Journal of Physics:

Conference Series

PAPER • OPEN ACCESS You may also like

- Investigation on the Effect of the Feature
Traffic sign detection based on classic visual Extraction Backbone for Small Object
Segmentation using Fully Convolutional
recognition models Neural Network in Traffic Signs Application
H M Elhawary, M I Shapiaib and A
Elfakharany
To cite this article: Jirui Yang 2023 J. Phys.: Conf. Ser. 2646 012024 - Attention feature fusion network for small
traffic sign detection
Miaozhi Wu, Jingmin Yang, Wenjie Zhang
et al.

- Traffic sign detection method based on

View the article online for updates and enhancements. Faster R-CNN
Linxiu Wu, Houjie Li, Jianjun He et al.

This content was downloaded from IP address 173.212.198.70 on 07/11/2024 at 13:30

International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

Traffic sign detection based on classic visual recognition

models

Jirui Yang
College of ENGINEERING, COMPUTING AND CYBERNETICS, The Australian
National University, Canberra, Australian

[email protected]

Abstract. This paper aims to assess the effectiveness of various object detection high-level
architectures, including Faster R-CNN, R-FCN, SSD, and YOLO, in recognizing traffic signs.
Since traffic sign recognition is a critical element of driver-assistance and autonomous driving
systems, this study explores how five distinct architectures for feature extracting (ResNet V1 50,
ResNet V1 101, Inception V2, Inception-ResNet-V2, and Darknet-19) can enhance the
performance of these meta-architectures. To evaluate the models, the authors fine-tuned a pre-
trained object recognition algorithm on the German Traffic Sign Detection Benchmark (GTSDB)
dataset. While the Faster R-CNN Inception Resnet v2 proved to be the most accurate among the
models, the unbalanced distribution of mandatory, prohibitory, and danger signs in the GTSDB
dataset may have caused biased results in object detection. Overall, the result obtained from this
study can help improve the accuracy of driver-assistance and autonomous driving systems by
providing insights into the performance of various object detection models in recognizing traffic
signs.

Keywords: Object detection, Traffic sign detection.

1. Introduction
"The TSR (traffic sign recognition) system plays a crucial role in advanced driver-assistance systems
(ADASs) and auto-driving systems (ADSs), typically comprising two components: traffic sign detection
(TSD) and traffic sign recognition" [1]. TSD aims to locate the position of the traffic sign, while TSR
classifies the category of the object in the position. Therefore, the accuracy of TSD is critical to the
overall accuracy of TSR.
Traffic signs are an essential form of instruction equipment on roads that offer drivers and pedestrians
guidance under different conditions, including day and night [2]. They are crucial in providing
information about road conditions and restrictions, particularly in vision-based vehicle guidance
systems. Due to their unique colors and shapes that stand out from the environment, traffic signs are
easy for humans to understand. The RGB camera and LIDAR are the two primary sensors used to
perceive the road environment. While the RGB camera observes the color, structure, and visual
characteristics of traffic signs, it does not provide 3D geometric information, which the LIDAR sensor
can offer. Combining the two sensors enables accurate detecting and classifying traffic sign even in
complex environments [3].
There exist multiple classical leading algorithms that leverage convolutional neural networks for

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

feature extraction. These algorithms include Faster R-CNN [4], R-FCN [5], SSD [6], and YOLO V2 [7].
Two categories of deep learning-based object detection detectors exist: the two-stage detector and the
one-stage detector. The two-stage detector filters the bounding box at a coarse level. With the filtered
bounding boxes, detectors with two stages, for example, Faster R-CNN and R-FCN are more accurate
at the expense of speed, while the one-stage detectors like SSD and YOLO V2 attempt to retrieve the
object at once [8]. When assessing the suitability of these detectors for certain tasks, engineers not only
consider accuracy metrics but also other metrics like computational complexity and memory
consumption. In the case of autonomous vehicles, decision speed is crucial, and the trade-off between
accuracy and decision speed must be evaluated carefully. In autonomous vehicle scenarios, the emphasis
is more on decision speed.
Object detection models are typically evaluated using different challenges, including Microsoft
COCO [9] and VOC Pascal [10], which comprise a variety of entities, such as humans, trains, and
vehicles, for the models to recognize. This paper examines four models that utilize distinct feature
extractors pre-learned on the Microsoft COCO dataset. Our focus is on evaluating the models'
performance by examining their meta-architecture and feature extractor. The paper presents four key
contributions: (1) a summary of the five different models and their meta-architecture and feature
extractors, (2) an assessment of the models' mean average precision (mAP) on the Microsoft COCO and
VOC PASCAL databases, as well as their ability to detect traffic signs at different scales, (3) a
comparison of the models' accuracy, revealing an unbalanced distribution problem in the GTSDB dataset
that causes biased detection, and (4) a comparison of the models' accuracy, demonstrating that the Faster
R-CNN Inception Resnet v2 achieves the highest level of accuracy.
The rest is separated into several sections. Section 2 introduces the dataset used for model pretraining,
while Section 3 presents details on previously leading models about traffic sign detection and
classification. Section 4 shows and analyzes the results generated from the Microsoft COCO and VOC
PASCAL databases. Finally, Section 5 draws conclusions from the experiment and discusses the
implications of the results regarding detecting and recognizing traffic signs.

2. Dataset
This study utilized the German Traffic Sign Detection Benchmark (GTSDB) dataset [11] for conducting
the experiments described in this paper. The dataset comprises 900 images containing a total of 1206
traffic signs, which were split into training and testing set. The training set consists of 600 images
containing 846 traffic signs, while the testing set consists of 300 images containing 360 traffic signs.
The dataset images were captured from diverse traffic scenes with various road conditions, weather
conditions, and lighting conditions. The traffic signs has four groups: prohibitory, mandatory, danger,
and other. The "other" category was not taken into account during the experiments because of its small
number. The distribution of the remaining three categories was uneven. The training set had 396
prohibitory signs (59.5%), 156 danger signs (23.4%), and 114 mandatory signs (17.1%). Similarly, the
test set included 161 images of prohibitory signs, 63 images of danger signs, and 49 images of mandatory
signs.

3. Diverse traffic sign detection model

This section provides a detailed introduction to four CNN-based model architectures and six feature
extractors. Models are pre-learned on the Microsoft COCO database [9] and are used for direct
evaluation, without considering computational complexity. The pre-trained models classify traffic signs
into three classes: prohibitory, mandatory, and danger. Table 1 shows the combination between meta-
architecture and feature extraction. The reason for not pairing the meta-architecture with the feature
extractor in a pairwise manner is that each feature extractor requires customization and adaptation before
being integrated into a meta-architecture. Due to its computational complexity, only this pre-trained
model is selected for evaluation.

2
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

Tabel 1. Combination of Meta-architecture and Feature extractor.

Faster R-CNN R-FCN SSD YOLO
Resnet v1 101 √
Resnet v1 50 √
Inception v2 √
Inception Resnet v2 √
Darknet-19 √

3.1. Model architecture

This section will present and compare the characteristics of every meta-architecture, namely Faster R-
CNN, R-FCN, SSD, and YOLO V2.

3.1.1. Faster R-CNN. Faster R-CNN is an exemplar of a object detector with two stages that utilizes a
"coarse to fine" strategy. The method initially filters bounding boxes in a coarse manner, followed by
extracting classification features from these filtered boxes to achieve greater accuracy. The Region
Proposal Network (RPN) was proposed by Faster R-CNN as a fully convolutional network (FCN) [12]
capable of end-to-end learning for producing detection proposals. This approach is novel and not present
in previous two-stage object detection models. To generate detection proposals, the RPN slides windows
on the convolutional feature map, which can be a network like ResNet or Inception. At each window
position, the RPN generates a group of object candidates, each having an objectness score that measured
the likelihood of the proposal containing the object of interest or simply being a background.
The Region Proposal Network (RPN) generates multiple region proposals for each anchor box. An
anchor box refers to a set of reference boxes, and the RPN predicts proposals by considering the position
of each anchor box. The quantity of detection proposals generated for every location is determined by
the number of these anchors. To accommodate objects of different sizes and shapes, anchor boxes have
multiple scales and aspect ratios. These boxes form a pyramid of anchors that serve as a reference for
bounding box regression. To avoid redundancy in detections, proposed regions undergo non-maximum
suppression (NMS) based on their objectness scores, which filters out lower-scoring detections when
multiple detections overlap. The detection network then receives proposal regions ranked in the top-N
for object categorization and the estimation of bounding box positions.

3.1.2. R-FCN. The Region-Based Fully Convolutional Network (R-FCN) is an object detection model
that operates in two stages and has a comparable architecture to Faster R-CNN. R-FCN, like Faster R-
CNN, incorporates proposal of regions and classification of regions. The feature map produced by the
fully convolutional neural network (CNN) of Faster R-CNN is shared with both RPN and R-FCN in R-
FCN's architecture, and is processed by the region proposal network to produce candidate regions. In R-
FCN, parameterized layers are trained based on the entire image, which reduces computations on each
region. Additionally, the author of the paper demonstrates the trade-off between transition invariance
and variance through their experiments.

3.1.3. SSD. Apart from the detectors like Faster R-CNN and R-FCN with two stages, there is a detector
named the single-shot multi-box detector (SSD) with one stage. Unlike its two-stage counterparts, SSD
employs a feed-forward network to generate a set of fix-size bounding boxes along with their respective
objectness scores directly. For each detected object, SSD only requires a ground truth bounding box and
generates a feature map pyramid, with a predetermined set of bounding boxes assigned to each feature
cell, which are then used to calculate the confidence scores and shape offset for all object categories.
The SSD model adheres to a conventional design for producing top-notch image classification
results. However, it omits the classification layer and incorporates extra layers for capturing crucial
features, which will be elaborated on in subsequent discussion.

3
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

The SSD model comprises multiple convolutional feature layers, which have the ability to generate a
range of feature maps at different scales, forming a pyramid of features. To obtain a classification score
or an offset in shape associated with the coordinates of the default box., a small 3x3xP kernel is applied
to each feature layer, which has an input size of m x n with p channels. Each feature map cell is assigned
a group of default detecting boxes, which are akin to anchor boxes, and their positions are fixed to their
corresponding cells. These default boxes are selected carefully to handle various tasks that involve
different target object scales. In summary, the SSD architecture is capable of predicting at multiple
scales.

3.1.4. YOLO V2. Firstly, the original YOLO algorithm [13] will be discussed, which takes 448×448
fixed input size. The input image is separated into 13×13 grid cells and estimate whether the cell contains
an object with a confidence score. Bounding box regression is then applied to generate bounding boxes
around the objects with a metric measuring their position and objectness score. The Intersection over
Union (IOU) filter out bounding boxes that only partially contain an object. Last, non-maximum
suppression (NMS) merges all bounding boxes containing the same object into a single bounding box
with the highest confidence score.
Next, the updates of YOLO v2 on the original YOLO will be introduced. YOLOv2 improved the
original YOLO algorithm by introducing anchor boxes for better object localization and adding batch
normalization for faster training. Batch normalization is significantly effective in improving
convergence [14]. In order to obtain accurate bounding boxes, a specially designed distance metric is
employed to utilize k-means clustering on the detection rectangles of the training set, as opposed to
depending on manually chosen priors. (Eq. (1))
d(box, centroid) = 1 − IOU(box, centroid) (1 )
To avoid model instability, which can lead to anchor boxes being placed at arbitrary positions in the
image irrespective of the location where the box is predicted, the model predicts location coordinates
respect to the grid cell's location apart from predicting offsets.
The Darknet-19 which will be discussed in Section 3.2.4 is a classification model recently created to
serve as the basis for YOLOv2. During the training process, the model is customized and refined for
object detection. To enhance detection accuracy, conventional augmentation methods like random crops,
rotations, hue and saturation adjustments, and exposure shifts are utilized [7]. Additionally, the network
incorporates training across multiple scales by selecting a different image size every 10 batches. This
approach ensures that the network learns features from various resolutions.

3.2. Feature extractor

In this section, some classic convolutional neural networks (Resnet V1 50 / Resnet V1 101, Inception
V2, Inception Resnet V2 and Darknet-19) are adopted as feature extractors for high-level features.

3.2.1. Resnet V1 50 / Resnet v1 101. ResNet V1 50 and ResNet V1 101 are deep neural networks that
achieved success in the ILSVRC and COCO 2015 competitions, including tasks such as segmentation,
detection, and classification [15]. In the Faster R-CNN and R-FCN frameworks, ResNet is used as a
feature generator, which is split into two parts. The first part extracts features for the Region Proposal
Network (RPN), while the latter segment extracts features for the box classifier.
Both two Resnets are in the structure of four blocks. The first three blocks exact a feature map that
is shared by RPN and region of interest (RoI) layer in both Faster R-CNN and R-FCN while the last
layer in third block predicts region proposal. Last, the box classifier feature is extracted by the last block.

3.2.2. Inception V2. Inception V2 [13] set the state-of-the-art performance in both the ILSVRC2014
classification and detection challenges. One of the key features of Inception V2 is the use of inception
units, which are composed of a group of parallel convolutional sorts with different kernel sizes, and a

4
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

pooling layer. These units allow the network to increase its width and depth without increasing
computational complexity, which can be beneficial for training and inference efficiency.
As discussed in Section 3.1.3, SSD adopts a one-stage detection approach where region proposals
are not needed. Therefore, instead of splitting the Inception V2 model into distinct modules for PRN
feature extraction and box classifier feature extraction, the entire model is employed as the feature
extractor.

3.2.3. Inception Resnet V2d. For the Inception Resnet V2 [16], in the experiment, only the Faster R-
CNN high-level architecture is combined with this feature encoder. Faster R-CNN, being a detector with
two stages is partitioned into two components. The computational efficiency of the inception unit is
combined with the optimization effect by residual connect.
The Inception-ResNet-V2 consists of multiple modules, including the Inception-ResNet-B and
Inception-ResNet-C modules, which are used for feature extraction for the Region Proposal Network
(RPN) and box classifier, respectively. The Inception-ResNet-B module operates on a 17x17 grid, while
the Inception-ResNet-C operates on an 8x8 grid.

3.2.4. Darknet-19. As explained in Section 3.1.4, Darknet-19 is a deep learning model initially designed
for object classification, but it has been adapted to perform object detection through modifications.
The Darknet-19 model follows a similar architecture as the VGG models, where the pooling increases
the channel’s number by twice, and a 3x3 kernel is utilized. Then the global average pooling is applied
along with 1×1 kernels to compress the space dimension of the feature map produced by the 3×3
convolutions. Also, Batch normalization is applied to improve the model regularization, stabilize
training and accelerate convergence.
Initially, the model is trained with the 224×224 input image size on ImageNet. Then the model is
fine-tuned in the larger dataset for a few epochs with a relatively low learning rate which gives the model
better performance on the higher-resolution images.
Finally, the last convolutional layer in the pre-trained feature extractor is replaced by several fully
connected layers to generate region proposals, and then those proposals are fed into a separate head
network which consists of three 3x3 convolutional layers with 1024 filters each, with two layers
afterwards: one component for identifying object categories and another component for predicting
object locations through bounding boxes. The the final convolutional layer's output size is decided by
the quantity of object categories that the model is designed to recognize.

4. Result
Here, the performance of each meta-architecture with diverse feature extractors that are introduced in
Section 3 will be analyzed. Measurement on mean average precision (mAP) and Intersection of Union
(IoU) in the experiment will be compared. Before comparing each metric, definitions are introduced in
the following.

4.1. Accuracy measurement

4.1.1. Mean Average Precision (mAP). The performance of object detection systems is evaluated using
the mean average precision (mAP) obtained from PASCAL VOC 2010 [10].To introduce mAP, average
precision (AP) will be conducted. AP summarises the precision/recall recurve to one scalar value which
it is the area under the precision/recall curve (Eq (2)).
1
𝐴𝑣𝑒𝑟𝑎𝑔𝑒 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 (𝐴𝑃) = ∫ 𝑝(𝑟)𝑑𝑟 (2 )
𝑟=0

where 𝑝(𝑟) denotes the precision measured at a given recall level r.

The function of precision corresponding to recall r is defined as the highest precision attainable at or
beyond a recall level of r (ℎ ≥ 𝑟).

5
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

𝑝(𝑟) = max 𝑝(ℎ) (3 )

ℎ:ℎ>𝑟

Finally, mAP is obtained by taking the mean value of each category's average precision.

4.1.2. Intersection of Union (IoU). The accuracy of predictions is assessed using a metric denoted as
Intersection of Union (IoU). The IoU metric computes the percentage of the overlap between the
predicted (𝐵𝑝 ) and the ground truth bounding box (𝐵𝑔𝑡 ) over the union area of both boxes (𝐵𝑔𝑡 ∪ 𝐵𝑝 )
(Eq 4.). A threshold of 0.5 is set to determine the correctness of a prediction. If the Intersection over
Union (IoU) score is higher than 0.5, the prediction is deemed accurate, while those with an IoU less
than 0.5 are false positives. Objects in the ground truth data that do not have matching predicted
bounding boxes are considered false negatives, while redundant object detections are recognized as the
false positives.
𝑎𝑟𝑒𝑎(𝐵𝑝 ∩ 𝐵𝑔𝑡 )
𝐼𝑜𝑈 = (4 )
𝑎𝑟𝑒𝑎(𝐵𝑝 ∪ 𝐵𝑔𝑡 )

4.2. Evaluation on AP over different IoU

(a) Faster R-CNN Inception Resnet v2 (b) Faster R-CNN Resnet v1 50

(c) R-FCN Resnet v1 101 (d) SSD Inception v2

6
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

(e) YOLO V2
Figure 1. (a-e) Average precision of different model of different IoU threshold.
Average precision is a good estimator of the accuracy of the model prediction and the IoU detects
the bias between prediction and ground truth. In Fig. 1., there are five scatter plots on average precision
with different IoU thresholds in different categories and one vertical line at IoU = 0.75 is drawn to
compare different model precision at high accuracy standards. The average precision is obtained through
the pre-learned and fine-tuned model on the GTSDB dataset that is tested on the Microsoft COCO
dataset.
From the five plots above, the average precision on mandatory traffic signs is significantly lower
than the other two categories while the prohibitory has the highest AP at IoU = 0.5 which is the threshold
used for training. With the Inception Resnet v2 as feature encoder, in Faster R-CNN the average
precision for prohibitory even reaches 1.00. This result is highly correlated with the bias distribution of
three categories of traffic signs described in Section 2.

4.3. Evalution on mAP over diverse model

Figure 2. Accuracy in different size on 5 different detectors.

7
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

To be noticed, the size of the traffic sign influence negatively on the accuracy. The ground truth
traffic images are split into three parts according to the width of the image. There are two thresholds to
split the images. The first threshold is a width of 32 pixels, which separates the small group of 89 images
from the other two groups. Any image in the dataset with a width smaller than 32 pixels will be assigned
to the small group. The second threshold is a width of 45 pixels, which separates the medium group of
93 images from the large group of 91 images. Any image in the dataset with a width between 32 pixels
and 45 pixels (inclusive) will be assigned to the medium group, while any image with a width greater
than 45 pixels will be assigned to the large group. Since the size of the three groups are relatively large
and the difference between each other is small. The effect of the group size on the mAP can be ignored.
From Fig. 2., all the detectors perform worse in the small group than in other groups. This may be
caused by the feature extractor that is pre-trained by the Microsoft COCO dataset which the traffic sign
size is large and located in the middle while in reality the traffic sign usually appears at the edge of the
image and small. Then one horizontal line is drawn to compare the mAP. It is clear that the SSD
Inception v2 and YOLO v2 perform significantly worse than the other three models while their
performance on medium or large image sizes is similar to the other three models. This drawback can be
observed in the original paper [6,7]. For instance, the SSD detector, trained on Microsoft COCO and
finetuned on PASCAL VOC2012 and PASCAL VOC2007 dataset, has a lower performance on the small
object like plants (57.2%AP) and bottles (60.8% AP) while on the larger objects, the performance is
better, such as aero (90.7%AP) and horse (89.0%AP).
Finally, regardless of the recognizing speed, on the performance of accuracy, the Faster R-CNN with
Inception Resnet v2 and R-FCN with Resnet v1 50 have similar performance and work best within the
five detectors.

5. Conclusion
In this paper, an experiment on comparison of five traffic sign deep-learning-based detectors with
different feature extractors is present. The accuracy is the main aspect compared in this paper. The
models compared were pre-learned from the Microsoft COCO dataset and then optimized using the
GTSDB dataset with the aim of traffic sign detection and classification. The traffic signs were
categorized as three groups: mandatory, prohibitory, and danger signs.
Accuracy which is measured by mAP is calculated through the dataset PASCAL VOC 2010. With
Inception Resnet V2 feature extractor, the Faster R-CNN gain the mAP of 95.77% at top level and the
R-FCN with Resnet v1 101 extractor has a similar mAP of 95.17% as the first rank detector. Then, the
SSD with Resnet v1 50 and YOLO v2 have poor performance in detecting small objects which is the
drawback shown in the original paper. Finally, the bias of the GTSDB dataset is observed that causes all
five detectors to perform worse in mandatory traffic signs than the other two.
In future work, other deep learning-based architectures in recent years that have good performance
in object detection can be researched such as YOLO v7 [17] and adjustment to the traffic sign detection
will be applied for analysis.

References
[1] Chunsheng Liu, Shuang Li, Faliang Chang, & Yinhai Wang. (2019). Machine Vision Based
Traffic Sign Detection Methods: Review, Analyses and Perspectives. IEEE Access, 7, 86578–
86596. https://doi.org/10.1109/ACCESS.2019.2924947
[2] de la Escalera, Moreno, L. E., Salichs, M. A., & Armingol, J. M. (1997). Road traffic sign
detection and classification. IEEE Transactions on Industrial Electronics (1982), 44(6), 848–
859. https://doi.org/10.1109/41.649946
[3] Zhou, & Deng, Z. (2014). LIDAR and vision-based real-time traffic sign detection and
recognition algorithm for intelligent vehicle. 17th International IEEE Conference on
Intelligent Transportation Systems (ITSC), 578–583.
https://doi.org/10.1109/ITSC.2014.6957752
[4] Shaoqing Ren, Kaiming He, Girshick, R., & Jian Sun. (2017). Faster R-CNN: Towards

8
International Conference on Software Engineering and Machine Learning IOP Publishing
Journal of Physics: Conference Series 2646 (2023) 012024 doi:10.1088/1742-6596/2646/1/012024

Real-Time Object Detection with Region Proposal Networks. IEEE Transactions on P

attern Analysis and Machine Intelligence, 39(6), 1137–1149. https://doi.org/10.1109/TP
AMI.2016.2577031
[5] Dai, Li, Y., He, K., & Sun, J. (2016). R-FCN: Object Detection via Region-based Fully
Convolutional Networks. arXiv.org.
[6] Liu, Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., & Berg, A. C. (2016). SSD: Single
Shot MultiBox Detector. Computer Vision – ECCV 2016, 21–37. https://doi.org/10.1007/978-
3-319-46448-0_2
[7] Redmon, & Farhadi, A. (2016). YOLO9000: Better, Faster, Stronger. arXiv.org.
[8] Zou, Chen, K., Shi, Z., Guo, Y., & Ye, J. (2023). Object Detection in 20 Years: A Survey.
Proceedings of the IEEE, 1–20. https://doi.org/10.1109/JPROC.2023.3238524
[9] Lin, Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (n.d.).
Microsoft COCO: Common Objects in Context. In Computer Vision – ECCV 2014 (pp. 740–
755). Springer International Publishing. https://doi.org/10.1007/978-3-319-10602-1_48
[10] Everingham, Van Gool, L., Williams, C. K. I., Winn, J., & Zisserman, A. (2010). The Pascal
Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2),
303–338. https://doi.org/10.1007/s11263-009-0275-4
[11] Houben, Stallkamp, J., Salmen, J., Schlipsing, M., & Igel, C. (2013). Detection of traffic signs in
real-world images: The German traffic sign detection benchmark. The 2013 International Joint
Conference on Neural Networks (IJCNN), 1–8. https://doi.org/10.1109/IJCNN.2013.6706807
[12] Shelhamer, Long, J., & Darrell, T. (2017). Fully Convolutional Networks for Semantic
Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4), 640–
651. https://doi.org/10.1109/TPAMI.2016.2572683
[13] Redmon, Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified, Real-
Time Object Detection. 2016 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 779–788. https://doi.org/10.1109/CVPR.2016.91
[14] Ioffe, & Szegedy, C. (2015). Batch Normalization: Accelerating Deep Network Training by
Reducing Internal Covariate Shift. arXiv.org.
[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. (2016). Deep Residual Learning for
Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 770–778. https://doi.org/10.1109/CVPR.2016.90
[16] Szegedy, Ioffe, S., Vanhoucke, V., & Alemi, A. (2016). Inception-v4, Inception-ResNet and the
Impact of Residual Connections on Learning. arXiv.org.
[17] Chien-Yao, Bochkovskiy, A., & Hong-Yuan, M. L. (2022). YOLOv7: Trainable bag-of-freebies
sets new state-of-the-art for real-time object detectors. arXiv.org.