Using Continual Learning on Edge Devices for Cost-Effective, Efficient
License Plate Detection
Reshawn Ramjattan
1
, Rajeev Ratan
2
, Shiva Ramoudith
3
, Patrick Hosein
3
and Daniele Mazzei
1
1
Department of Computer Science, University of Pisa, Largo B. Pontecorvo 3, 56127 Pisa, Italy
2
BPP University, London, U.K.
3
The University of the West Indies, St. Augustine, Trinidad
daniele.mazzei@unipi.it
Keywords:
Deep Learning, Continual Learning, Object Detection, Edge Computing.
Abstract:
Deep learning networks for license plate detection can produce exceptional results. However, the challenge
lies in real-world use where model performance suffers when exposed to new variations and distortions of
images. Rain occlusion, low lighting, glare, motion blur and varying camera quality are a few among many
possible data shifts that can occur. If portable edge devices are being used then the change in location or the
angle of the device also results in reduced performance. Continual learning (CL) aims to handle shifts by
helping models learn from new data without forgetting old knowledge. This is particularly useful for deep
learning on edge devices where resources are limited. Gdumb is a simple CL method that achieves state-of-
the-art performance results. We explore the potential of using continual learning for license plate detection
through experiments using an adapted Gdumb approach. Our data was collected for a license plate recognition
system using edge devices and consists of images split into 3 categories by quality and distance. We evaluate
the application for data shifts, forward/backward transfer, accuracy and forgetting. Our results show that a CL
approach under limited resources can attain results close to full retraining for our application.
1 INTRODUCTION
Machine learning (ML) solutions have seen over-
whelmingly positive results over recent years. Even
so, in real-world use, it is common for ML models
to be exposed to data that is significantly different
from what they were trained on. It would be ideal
if those models could swiftly learn from that data
and improve, but this is an immensely challenging
task. As the model is changed to handle new data,
old but meaningful knowledge would likely be lost
and this is called catastrophic forgetting (Goodfel-
low et al., 2013). Continual learning (CL) is a cat-
egory of ML that strives to address these issues and
has been receiving a noteworthy amount of recent at-
tention (Biesialska et al., 2020; Cossu et al., 2021;
Khetarpal et al., 2020; Delange et al., 2021). CL
addresses the problem of learning from a continuous
stream of data while preserving and extending the ac-
quired knowledge. This can provide significant bene-
fits such as the quick adaptability of models to new
shifts in data distribution, reduced cost in terms of
computing and infrastructure overhead and more ef-
ficient use of data through forward/backward transfer
of knowledge.
These benefits are especially important for the
space of ML on edge devices and have inspired a
specific focus on CL for the edge where comput-
ing power, memory and sometimes even internet con-
nectivity are limited (Demosthenous and Vassiliades,
2021; Kwon et al., 2021; Piyasena et al., 2020; Pelle-
grini et al., 2021). Thus, we saw it fit to explore the
potential of CL for one such “ML on the edge” use
case, license plate detection, using real-world data.
Consider a license plate recognition system for se-
curity, comprising a plate detection model, an opti-
cal character recognition model and software for han-
dling input, orchestrating the models and interacting
with a database. The recognition system runs on a
setup consisting of an Nvidia Jetson Nano module
with a CSI camera mounted onto it and can be used
in many environments that present varying types of
input images. The unique variety of images is diffi-
cult to foresee and can affect the performance of the
license plate detection step. Given the computational
load, and limited computing power coming from a de-
544
Ramjattan, R., Ratan, R., Ramoudith, S., Hosein, P. and Mazzei, D.
Using Continual Learning on Edge Devices for Cost-Effective, Efficient License Plate Detection.
DOI: 10.5220/0011656300003417
In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages
544-550
ISBN: 978-989-758-634-7; ISSN: 2184-4321
Copyright
c
2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
sire to maintain an affordable product, a low-cost CL
approach to quickly respond to the data shifts of a new
environment would be of great benefit.
Moreover, even if a large model by a license
plate recognition product provider was successfully
deployed on an edge-based setup, then, with enough
data, the model could still be trained to handle these
adjustments. However, for a large model to be re-
trained for every new environment with seemingly
minor variations is inconvenient, expensive and can
benefit from continual learning.
Despite the crucial advances made in open-source
tools and community-building for CL (Lomonaco
et al., 2021), it is still a young field. Implement-
ing state-of-the-art approaches on embedded systems,
for a task like object detection with complex mod-
els, training methods and output shapes is no minor
task. Hence, exploring the potential of CL for our use
case through simpler means can provide a motivat-
ing foundation of empirical evidence to drive forward
further work in the space. Moreover, Gdumb (Prabhu
et al., 2020) aligns well with our needs and has shown
results comparable to state-of-the-art methods (Mai
et al., 2022). Gdumb is a simple approach to CL that
has impractical computing and memory demands but
is ideal for benchmarking.
The Gdumb approach greedily stores samples in a
memory buffer as they come and then trains a model
from scratch using the buffer at test time. Our ap-
proach to this style fills the buffer with samples of
a specific environment scenario and uses it for in-
cremental training. In order to produce an upper
bound for comparison, we also adapted this setup with
a fully accumulating buffer that collects all samples
over time for all scenarios. Such experiments will al-
low one to check the model’s ability to learn incre-
mentally and empirically observe the impact of data
changes on the model.
Our work has the following research goals: (1) to
observe the data shift problem for license plate detec-
tion in different environments using real data and (2)
to estimate the practical benefits of using CL to ad-
dress this problem. We aim to meet these goals using
Gdumb-styled experiments. Our contributions, there-
fore, include, showing the impact of small variations
on a modern deep learning model designed for the
edge and exploring the potential of continual learn-
ing to ease adaptations without complete retraining
by looking at a strong benchmark. To the best of our
knowledge, this is also the first evaluation of Gdumb
for continual object detection.
In this paper, we first review the existing relevant
literature. We then describe our base model and the
data and its different scenarios. Afterwards, we put
forward experiments for evaluating the potential of
CL and observing data shift and forgetting. Finally,
we present and discuss the results before concluding.
2 RELATED WORK
Deep learning for edge devices has been well studied
and widely successful (Li et al., 2018; Chen and Ran,
2019; Wang et al., 2020), especially for computer vi-
sion problems through convolutional neural networks
(CNNs) (Huang et al., 2021; Sufian et al., 2021). This
success has naturally extended to the case of license
plate detection on edge devices (Jamtsho et al., 2021;
Alam et al., 2021) through object detection algorithms
based on CNNs such as YOLO (Redmon et al., 2016)
and EfficientDet (Tan et al., 2020). It must be noted
though that these cases train using ideal, close and
clear images of vehicles and plates. Work by (Silva
and Jung, 2018; Xie et al., 2018) examines license
plate detection on images with oblique plates or ones
from various countries, but real-world scenarios can
be even more challenging with varying camera qual-
ities, distant, moving and out-of-focus vehicles, rain
or snow occlusion, image noise issues, over/under-
exposure, motion blur, etc. Furthermore, these works
do not explore a continual learning perspective to ad-
dress the challenges that come with a product that
must perform in many different environments.
There is an extensive number of CL approaches
and methods and they can be loosely categorized
as follows: (1) regularization-based (2) architectural
(3) replay-based (Parisi et al., 2019; Delange et al.,
2021). Gdumb on the other hand is a simple baseline
metric. It greedily stores all samples as the scenar-
ios change and completely retrains the base model on
them. It demands heavy resources and is not suitable
for use in most real-time situations.
However, it provides a useful benchmark compar-
ison for complicated CL techniques. Moreover, stud-
ies have shown this approach outperforms many state-
of-the-art CL methods (Mai et al., 2022; Prabhu et al.,
2020). Our research goals are to show the impact of
catastrophic forgetting in a seemingly easy-to-solve
application and to observe the potential of CL in ad-
dressing it. Thus, we consider experiments around
adapting Gdumb to be a simple and effective way to
evaluate the potential of CL in license plate detection.
Some interesting applications of on-edge CL have
demonstrated its ability to ease the path towards high-
quality performance at low computational cost (Kwon
et al., 2021; Shao et al., 2021). Additionally, recent
investigations of CL for object detection have shown
promising results with the adaptability that can be
Using Continual Learning on Edge Devices for Cost-Effective, Efficient License Plate Detection
545
achieved (Liu et al., 2020; Wang et al., 2021).
While there has been a growing interest in CL ap-
proaches for object detection problems, there is rel-
atively little work done in the area, both in terms of
empirical evaluations and use case applications. So a
simple approach to evaluating the potential benefits of
CL to detection use cases can be of benefit. We eval-
uate such an approach to the problem of license plate
detection under various challenging conditions.
3 EXPERIMENTAL SETUP
3.1 Dataset
The dataset is comprised of 1223 images of vehicles
and license plates. Each image has an associated la-
bel file storing the details of the bounding box loca-
tions for each license plate. The images were gathered
and labelled as part of work done for a license plate
recognition product. They were acquired using var-
ious edge devices, such as smartphones and CCTV
cameras, in three different types of scenarios. Most
of the images were taken from oblique angles. While
there are many possible scenarios, we chose three of
the most likely types to occur. Examples of each are
illustrated in figure 1.
Figure 1: Samples of images under varying conditions.
3.1.1 Scenario 1: High Quality and Close
This subset is made up of 360 images and repre-
sents the ideal situation. License plates and vehicles
are captured from close range using relatively high-
quality cameras that are still feasible for use in an em-
bedded system product.
3.1.2 Scenario 2: Low Quality and Blurred
Scenario 2 focuses on lower-quality input that has
slight motion blur and contains 300 images. Aside
from the affordability and availability of lower-
quality cameras, consider the following example as
an illustration of why this input type might be a com-
mon occurrence. The license plate recognition system
is used for allowing or rejecting vehicles at a point of
entry. To secure the product, it is placed at a high
point, perhaps on a pole as is commonly done for
CCTV cameras. The image is then zoomed in to a
key physical point to capture only the vehicle under
review. With the distance and zoom ability of an af-
fordable camera, the resulting image can be of lower
quality and contain slight motion blur.
3.1.3 Scenario 3: Unisolated and Distant
The third set looks at 563 photos taken from a dis-
tance where the vehicle may be one of many or only
a small fraction of the entire picture. Elaborating on
the prior example, if the same highly placed camera
needs to now cover multiple entry points or is to be
used for traffic analytics at a wider angle, then the ve-
hicle positions can no longer be precisely anticipated.
This results in multiple distant vehicles in a single im-
age, greater skew of plates and generally less isolated
vehicles.
3.2 Base Model
The object detection algorithm chosen for the exper-
iments was YOLOv5 (Jocher et al., 2022) because
of its impressive benchmark results, well-established
community support and, thorough documentation.
The problem of detecting license plates shares under-
lying similarities with problems such as detecting ve-
hicles and recognizing backgrounds. Moreover, our
dataset of scenarios contains a small number of sam-
ples. Hence, we chose to use weights pre-trained on
the COCO dataset (Lin et al., 2014). Since the model
needs to run on an edge device with limited resources,
the YOLOv5s model version was used as its smaller
size was designed for the edge.
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
546
4 METHOD AND RESULTS
We conducted three experiments to observe data dis-
tribution shifts, forget, accuracy, forward transfer and
backward transfer. Each test involved multiple train-
ing steps, with access to different levels of data and
tested against each scenario. At each training step,
the model was trained for 50 epochs after which the
optimal weights were selected. As per the style of
Gdumb, zero hyperparameter tuning was done at any
stage. Due to its ideal characteristics, images from
scenario 1 (high quality and close) were used first for
training at each test.
4.1 Experiment 1
We first established contextual upper-bound metrics
for our models by using all 3 scenarios combined for
training and testing. The YOLOv5s model gave an
end result of 92.6% accuracy.
4.2 Experiment 2
The model was trained on each scenario separately.
That is, the same pre-trained weights were used at
each step with a single scenario’s data, and the result
was evaluated against the test sets of all settings. The
results of this are presented in table 1.
4.3 Experiment 3
The YOLOv5s model was trained on each scenario
incrementally. It was done in the order of scenarios
1 to 3 and each step used the resulting weights of the
prior step. Table 1 shows the results of each step.
4.4 Experiment 4
Lastly, for a greedy approach, we trained the model on
incremental scenarios while keeping old data at each
step, starting from the pre-trained weights. So, first,
it used scenario 1 with the initial pre-trained weights,
then the second step used scenarios 1 and 2 with the
initial pre-trained weights, and the third step used all
three settings with the same initial weights. The re-
sults are shown in table 1.
5 DISCUSSION
The overall accuracy results for our problem in exper-
iment 1 are positive, at 92.6% accuracy, considering
the challenging cases and relatively small dataset.
The remaining experiments are geared towards
evaluating the CL concerns of data distribution shift,
forgetting, and forward and backward transfer. For
the sake of clarity, this section will discuss the results
for each of those individually.
5.1 Data Distribution Shift
Data shifts are the crux of the problem CL tries to
solve. They represent a change that an existing model
was not trained to handle, which translates into poor
performance in production. Therefore, examining
data shifts is crucial in evaluating the applicability of
CL.
From experiment 2, the data shifts can be empiri-
cally seen. We observe differences in accuracy as fol-
lows: 12.3% between the close-set (scenario 1) and
the blurry set (scenario 2), 41.6% from the close set
to the distant set (scenario 3) and lastly, 74.7% be-
tween the blurry and distant sets. These results illus-
trate the difference between the close or blurry sets
compared to the distant sets. Experiments 3 and 4 re-
veal similar evidence. The model trained on clean and
blurry data incrementally with the same weights per-
formed 57.4% worse on distant predictions and the
model trained on close and blurry at the same time
performed 45% worse.
5.2 Forgetting
Catastrophic forgetting is the primary challenge
standing in the path of CLs ability to cope with data
shifts. Thus, an estimation of this issue is also of inter-
est to our evaluation. In experiment 3, weights were
incrementally trained one scenario at a time in the or-
der of close, blurry and distant. At stage 2 when the
blurry set was used, a decrease of 0.5% on the close-
set is seen. Meanwhile, when the distant set was in-
troduced, the blurry and close results dropped by 7%
and 4.8% respectively. These changes due to forget-
ting may seem small in comparison to the data shifts
discussed, but ideally, through CL the model should
retain prior knowledge and even improve with the new
knowledge.
5.3 Backward Transfer
Improving on old samples due to learning from new
ones is also known as backward transfer. In experi-
ment 4, when access to data is unrestrained we can
see small backward transfer improvements on prior
scenarios as new ones are introduced. These improve-
ments are 0.2% on both the close and blurry scenarios.
Using Continual Learning on Edge Devices for Cost-Effective, Efficient License Plate Detection
547
Table 1: Accuracy results for all experiments.
Train
Test
Scenario 1 Scenario 2 Scenario 3 Average
Exp 2: Training from scratch for each scenario
Scenario 1 99.3 87.0 57.7 81.3
Scenario 2 89.1 98.8 24.1 70.7
Scenario 3 92.6 62.3 79.4 78.1
Exp 3: Incrementally training the same weights without storing scenarios
Scenario 1 99.3 87.0 57.7 81.3
Scenario 2 98.0 98.9 41.1 79.3
Scenario 3 93.2 91.9 82.1 89.1
Exp 4: Incrementally training the same weights while storing scenarios
Scenario 1 99.3 87.0 57.7 81.3
Scenario 1 + 2 99.5 98.8 54.1 84.1
Scenario 1 + 2 + 3 99.5 99.0 79.2 92.6
5.4 Forward Transfer
Forward transfer refers to the ability to learn a new
task easier because of learning another task first. If
we compare the model training on the distant setting
first versus training on that set after learning from the
close and blurry settings, we see a 2.7% improvement.
Similarly, a 0.1% increase is observed if the blurry set
is used for training after the close one.
A more noticeable example of forward transfer
is noted when looking at the data distribution shifts.
There is a moderate shift of 12.3% between close and
blurry changes to 9.7% if the blurry set is trained be-
fore the close. Likewise, the vast difference of 74.7%
between distant and blurry goes to 17.1% if the distant
set is seen before the blurry.
5.5 Accuracy
The average accuracy across all scenarios for the dif-
ferent stages of each experiment was also calculated.
The highest 3-setting average for a model trained on
one scenario came from the close set rather than the
distant set which was the most challenging. Whereas
the lowest result came from training on the blurry
only. The best result attained, as one might expect,
came from training on all 3 scenarios with no memory
constraints at 92.6%. However, incrementally train-
ing weights on the 3 settings resulted in an average
accuracy of 89.1%, only 3.5% less, which is a posi-
tive sign for the continual approaches to this problem.
6 CONCLUSIONS
A license plate recognition system in a production en-
vironment can see dynamic changes in its input data.
The ideal scenario of ‘clean’ images can quickly shift
to a more challenging data distribution. By procur-
ing a dataset of 3 likely settings, Gdumb-based tests
allowed us to empirically observe the data shift, for-
getting, accuracy and forward/backward transfer for
our plate detection problem.
The data shifts proved to be a problem that can-
not be ignored with as much as a 74.7% performance
drop. An adapted Gdumb approach to addressing this
resulted in a final average accuracy of 89.1% versus
92.6% from regular training. On the other hand, the
degree of forgetting perceived in comparison to the
backward and forward knowledge transfer observed
was less than ideal. Therefore, we can conclude that
an incremental continual learning approach to license
plate detection shows potential for low-compute on-
edge solutions but still has room for improvement.
In terms of future work, while Gdumb is an excel-
lent benchmark method and has demonstrated similar
or better performance than recent CL approaches, it
is still among the most resource-demanding. Our in-
cremental training approach, with a clear comparison
to unconstrained retraining on full data over time, can
reduce the load of what must be collected on edge de-
vices. Moreover, it shows the accuracy trade-off of
a continual approach that is more cost-effective and
efficient as a solution. A foundation that can be im-
proved upon by leveraging more advanced CL for-
mulations in this complex model setting. Therefore,
comparing the presented results to other strong meth-
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
548
ods like iCARL or Maximal Interfered Retrieval (Re-
buf et al., 2017; Aljundi et al., 2019) will be use-
ful. These methods can then be deployed on the Jet-
son Nano setup itself towards an on-edge solution that
continually adapts with minimal manual intervention.
Furthermore, there have been interesting advances
in the area of Unsupervised Continual Learning (Rao
et al., 2019; He and Zhu, 2021; Bertugli et al., 2020).
Such an approach could not only aid in solving the
issues discussed in this paper but can also allow for
the product using the detection model to be much
more scalable to new environments without manual
labelling. Hence, comparing the results and practi-
cality of an unsupervised method to the incremental
labelling approach is another area for further investi-
gation.
ACKNOWLEDGEMENTS
This research is supported by the Ministry of Uni-
versity and Research (MUR) as part of the PON
2014-2020 “Research and Innovation” resources
Green/Innovation Action DM MUR 1061/2022.
This research was also supported by the Nvidia Aca-
demic Hardware Grant Program.
REFERENCES
Alam, N.-A., Ahsan, M., Based, M. A., and Haider, J.
(2021). Intelligent system for vehicles number plate
detection and recognition using convolutional neural
networks. Technologies, 9(1):9.
Aljundi, R., Belilovsky, E., Tuytelaars, T., Charlin, L., Cac-
cia, M., Lin, M., and Page-Caccia, L. (2019). Online
continual learning with maximal interfered retrieval.
Advances in neural information processing systems,
32.
Bertugli, A., Vincenzi, S., Calderara, S., and Passerini,
A. (2020). Few-shot unsupervised continual
learning through meta-examples. arXiv preprint
arXiv:2009.08107.
Biesialska, M., Biesialska, K., and Costa-juss
`
a, M. R.
(2020). Continual lifelong learning in natural
language processing: A survey. arXiv preprint
arXiv:2012.09823.
Chen, J. and Ran, X. (2019). Deep learning with edge
computing: A review. Proceedings of the IEEE,
107(8):1655–1674.
Cossu, A., Carta, A., Lomonaco, V., and Bacciu, D. (2021).
Continual learning for recurrent neural networks: an
empirical evaluation. Neural Networks, 143:607–627.
Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia,
X., Leonardis, A., Slabaugh, G., and Tuytelaars, T.
(2021). A continual learning survey: Defying forget-
ting in classification tasks. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence.
Demosthenous, G. and Vassiliades, V. (2021). Contin-
ual learning on the edge with tensorflow lite. arXiv
preprint arXiv:2105.01946.
Goodfellow, I. J., Mirza, M., Xiao, D., Courville, A., and
Bengio, Y. (2013). An empirical investigation of
catastrophic forgetting in gradient-based neural net-
works. arXiv preprint arXiv:1312.6211.
He, J. and Zhu, F. (2021). Unsupervised continual learning
via pseudo labels. arXiv preprint arXiv:2104.07164.
Huang, Z., Yang, S., Zhou, M., Gong, Z., Abusorrah, A.,
Lin, C., and Huang, Z. (2021). Making accurate ob-
ject detection at the edge: review and new approach.
Artificial Intelligence Review, pages 1–30.
Jamtsho, Y., Riyamongkol, P., and Waranusast, R. (2021).
Real-time license plate detection for non-helmeted
motorcyclist using yolo. Ict Express, 7(1):104–109.
Jocher, G., Chaurasia, A., Stoken, A., Borovec, J.,
NanoCode012, Kwon, Y., TaoXie, Fang, J., imyhxy,
Michael, K., Lorna, V, A., Montes, D., Nadar, J.,
Laughing, tkianai, yxNONG, Skalski, P., Wang, Z.,
Hogan, A., Fati, C., Mammana, L., AlexWang1900,
Patel, D., Yiwei, D., You, F., Hajek, J., Diaconu, L.,
and Minh, M. T. (2022). ultralytics/yolov5: v6.1 -
TensorRT, TensorFlow Edge TPU and OpenVINO Ex-
port and Inference.
Khetarpal, K., Riemer, M., Rish, I., and Precup, D. (2020).
Towards continual reinforcement learning: A review
and perspectives. arXiv preprint arXiv:2012.13490.
Kwon, Y. D., Chauhan, J., Kumar, A., Hui, P., and Mascolo,
C. (2021). Exploring system performance of contin-
ual learning for mobile and embedded sensing appli-
cations. arXiv preprint arXiv:2110.13290.
Li, H., Ota, K., and Dong, M. (2018). Learning iot in
edge: Deep learning for the internet of things with
edge computing. IEEE network, 32(1):96–101.
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,
Ramanan, D., Doll
´
ar, P., and Zitnick, C. L. (2014).
Microsoft coco: Common objects in context. In Euro-
pean conference on computer vision, pages 740–755.
Springer.
Liu, X., Yang, H., Ravichandran, A., Bhotika, R., and
Soatto, S. (2020). Continual universal object detec-
tion. arXiv preprint arXiv:2002.05347.
Lomonaco, V., Pellegrini, L., Cossu, A., Carta, A., Graffi-
eti, G., Hayes, T. L., De Lange, M., Masana, M., Pom-
poni, J., Van de Ven, G. M., et al. (2021). Avalanche:
an end-to-end library for continual learning. In Pro-
ceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, pages 3600–3610.
Mai, Z., Li, R., Jeong, J., Quispe, D., Kim, H., and San-
ner, S. (2022). Online continual learning in image
classification: An empirical survey. Neurocomputing,
469:28–51.
Parisi, G. I., Kemker, R., Part, J. L., Kanan, C., and
Wermter, S. (2019). Continual lifelong learning with
neural networks: A review. Neural Networks, 113:54–
71.
Using Continual Learning on Edge Devices for Cost-Effective, Efficient License Plate Detection
549
Pellegrini, L., Lomonaco, V., Graffieti, G., and Maltoni,
D. (2021). Continual learning at the edge: Real-
time training on smartphone devices. arXiv preprint
arXiv:2105.13127.
Piyasena, D., Thathsara, M., Kanagarajah, S., Lam, S. K.,
and Wu, M. (2020). Dynamically growing neural net-
work architecture for lifelong deep learning on the
edge. In 2020 30th International Conference on Field-
Programmable Logic and Applications (FPL), pages
262–268. IEEE.
Prabhu, A., Torr, P. H., and Dokania, P. K. (2020). Gdumb:
A simple approach that questions our progress in con-
tinual learning. In European conference on computer
vision, pages 524–540. Springer.
Rao, D., Visin, F., Rusu, A., Pascanu, R., Teh, Y. W., and
Hadsell, R. (2019). Continual unsupervised represen-
tation learning. Advances in Neural Information Pro-
cessing Systems, 32.
Rebuffi, S.-A., Kolesnikov, A., Sperl, G., and Lampert,
C. H. (2017). icarl: Incremental classifier and rep-
resentation learning. In Proceedings of the IEEE con-
ference on Computer Vision and Pattern Recognition,
pages 2001–2010.
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.
(2016). You only look once: Unified, real-time object
detection. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 779–
788.
Shao, Y., Zhao, Y., Yu, F., Zhu, H., and Fang, J. (2021). The
traffic flow prediction method using the incremental
learning-based cnn-ltsm model: the solution of mobile
application. Mobile Information Systems, 2021.
Silva, S. M. and Jung, C. R. (2018). License plate detec-
tion and recognition in unconstrained scenarios. In
Proceedings of the European conference on computer
vision (ECCV), pages 580–596.
Sufian, A., Alam, E., Ghosh, A., Sultana, F., De, D., and
Dong, M. (2021). Deep learning in computer vision
through mobile edge computing for iot. In Mobile
Edge Computing, pages 443–471. Springer.
Tan, M., Pang, R., and Le, Q. V. (2020). Efficientdet: Scal-
able and efficient object detection. In Proceedings
of the IEEE/CVF conference on computer vision and
pattern recognition, pages 10781–10790.
Wang, F., Zhang, M., Wang, X., Ma, X., and Liu, J. (2020).
Deep learning for edge computing applications: A
state-of-the-art survey. IEEE Access, 8:58322–58336.
Wang, J., Wang, X., Shang-Guan, Y., and Gupta, A. (2021).
Wanderlust: Online continual object detection in the
real world. In Proceedings of the IEEE/CVF Interna-
tional Conference on Computer Vision, pages 10829–
10838.
Xie, L., Ahmad, T., Jin, L., Liu, Y., and Zhang, S. (2018).
A new cnn-based method for multi-directional car li-
cense plate detection. IEEE Transactions on Intelli-
gent Transportation Systems, 19(2):507–517.
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
550