, for the ﬁrst RBN present in Figure 4, it will be

considered a correct recognition only if the system

outputs 3628. The computation of those metrics re-

quired the number of correctly recognized bibs or true

positives (TP) when all its digits were identiﬁed, the

number of incorrectly recognized bibs or false posi-

tives (FP), when at least one but not all the digits were

identiﬁed, and the number of non-classiﬁed bib num-

bers or false negatives (FN), those without any digit

located. The recall (R), precision (P), and F-score (F)

are deﬁned respectively as:

P =

T P

T P + FP

R =

T P

T P + FN

F = 2 ·

P · R

P + R

(1)

The results achieved for RBNR dataset are sum-

marized in Table 1. We compare the performance of

our approach, which uses the ﬁne-tuned YOLOv7-

based digit detector, with the mentioned OCRs and

the latest published result for this dataset that we

are aware of (Nag et al., 2020), which refers to the

work (Bartz et al., 2018).

Among the OCRs, Tesseract produced signiﬁ-

cantly poor results, but EasyOCR provided fairly

competitive results compared to the latest reported

state-of-the-art for RBNR (Bartz et al., 2018), al-

though still lower. However, the results achieved by

the speciﬁcally trained digit detector signiﬁcantly out-

performed any other approach applied to this dataset.

Table 1: Bibs recognition results for RBNR.

Approach P R F

EasyOCR 0.67 0.54 0.60

Tesseract 0.33 0.05 0.08

See (Bartz et al., 2018) 0.76 0.77 0.77

YOLOv7 0.91 0.91 0.91

With those results, we decided to explore the er-

rors of the YOLOv7 digit detector, focusing speciﬁ-

cally on FPs, to understand why they occur. As men-

tioned above, FPs are outputs that do not match the

annotated RBN completely. Upon observing the FPs

in the evaluation, we noticed that in most cases, even

when the entire bib number is not recognized, most

detected digits are correctly identiﬁed, as shown in

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

914

Figure 5: Partially recognized bib numbers, just three digits of four were detected by the digit detector in both samples. The

previously adopted metric considers both cases as FPs. A metric evaluating the digit recognition performance provides a

better insight into the approach validity.

Figure 6: Images that reported a digit FP according to RBNR ground annotation. Upper row: the YOLOv7 digit detector

confused a 3521 with 3527. Second row: bibs annotated as 80653, 891, 359, and 764 were recognized as 80635, 0891, 3594,

and 3764, respectively. Indeed, those recognized bibs must be considered correct.

Figure 5. Fixing a detection conﬁdence threshold of

0.5, we found that in both images, out of the eight dig-

its present, only two of them were not located. As the

bib number is not correctly matched, they are counted

as FPs. That consideration is not fair to establish a

real usability of the digit detector. Therefore, we pro-

pose a novel metric approach to evaluate the perfor-

mance more precisely for the RBN recognition prob-

lem: considering digit recognition instead of bib num-

ber recognition. By adopting this evaluation strategy,

the values for precision (P), recall (R), and F-score (F)

for the YOLOv7 ﬁne-tuned digit detector signiﬁcantly

increase to 0.992, 0.975, and 0.984, respectively.

These results suggest that the RBNR dataset is rel-

atively simple for current general object detectors. An

example of the type of FNs in the digit detector is vis-

ible in the previously mentioned Figure 5. As for the

still remaining digit FPs, only ﬁve bib numbers were

reported to have digit FPs. However, upon closer in-

spection, it was revealed that only one of them was a

true digit FP, while the rest correspond to annotation

errors, as seen in Figure 6. This means that the actual

value of precision (P) would be indeed higher, indi-

cating that the RBNR dataset is certainly trivial for

current general object detectors.

Now that we have established that RBNR is a rel-

atively simple dataset for existing tools, we evaluated

the best approaches for bib number recognition on

TGCRBNW. Before discussing the results, it is im-

portant to note that TGCRBNW contains 3223 an-

notated bibs captured in both low-light and daylight

conditions, with a split of 69% and 31%, respectively.

Those images correspond to video frames, thus there

is no human photographer behind the camera adjust-

ing the camera settings for each sample. Additionally,

it is worth mentioning that the bib number is not nec-

essarily visible in the images, as the human annotators

had access to the whole runner tracklet to annotate the

identity, i.e., his/her bib number. The results of bib

number recognition are summarized in Table 2.

Indeed, those results are signiﬁcantly worse than

those achieved for the RBNR dataset, suggesting

that the TGCRBNW dataset is considerably more

challenging. A ﬁrst impression suggests that the

YOLOv7-based recognizer has a lower precision (P),

i.e., it exhibits a higher number of false positives

(FPs). However, if, similarly to the consideration

made for RBNR, we adopt a digit-based metric in-

An Evaluation of General-Purpose Optical Character Recognizers and Digit Detectors for Race Bib Number Recognition

915

stead of a bib number-based metric, the reported pre-

cision (P), recall (R), and F-score (F) for the YOLOv7

detector are 0.776, 0.779, and 0.778, respectively,

which suggests promising performance also in severe

conditions. We remind again that among the 3223 bib

numbers contained in TGCRBNW, only 1000 were

captured in daylight. The reported values, as seen in

Table 3, suggest a slightly easier scenario with day-

light illumination conditions; however, there is still

much room for improvement.

Table 2: Bibs recognition results for TGCRBNW.

Approach P R F

EasyOCR 0.259 0.002 0.004

YOLOv7 0.216 0.047 0.077

Table 3: Digits recognition results for TGCRBNW includ-

ing overall and daylight and nightlight splits metrics.

TGCRBNW split P R F

Complete 0.776 0.779 0.777

Nightlight 0.714 0.714 0.714

Daylight 0.825 0.831 0.827

5 CONCLUSIONS

In this paper, we analyzed RBN recognition using

two available datasets: RBNR and TGCRBNW. To

achieve this, we evaluated two well-known general-

purpose OCRs and a speciﬁc digit detector trained

with samples from a different scenario, speciﬁcally

street numbers, using a general object detection ar-

chitecture. After an initial evaluation of RBNR, we

deﬁned a new digit-based metric instead of an RBN-

based one to get more precise evidence of the par-

tially recognized RBNs. The results suggest that

the scenario deﬁned by datasets similar to RBNR

is now solvable. However, the scenario presented

by TGCRBNW poses a real in-the-wild benchmark,

likely due to challenging features such as illumina-

tion conditions (nightlight, shadows, etc.), RBN reso-

lution, and motion blur, among others.

In any case, relying solely on RBN for recognition

is not enough due to the presence of occlusions, which

are common in such scenarios. Integrating other cues,

such as biometric traits, might help achieve coherent

recognition of speciﬁc runners, especially if a video

stream is available instead of single frames. However,

this concept is beyond the scope of the present paper.

ACKNOWLEDGEMENTS

This work is partially supported by the the Span-

ish Ministry of Science and Innovation under

project PID2021-122402OB-C22 and by the ACIISI-

Gobierno de Canarias and European FEDER funds

under projects ProID2021010012, ULPGC Facilities

Net, and Grant EIS 2021 04

REFERENCES

Apap, A. and Seychel, D. (2019). Marathon bib num-

ber recognition using deep learning. In 11th Inter-

national Symposium on Image and Signal Processing

and Analysis (ISPA), pages 21–26, Dubrovnik, Croa-

tia. IEEE.

Baek, Y., Lee, B., Han, D., Yun, S., and Lee, H. (2019).

Character region awareness for text detection. In Pro-

ceedings of the IEEE Conference on Computer Vision

and Pattern Recognition, pages 9365–9374.

Bartz, C., Yang, H., and Meinel, C. (2018). See: To-

wards semi-supervised end-to-end scene text recogni-

tion. Proceedings of the AAAI Conference on Artiﬁcial

Intelligence, 32(1).

Ben-Ami, I., (Basha), T. D., and Avidan, S. (2012). Racing

bib number recognition. In British Machine Vision

Conference, pages 1–10, Surrey, UK. British Machine

Vision Association.

Bochkovskiy, A., Wang, C.-Y., and Liao, H.-Y. M. (2020).

YOLOv4: Optimal speed and accuracy of object de-

tection. arXiv preprint arXiv:2004.10934.

Boonsim, N. (2018). Racing bib number localization on

complex backgrounds. WSEAS Transactions on Sys-

tems and Control, 13:226–231.

Carty, G., Raja, M. A., and Ryan, C. (2021). Running to get

recognised. In Thampi, S. M., Krishnan, S., Hegde,

R. M., Ciuonzo, D., Hanne, T., and Kannan R., J.,

editors, Advances in Signal Processing and Intelligent

Recognition Systems, pages 3–17, Singapore. Springer

Singapore.

Choi, Y., Napolean, Y., and van Gemert, J. C. (2021). The

arm-swing is discriminative in video gait recognition

for athlete re-identiﬁcation. In IEEE International

Conference on Image Processing.

de Jes

us, W. M. and Borges, D. L. (2018). An improved

stroke width transform to detect race bib numbers.

In Proceedings of the Mexican Conference on Pat-

tern Recognition, pages 267–276, Puebla, Mexico.

Springer.

Hern

andez-Carrascosa, P., Penate-Sanchez, A., Lorenzo-

Navarro, J. F. O., and David; Castrill

on-Santana, M.

(2020). TGCRBNW: A dataset for runner bib num-

ber detection (and recognition) in the wild. In Pro-

ceedings International Conference on Pattern Recog-

nition, Milan, Italy.

Ivarsson, E. and Mueller, R. M. (2019). Racing bib num-

ber recognition using deep learning. In Proceedings of

the 25th Americas Conference on Information Systems

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

916

(AMCIS), Canc

un, Mexico. Association for Informa-

tion Systems.

Kamlesh, P. X., Yang, Y., and Xu, Y. (2017). Person re-

identiﬁcation with end-to-end scene text recognition.

In Chinese Conference on Computer Vision, pages

363–374, Tianjin, China. Springer.

Li, G. and Zhang, C. (2019). Automatic detection tech-

nology of sports athletes based on image recognition

technology. EURASIP Journal on Image and Video

Processing, 2019(1):15.

Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Doll

ar, P.

(2017). Focal loss for dense object detection. In In

IEEE International Conference on Computer Vision

(ICCV).

Liu, H. and Bhanu, B. (2019). Pose-guided R-CNN for jer-

sey number recognition in sports. In Conference on

Computer Vision and Pattern Recognition Workshops

(CVPRW), pages 2457–2466, Long Beach, CA, USA.

IEEE.

Liu, X., Liang, D., Yan, S., Chen, D., Qiao, Y., and Yan,

J. (2018). FOTS: Fast oriented text spotting with a

uniﬁed network. In IEEE Conf. on Computer Vision

and Pattern Recognition.

Luo, C., Jin, L., and Sun, Z. (2019). Moran: A multi-object

rectiﬁed attention network for scene text recognition.

Pattern Recognition, 90:109–118.

Mendes-Neves, T., Meireles, L., and Mendes-Moreira, J.

(2023). A survey of advanced computer vision tech-

niques for sports.

Minghui Liao, B. S. and Bai, X. (2018). TextBoxes++: A

single-shot oriented scene text detector. IEEE Trans-

actions on Image Processing, 27(8):3676–3690.

Nag, S., Ramachandra, R., Shivakumara, P., Pal, U., Lu,

T., and Kankanhall, M. (2019). Crnn based jersey-bib

number/text recognition in sports and marathon im-

ages. In International Conference on Document Anal-

ysis and Recognition (ICDAR), pages 1149–1156,

Sydney, Australia. IEEE.

Nag, S., Shivakumara, P., Pal, U., Lu, T., and Blumenstein,

M. (2020). A new uniﬁed method for detecting text

from marathon runners and sports players in video (pr-

d-19-01078r2). Pattern Recognition, 107:107476.

Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and

Ng, A. Y. (2011). Reading digits in natural images

with unsupervised feature learning. In NIPS Workshop

on Deep Learning and Unsupervised Feature Learn-

ing.

Penate-Sanchez, A., Freire-Obreg

on, D., Lorenzo-Meli

an,

A., Lorenzo-Navarro, J., and Castrill

on-Santana, M.

(2020). TGC20ReId: A dataset for sport event re-

identiﬁcation in the wild. Pattern Recognition Letters,

138:355–361.

Shivakumara, P., Raghavendra, R., Qin, L., B.Raja, K., Luc,

T., and Pal, U. (2017). A new multi-modal approach to

bib number/text detection and recognition in marathon

images. Pattern Recognition, 61:479–491.

Thomas, G., Gade, R., Moeslund, T. B., Carr, P., and Hilton,

A. (2017). Computer vision for sports: Current ap-

plications and research topics. Computer Vision and

Image Understanding, 159:3 – 18.

Wang, C.-Y., Liao, H.-Y. M., and Yeh, I.-H. (2023). Design-

ing network design strategies through gradient path

analysis. Journal of Information Science and Engi-

neering.

Wong, Y. C., Choi, L. J., Singh, R. S. S., Zhang, H., and

Syafeeza, A. R. (2019). Deep learning based rac-

ing bib number detection and recognition. Jorda-

nian Journal of Computers and Information Technol-

ogy (JJCIT), 5(3):(3):181–194.

Wro

nska, A., Sarnacki, K., and Saeed, K. (2017). Athlete

number detection on the basis of their face images. In

Proceedings International Conference on Biometrics

and Kansei Engineering, pages 84–89, Kyoto, Japan.

IEEE.

An Evaluation of General-Purpose Optical Character Recognizers and Digit Detectors for Race Bib Number Recognition

917