APSIPA Transactions on Signal and Information Processing, 2024, 13, e29

This is an Open Access article, distributed under the terms of the Creative Commons
Attribution licence (http:// cr eativecommons.or g/ licenses/by-nc/ 4.0/ ), which permits
unrestricted re-use, distribution, and reproduction in any medium, for non-commercial use,
provided the original work is properly cited.

Overview Paper
YOLOv1 to YOLOv10: The Fastest and
Most Accurate Real-time Object
Detection Systems
Chien-Yao Wang1,2* and Hong-Yuan Mark Liao1,2,3
1
Institute of Information Science, Academia Sinica, Taiwan
2
National Taipei University of Technology, Taiwan
3
National Chung Hsing University, Taiwan

ABSTRACT
This is a comprehensive review of the YOLO series of systems.
Different from previous literature surveys, this review article re-
examines the characteristics of the YOLO series from the latest
technical point of view. At the same time, we also analyzed
how the YOLO series continued to influence and promote real-
time computer vision-related research and led to the subsequent
development of computer vision and language models. We take
a closer look at how the methods proposed by the YOLO series
in the past ten years have affected the development of subsequent
technologies and show the applications of YOLO in various fields.
We hope this article can play a good guiding role in subsequent
real-time computer vision development.

Keywords: YOLO, computer vision, real-time object detection

1 Introduction

Object detection is a fundamental computer vision task that can support a

wide range of downstream tasks. For example, it can be used to assist instance
∗ Corresponding author: [email protected].

Received 17 August 2024; revised 20 September 2024; accepted 09 October 2024

ISSN 2048-7703; DOI 10.1561/116.20240058
© 2024 C.-Y. Wang and H.-Y. M. Liao
2 Wang and Liao

segmentation, multi-object tracking, behavior analysis and recognition, face

recognition, etc. Therefore, it has been a popular research topic in the
past few decades. In recent years, due to the popularity of mobile devices,
the ability to perform real-time object detection on the edge has become a
necessary component for various real-world applications. Tasks belonging
to such applications include autonomous driving, industrial robots, identity
authentication, smart health care, visual surveillance, etc. Among the many
real-time object detection algorithms, the YOLO (You Only Look Once) series
(from v1 to v10) [2, 26, 37, 57, 82–84, 102, 105, 109] developed in recent years
is particularly outstanding. It has greatly and extensively affected various
research in the field of computer vision. This paper will review the YOLO
family of technologies and their impact on the development of contemporary
real-time computer vision systems.
The first deep learning-based method to achieve breakthrough success
in the field of object detection was R-CNN [24]. R-CNN is a two-stage
object detection method, which divides the object detection process into two
stages: object proposal generation and object proposal classification. What
R-CNN does is to first use selective search [99], which is commonly used in
image processing, to extract proposals. At this stage, CNN is only used as a
feature extractor to extract features of proposals. As for the recognition part,
SVM [49] is used. The subsequent development of Fast R-CNN [23] and Faster
R-CNN [85] respectively used SPPNet [48] to accelerate feature extraction and
proposed Region Proposal Networks to gradually convert object detection into
the end-to-end format. YOLO [82] was proposed by Joseph Radmon in 2015.
It uses per gird prediction to complete object detection in one step. This is a
groundbreaking approach that brings the field of real-time object detection
to a whole new level. The subsequent development of classic one-stage object
detection systems includes SSD [67], RetinaNet [62], FCOS [98], etc.
Although the one-stage object detection method can detect objects in real
time, there is still a gap in accuracy from the two-stage object detection method.
The one-stage detection systems such as RetinaNet [62] and YOLOv3 [84]
have made significant progress on this issue, and they both achieved sufficient
accuracy. YOLO series have become the most preferred method for industry
and all academia and research centers that require real-time object analysis. In
2020, scaled-YOLOv4 [104] further designed a very effective object detection
model scaling method. For the first time, the accuracy of the one-stage
object detection method in the field of general object detection surpassed all
contemporary two-stage object detection methods, and this achievement also
led to many subsequent related research based on YOLO series methods.
In addition to object detection, YOLO series is also used in other computer
vision fields as a basis for developing real-time systems. Currently in instance
segmentation, pose estimation, image segmentation, 3D object detection, open-
vocabulary object detection, etc., YOLO still plays a pivotal role in real-time
systems.
YOLOv1 to YOLOv10 3

In this review article, we will introduce the following issues in order:

• Introduction to the YOLO series methods and important related litera-

ture.
• The impact of the YOLO family of methods on the contemporary field
of computer vision.
• Important methods for applying YOLO in different computer vision
fields.

2 YOLO series

YOLO is synonymous with the most advanced real-time object detector of our
time. The biggest difference between YOLO and traditional object detection
systems is that it abandons the previous two-stage object detection method that
requires first finding the locations where objects may be located in the image,
and then analyzing the content of these locations individually. YOLO proposes
a unified one-stage object detection method, and this method is streamlined
and efficient, which makes YOLO widely used in various edge devices and
real-time applications. Next we will introduce several representative YOLO
versions, as listed in Table 1, and this literature review is different from the
previous ones. We will put our emphasis on the state-of-the-art object detection
methods and review the advantages and disadvantages of these methods.

2.1 YOLO (YOLOv1)

YOLOv1: Redmon et al. [82] was the first one who proposed the one-stage
object detector in 2015, and the architecture of YOLOv1 is illustrated in
Figure 1. As shown in the figure, an input image first passes through CNN
for feature extraction, and then passes through two fully connected layers to
obtain global features. Then, the aforementioned global features are reshaped
back to the two-dimensional space for per grid prediction. YOLOv1 has the
following important features:
One-Stage Object Detector. As shown in Figure 1, YOLOv1 directly
classifies each grid of feature map, and also predicts B bounding boxes. Each
bounding box will predict the object center (bx , by ), object size (bw , bh ), and
object score (bobj ) respectively. The one stage prediction method does not need
to rely on the selective search that must be executed in the object proposal
generation stage, which can avoid missed detections caused by insufficient
manual design clues. In addition, the one-stage method can avoid the large
number of parameters and calculations generated by fully connected layers
4 Wang and Liao

Table 1: List of YOLO series papers and publication time.

Model Date Publication Citation

YOLO (YOLO1) 2015.06 CVPR 2016 [82]
YOLO9000 (YOLOv2) 2016.12 CVPR 2017 [83]
YOLOv3 2018.04 arXiv 2018 [84]
Gaussian YOLOv3 2019.03 ICCV 2019 [14]
YOLOv4 2020.01 arXiv 2020 [2]
YOLOv5 2020.05 – [26]
Scaled-YOLOv4 2020.06 CVPR 2021 [104]
YOLOv5 r1.0 2020.06 – [27]
PP-YOLO 2020.07 arXiv 2020 [68]
YOLOv5 r2.0 2020.07 – [28]
YOLOv5 r3.0 2020.08 – [29]
Scaled-YOLOv4 P6 2020.08 CVPR 2021 [104]
YOLOv5 r3.1 2020.10 – [30]
YOLOv5 r4.0 2021.01 – [31]
PP-YOLOv2 2021.04 arXiv 2021 [51]
YOLOv5 r5.0 2021.04 – [32]
YOLOR 2021.05 JISE 2023 [110]
YOLOX 2021.07 arXiv 2021 [22]
YOLOv5 r6.0 2021.10 – [33]
PP-PicoDet 2021.11 arXiv 2021 [119]
YOLOv5 r6.1 2022.02 – [34]
PP-YOLOE 2022.03 arXiv 2022 [116]
YOLOv6 2022.06 – [73]
YOLOv7 2022.07 CVPR 2023 [105]
YOLOv5 r6.2 2022.08 – [35]
YOLOv6 2.0 2022.09 arXiv 2022 [57]
YOLOv7 AF 2022.11 CVPR 2023 [105]
DAMO-YOLO 2022.11 arXiv 2022 [117]
YOLOv5 r7.0 2022.11 – [36]
YOLOv8 2023.01 – [37]
YOLOv6 3.0 2023.01 arXiv 2023 [56]
YOLOv6 4.0 2023.04 OpenReview 2023 [58]
YOLO-NAS 2023.05 – [93]
Gold-YOLO 2023.09 NeurIPS 2023 [103]
YOLOv8 r1.0 2024.01 – [38]
YOLOv9 2024.02 ECCV 2024 [109]
YOLOv8 r2.0’ 2024.04 – [40]
YOLOv10 2024.05 NeurIPS 2024 [102]
YOLOv8 r2.0 2024.06 – [39]
YOLOv8 r3.0 2024.09 – [41]

in the second stage, and it can avoid the irregular operations required when
connecting two stages of RoI operations. Therefore, YOLO’s design can
capture features and make predictions more timely and effectively. Below we
will take a closer look at the most important concepts in YOLOv1, which are
anchor-free bounding box regression, IoU-aware objectness, and global context
features.
YOLOv1 to YOLOv10 5

Figure 1: Architecture of YOLOv1. The architecture of YOLOv1 allows users to extract

global features from fully-connected layers, predict B anchor-free bounding boxes at each
position, and use IoU-aware objectness score as the filtering basis.

Anchor-free Bounding Box Regression. In Equation 1, YOLOv1 directly

predicts the proportion of the length and width of the object in the entire
image. Although the anchor-free method requires optimization of a large
dynamic range of length and width, which makes convergence more difficult,
it also has the advantage of being able to predict some special examples more
accurately because it is not restricted by anchors.

bx = tx + cx ,
by = ty + cy ,
(1)
bw = tw 2 ,
bh = th 2
IoU-aware Objectness. In order to more accurately measure the quality
of bounding box prediction, the method proposed by YOLOv1 is to predict
the IoU value between a certain bounding box and the assigned ground-truth
bounding box, and use this as the soft label of the objectness predicted by
IoU-aware branch. Finally, the confidence score of bounding box is determined
by the product of objectness score and classification probability.
Global Context Feature. To ensure that a grid doesn’t only see the
local feature and cause prediction errors, YOLOv1 uses fully connected layer
to retrieve global context features. In such a design, no matter what the
underlying CNN architecture is, each grid can see a sufficient range of features
to predict the target object during prediction. Compared with fast R-CNN,
this design effectively reduces background error by more than half.

2.2 YOLO9000 (YOLOv2)

In addition to proposing many insightful new methods, Redmon and Farhadi
[83] also integrate various existing techniques. They designed an object detector
that combines high accuracy and speed, as shown in Figure 2. They converted
the entire object detection architecture to full convolutional network. They
6 Wang and Liao

Figure 2: Architecture of YOLOv2. The architecture of YOLOv2 is to extract mix-resolution

features by passthrough layer, and then predict the objects corresponding to B anchors for
each position.

then combined high-resolution and low-resolution features, and finally use

anchor-based for prediction. Due to its simple input and output formats,
YOLOv2 is still one of the mainstream object detection methods commonly
used in maintenance and development of many industrial scenes, especially on
low-end devices with very limited computing resources. Below we will discuss
the most essential parts of YOLOv2 respectively regarding dimension cluster,
direct location prediction, fine-grained feature, resolution calibration, and joint
training with WordTree issues.
Dimension Cluster. YOLOv2 proposed to use IoU distance as the basis to
k-means clustering on ground-truth bounding boxes to obtain anchors. On
the one hand, the anchor obtained by using dimension cluster can avoid the
original manually set aspect ratio, which is difficult to learn the bounding box
prediction of the object. On the other hand, it is also easier to converge than
the bounding box regression of the anchor-free approach.
Direct Location Prediction. Faster R-CNN uses the anchor center as the
basis to predict the offset between the object center and the anchor center.
The above approach is very unstable during early training. YOLOv2 follows
the object center regression method of YOLOv1 and directly predicts the
true position of the object center based on the upper left corner of the grid
responsible for predicting the object.
Fine-grained Feature. Passthrough layer predicts by reorganizing high-
resolution features into lossless spatial-to-depth and combining them with
low-resolution features. This can enhance fine-grained small object detection
capabilities through high-resolution features while taking into account speed
simultaneously.
Resolution Calibration. Since the backbone CNN often uses lower-resolution
images for image classification pre-training than those used for object detection
YOLOv1 to YOLOv10 7

training, the pre-trained model has never seen the state of larger objects.
YOLOv2 uses the image classification pre-train of the same training size
image, so that the object detection training process does not require additional
learning of new size object information.
Joint Training with WordTree. YOLOv2 designed the training of group
softmax using ImageNet with a similar hierarchy as WordTree, and then
integrated the categories of COCO and ImageNet using Word- Tree. In the
end, this technology requires joint training of ImageNet’s image classification
and COCO’s object detection tasks. Because of the above design, YOLOv2
has the ability to detect 9000 categories of objects.

2.3 YOLOv3

YOLOv3 [84] was proposed by Redmon and Farhadi in 2018. They integrated
the advanced technology of existing object detection and made corresponding
optimizations to one-stage object detectors. As shown in Figure 3, in terms
of architecture, YOLOv3 mainly combines FPN [61] to enable prediction of
multiple scales at the same time. It also introduces the residual network archi-
tecture and designs DarkNet53. In addition, YOLOv3 also made significant
changes to the label assignment task. The first change is that a ground truth
will only be assigned to one anchor, while the second change is to change from
soft label to hard label for IoU aware objectness. To this day YOLOv3 is still
the most popular version of YOLO series. In what follows, let us detail the
special designs of YOLOv3, namely prediction across scales, high GPU utility,
and SPP.

Figure 3: Architecture of YOLOv3, YOLOv5, and PP-YOLO. The design of YOLOv3 mainly
changes the feature extraction to use SPP (optional) and FPN to extract multi-resolution
features. The initial versions of YOLOv5 and PP-YOLO also follow this architecture.

Predictions Across Scales. YOLOv3 combines FPN to achieve prediction

across scales, which can greatly improve the detection ability of small objects.
High GPU Utility. In the time of 2018, mainstream network architecture
design focuses on reducing the amount of calculations and parameters. The
design of Draknet53 has higher GPU hardware utilization than other architec-
tures, so it has faster inference speed under the same amount of calculations.
8 Wang and Liao

Such a design has also led to subsequent architectural research focusing on

actual hardware inference speed.
SPP. YOLOv1 uses fully connected layer to obtain global context features,
while YOLOv2 uses passthrough layer to combine multiple resolution features.
YOLOv3 designed multiple maximum pooling layers with a stride of 1 for
kernel size from local to global. This design allows each grid to obtain multiple
resolution features from local to global. SPP has been proven to be a simple,
efficient method that can greatly improve accuracy.

2.4 Gaussian YOLOv3

Gaussian YOLOv3 [14] proposed a great way to significantly reduce the false
positive of an object detection process. Gaussian YOLOv3 mainly changes the
decoding method of the prediction head, and the method used is to convert
the bounding box numerical regression problem into predicting its distribution.
Its architecture is shown in Figure 4.

Figure 4: Architecture of Gaussian YOLOv3. Gaussian YOLOv3 changes regression head of

bounding box to Gaussian distribution-based regressor.

Distribution-Based Bounding Box Regression. The distribution-based

bounding box regression module included in the figure is the uncertainty of
predicting bounding box (x, y, w, h) of a Gaussian distribution. This prediction
method can significantly reduce the false positive of object detection.

2.5 YOLOv4

Since Joseph Redmon withdrew from computer vision research for some reason,
subsequent versions of YOLO were mainly released on the open source platform
GitHub. As for the publication time of the paper, it is later than the open
source time. YOLOv4 [2] was submitted to Joseph Redmon as a draft by
Alexey Bochkovskiy, in early April 2020, and was officially released on April
23 2020. YOLOv4 mainly integrates various technologies in different fields
of computer vision in recent years to improve the learning effect of real-time
object detectors. The architectural change of YOLOv4 is to replace FPN with
PAN [66] and introduce CSPNet [106] as backbone, as shown in Figure 5.
Most of the subsequent similar YOLO architectures followed this architecture.
YOLOv1 to YOLOv10 9

Figure 5: Architecture of YOLOv4 [2], Scaled-YOLOv4 [104], YOLOv5 r1–r7 [27–36], and
PP-YOLOv2 [51]. YOLOv4 change the feature integration architecture to Path Aggregation
Network (PAN) [66], and almost all subsequent YOLO versions adopted this design.

In view of the rapid innovation of deep learning technology YOLOv4 was

not only developed based on DarkNet [1, 79], but also implemented on the
most successful PyTorch YOLOv3 [25] at the time. YOLOv4 successfully
demonstrated how to use one GPU to train an object detector that is as
accurate as those trained with more than 128 GPUs, and at the same time
has more than three times the inference speed. The excellent performance
of YOLOv4 has also led to many subsequent object detection research. The
following is a list of new features developed in YOLOv4.
Bag of Freebies. YOLOv4 introduces training techniques that only increase
training time but do not affect inference time mainly including loss function,
regularization methods, data augmentation methods, and label assignment
methods.
Bag of Specials. YOLOv4 also introduces methods that only slightly affect
the inference time but can greatly improve the accuracy, mainly including recep-
tive field modules, attention mechanism, activation function, and normalization
layers, and select useful combinations to add to the system.
Grid Sensitive Decoder. Users on the open source platform found that it
was difficult to predict accurately when the object center was near the grid
boundary. YOLOv4 analyzed the reason and found that the gradient from
Sigmoid function would approach zero at extreme values. The developers of
YOLOv4 then designed a decoding method as shown in Equation 2 to make
the predicted target values fall within the effective gradient range.

bx = (1 + sx )σ(tx ) − 0.5sx + cx ,
by = (1 + sy )σ(ty ) − 0.5sy + cy ,
(2)
bw = pw etw ,
bh = ph eth
Self-Adversarial Training. YOLOv4 also introduces self-adversarial sample
generation training to enhance the robustness of the object detection system.
Training with Memory Sharing. YOLOv4 is also designed to allow GPU
10 Wang and Liao

and CPU to share memory for storing the information required for gradient
updates. This design allows the trained batch size no longer be limited by
GPU memory.

2.6 Scaled-YOLOv4

In 2020, Wang et al. [104] continued the success achieved with YOLOv4
and continued to develop scaled-YOLOv4 that can be used on both edge
and cloud. Thanks to the activity of the DarkNet and PyTorch YOLOv3
communities, scaled-YOLOv4 can abandon the pre-train steps required by
ImageNet and directly use the train-from-scratch method to obtain high-
quality object detection results. In terms of architecture, scaled-YOLOv4
has also introduced CSPNet into PAN, which can comprehensively improve
the performance of speed, accuracy, number of parameters, and number of
calculations. Scaled-YOLOv4 also designs model scaling methods for various
edge devices and provides three types of models: P5, P6, and P7. In the training
part, scaled-YOLOv4 uses the decoder and label assignment strategy proposed
by the initial version of YOLOv5. Because of the various improvements
mentioned above, scaled-YOLOv4 has achieved the highest accuracy and
fastest inference speed of all object detectors. Below we list several unique
designs of scaled-YOLOv4:
Compound Model Scaling. Previous model scaling methods only considered
the integer hyperparameters of a given architecture. Scaled-YOLOv4 proposed
a model scaling that simultaneously considers the input image resolution and
receptive field matching, and uses the number of scaling model stages to design
a more efficient architecture that can be applied to high-resolution images.
Hardware Friendly Architecture. Taking into account ShuffleNet- v2
[70] and HardNet’s [8] analysis of hardware performance, the highly efficient
CSPDark module and CSPOSA module were designed.
Naïve Once For All Model. Since scaled-YOLOv4 is trained in the mode
of train-from-scratch, the problem of inconsistent resolution between the pre-
trained models and the detection model no longer exists. However, the problem
of inconsistency between user input images and training data still exists. The
model scaling method proposed in scaled-YOLOv4 allows users to obtain the
best accuracy without re-training during the inference stage, and only needs
to remove the output of the corresponding stage.

2.7 YOLOv5

YOLOv5 [26] continues the design concept of PyTorch YOLO- v3 and has
simplified and revised the overall architecture definition method. So far,
there are about 10 different versions. The initial version is designed with
an architecture similar to YOLOv3, while following EfficientDet’s [95] model
YOLOv1 to YOLOv10 11

scaling pattern to provide models with different specifications. PyTorch

YOLOv3 was developed from Erik’s open source codes [63], so it uses GPL3
license, and subsequent versions are adjusted to the more strict AGPL3
license. YOLOv5 inherits many functions of PyTorch YOLOv3, such as using
evolutionary algorithms for auto anchor and hyper-parameter search. When
YOLOv5 [26] was open sourced, its performance was slightly worse than
YOLOv3-SPP. After successively combining the CSPNet used by YOLOv4 and
the CSPPAN used by scaled-YOLOv4, the first version of YOLOv5 r1.0 [27]
was officially released in June 2020. Then the developers of YOLOv5 optimized
both the speed-accuracy trade-off of the CSP fusion layer and the activation
function, quoted YOLOR-based training hyper-parameters, used YOLOv5
r6.0 [33] in October 2021. The latest version of YOLOv5 is YOLOv5 r7.0 [36]
Glenn released in November 2022. Due to the continuous maintenance and
version updates by companies, YOLOv5 is currently the most popular YOLO
development platform. Let us point out some distinctive features of YOLOv5
as follows:
Power-based Decoder. The width-height regression system of YOLOv3 uses
exponential function to estimate offset, and this approach causes instability
during training in some datasets. YOLOv5 proposed power-based decoder
in Equation 3 to increase training stability. Since the output value range of
power-based decoder is limited to a certain scaling range of anchor, there will
be a theoretically bounded recall.

bx = 2σ(tx ) − 0.5 + cx ,
by = 2σ(ty ) − 0.5 + cy ,
(3)
bw = pw (2σ(tw ))2 ,
bh = ph (2σ(th ))2
Neighborhood Positive Samples. In order to make up for the deficiency
caused by recall, YOLOv5 proposed to add more neighbor grids as positive
samples. At the same time, in order to allow these neighbor grids to correctly
predict the center point, they also enlarged the sigmoid scaling coefficient of
the YOLOv4 center point decoder.

2.8 PP-YOLO

There are four versions of the PP-YOLO series, namely PP-YOLO [68], PP-
YOLOv2 [51], PP-PicoDet [119], and PP-YOLOE [116]. PP-YOLO is improved
based on YOLOv3. In addition to using a variety of YOLOv4 training tech-
niques, it also adds CoordConv [65], Matrix NMS [111], and better ImageNet
pre-trained model and other methods for improvement, while PP-YOLOv2
further introduces scaled-YOLOv4’s CSPPAN and other mechanisms. PP-
PicoDet uses neural architecture search as the basis to design the backbone,
12 Wang and Liao

and introduces YOLOX’s anchor-free decoder [22]. As for PP-YOLOE, it

has made major changes, as shown in Figure 6. It modified RepVGG and
designed CSPRepResStage and then used bounding box regression in TOOD’s
distribution-based regression process [20]. The YOLO series after YOLOv6
almost all follow the above format. Listed below are some design features of
PP-YOLO series.
Neural Architecture Search. PP-PicoDet is an architecture designed for
mobile devices. It combines ShuffleNetv2 [70] and GhostNet [46] for one shot
neural architecture search.
Reparameterization Module. PP-YOLOE applies RepVGG [17] to CSP-
Net, but removes the identity connection in the training phase.
Distribution-based Regression Raised. PP-YOLOE follows
TOOD to use DFL [59] for bounding box regression. DFL is different from
Gaussian YOLOv3 in that it does not need to limit the data to be Gaussian
distribution, and can also directly learn the distribution of real data.

Figure 6: Architecture of PP-YOLOE [116], YOLOv6 2.0 [56], YOLOv7 AF [105], YOLOv8
[37], and YOLO-NAS [93]. PP-YOLOE changed bounding box regression head to TOOD’s
[20] anchor-free distribution-based regressor, and subsequent YOLO versions adopted this
design.

2.9 YOLOR

YOLOR [110] is not an official version of the YOLO series, but its use of
Latent Variable Model (LVM) as implicit knowledge encoder can significantly
improve the detection effects of all YOLO series models, as shown in Figure 7.
YOLOR’s multi-task model has also been widely used in subsequent YOLO
versions, and the advanced training technology it proposed has been continued
and promoted in all subsequent versions. Below are some specially designed
features of YOLOR.
Implicit Knowledge Modeling. YOLOR proposed three LVMs to encode
implicit knowledge, including vector-based, neural network-based, and ma-
trix factorization-based. The above three encoding methods can effectively
enhance the feature alignment, prediction refinement, and multi-task learning
capabilities of deep neural networks.
YOLOv1 to YOLOv10 13

Figure 7: Architecture of YOLOR. YOLOR proposed refining bounding box regression head
to Latent Variable Models (LVM).

Multi-task Model. YOLOR provides models that can perform object

detection, image classification, and multiple object tracking at the same time,
and it also provides pose estimation models that based on YOLO-Pose [71].
Advanced Training Technique. YOLOR developed advanced autoML
technology [108], and its techniques for training hyperparameters are continued
to be used in the latest version of YOLO series. YOLOR also uses large dataset
pre-train, knowledge distillation, self-supervised learning, and self distillation
technologies in its model. Until now, YOLOR trained using the above method
is still the most accurate model of all YOLO series.

2.10 YOLOX

YOLOX [22], as shown in Figure 8, combined the most practical technologies

at the time, mainly based on the CSPNet architecture [106] and FCOS’ anchor-
free head [98], improved OTA [21] and proposed the SimOTA dynamic label
assignment method to replace the manual label assignment method that was
easily confusing. Subsequent versions of YOLO also began to use or design
different dynamic label assignment methods. The following describes the two
most important features of YOLOX.
Decoupled Head. YOLOX uses decoupled head of FCOS, and this design
makes classification and bounding box regression easier to learn.
Anchor-free Strikes Back. Due to the development of IoU-based loss,
anchor-free head is no longer affected by the loss imbalance caused by the
length and width of the object. With modern technology, anchor-free head
can also be well trained. As for YOLOX, it takes FCOS’ anchor-free head to
achieve the planned objective.
14 Wang and Liao

Figure 8: Architecture of YOLOX [22]. YOLOX proposed changing bounding box regression
head to anchor-free regressor of FCOS [98], which also led to the development of the YOLO
series towards anchor-free.

2.11 YOLOv6

The initial version of YOLOv6 [73] uses RepVGG [17] as the main architecture.
In versions after version 2.0, such as Li et al. [56, 57], CSPNet [106] was
introduced. YOLOv6 is a system specially designed for industry, so it has
put a lot of effort into quantization issues. The contributions of YOLOv6
include using RepOPT [16] to make the quantized model more stable, and
using quantization aware training (QAT) and knowledge distillation to enhance
the accuracy of the quantized model. YOLOv6 version 3.0 [56] proposed a
concept of anchor-aid training, as shown in Figure 9, to improve the accuracy
of the system. Later in YOLOv6 version 4.0 [58], a lightweight architecture
YOLOv6-lite based on depth-wise convolution was proposed to face lower-end
computing devices. The following lists some of the unique features proposed
by YOLOv6.
Reparameterizing Optimizer. YOLOv6 version 2.0 uses RepOPT to slow
down the accuracy lost after model quantization.
Quantization Aware Training. In YOLOv6 version 2.0, QAT is used to
improve the accuracy of the quantization model.
Knowledge Distillation. YOLOv6 version 2.0 uses self-distillation and
channel-wise distillation respectively to improve model accuracy, and it also
uses QAT to reduce the accuracy loss after model quantization.
Anchor-Aided Training. YOLOv6 version 3.0 proposed using anchor-based
head to assist anchor-free head learning, as shown in Figure 9, to improve
accuracy.

2.12 YOLOv7
YOLOv7 [105] introduces trainable auxiliary architectures that can be removed
or integrated during the inference stage, including YOLOR [110], the recently
popular RepVGG [17], and additional auxiliary losses. Architecturally, as
shown in Figure 10, YOLOv7 uses ELAN [107] to replace the CSPNet used by
YOLOv4, and proposes E-ELAN to design large models. YOLOv7 also provides
YOLOv1 to YOLOv10 15

Figure 9: Architecture of YOLOv6 3.0 and YOLOv6 4.0. YOLOv6 3.0 proposed to guide
the learning of anchor-free bounding box regression head with anchor-based bounding box
regression head.

Figure 10: Architecture of YOLOv7 [105]. YOLOv7 guides learning with coarse-to-fine
consistent auxliary head and inspires many consistent multiple prediction training research.

a variety of computer vision task-related models and supports anchor-based

and anchor-free architectures. The features of YOLOv7 are listed below.
Make RepVGG Great Again. The reparameterization method proposed by
RepVGG [17] allows simple network architectures to converge when deepening,
but it cannot be effectively applied to modern popular deep network archi-
tectures. The planned RepConv technology proposed by YOLOv7 allows the
reparameterization method to effectively bring gains to various residual-based
and concatenation-based architectures.
Consistent Label Assignment. The auxiliary loss method used in the
past will make the output targets of different branches inconsistent, which
will lead to confusion and instability when performing training. In response
to the maturity and popularity of the dynamic label assignment method,
YOLOv7 proposed the consistent label assignment mechanism to maintain the
consistency of the goals and feature learning directions of different branches.
Coarse to Fine Label Assignment. In the past, multi-stage refinement
architectures, such as Cascade R-CNN [6] and HTC [9], required additional
theoretical architectures to refine predictions step by step. YOLOv7 proposed
the coarse-to-fine label assignment mechanism, which can directly use auxiliary
16 Wang and Liao

loss to guide the coarse-to-fine characteristics in the feature space, providing

prediction refinement effects without changing the architecture.
Partial Auxiliary Loss. YOLOv7 allows some features to receive auxiliary
information updates, and the remaining parts are still focused on target task
learning. The developers of YOLOv7 found that this design has a good
improvement effect on the main tasks.
Various Vision Tasks. YOLOv7 provides models including object detection,
instance segmentation, and pose estimation, and has achieved real-time state-
of-the-art in these tasks.

2.13 DAMO-YOLO

DAMO-YOLO [117] have proposed improved methods in terms of backbone

architecture, feature integration, prediction head, and label assignment. The
block diagram of DAMO-YOLO is shown in Figure 11. Its features are
summarized as follows.
MAE-NAS. DAMO-YOLO uses MAE-NAS [92] to search CSPNet and ELAN
to achieve a more efficient architecture.
Efficient GFPN. DAMO-YOLO disassembled the queen-fusion of GFPN [53]
and retained the fusion layers of the trade-off that is designed for achieving the
best speed and accuracy, so as to combine it with ELAN and design RepGFPN.
ZeroHead. DAMO-YOLO simplifies the complex decoupled head into a
feature projection layer.
AlignedOTA. DAMO-YOLO proposed aligned OTA to solve the misalignment
problem of classification prediction, regression prediction, and dynamic label
assignment.

Figure 11: Architecture of DAMO-YOLO. DAMO-YOLO proposed changing the feature

integration architecture to GFPN.

2.14 YOLOv8

YOLOv8 [37] is a refactored version of YOLOv5 [26], which updates the way
the overall API is used and makes a lot of underlying code optimizations. It
architecturally changes YOLOv7’s ELAN, plus additional residual connection,
while its decoder is the same as YOLOv6 2.0. It is not so much a new
YOLOv1 to YOLOv10 17

YOLO version as it is a technology integration platform, and it basically

integrates the APIs of multiple downstream tasks and connects them in series.
YOLOv8 r2.0 [39] integrates the latest technologies such as YOLOv9 and
YOLO World [13]. Its most recent version is YOLOv8 r3.0 [41], which is also
known as YOLO11. It architecturally updates to combination of YOLOv9’s
Dark-ELAN and CSP-ELAN, and CSPNet and YOLOv10’s PSA to achieve
better speed-accuracy trade-off. Because program modification and API
usage are not very intuitive, many developers have not yet switched to this
platform. But for professional users, the performance improvements brought
by optimization of the underlying program code have also attracted many
R&D teams to use it. In what follows are two special features of YOLOv8:
Code Optimization. The optimization of the underlying program code
released by YOLOv8 has brought about 30% improvement in training perfor-
mance.
API for Down-stream Applications. YOLOv8 also provides a simple
API to connect the detection model with various downstream tasks, such as
segment anything, instance segmentation, pose estimation, multiple object
tracking, etc.

2.15 YOLO-NAS
YOLO-NAS [93] did not reveal too many technical details. It mainly uses its
own AutoNAC NAS to design the quantization friendly architecture and uses
a multi-stage training process, including pre-training on Object365, COCO
Pseudo-Labeled data, Knowledge Distillation (KD), and Distribution Focal
Loss (DFL).

2.16 Gold-YOLO
Gold-YOLO [103] The overall architecture of Gold-YOLO is similar to that of
YOLOv6 3.0. The main design is that the Gather-and-Distribute mechanism
replaces PAN in the architecture, and masked image modeling is pre-trained
during the training process.
Gather-and-Distribute Mechanism. The main architecture of
Gather-and-Distribute is shown in Figure 12. It mainly collects features from
each layer through two gather-and-distribute modules and integrates them
into global features using transformers. The integrated global features will
be distributed to the low-level and high-level layers respectively, and the
distribution method uses the information injection module to integrate the
global features with the features that have been distributed to layers.
18 Wang and Liao

Figure 12: Architecture of Gold-YOLO. Gold-YOLO proposed adding global features to the
feature integration architecture.

2.17 YOLOv9
YOLOv9 [109] proposed an important trustworthy technology – Programmable
Gradient Information (PGI), whose architecture is shown in Figure 13. The
design architecture in the figure can enhance the interpretability, robustness,
and versatility of the model. The design of PGI is to use the concepts of
reversible architecture and multi-level information to maximize the original
data that the model can retain and the information needed to complete the
target tasks. YOLOv9 extended ELAN to G-ELAN and used it to show how
PGI can achieve excellent accuracy, stability and inference speed on models
with low number of parameters. Several outstanding features of YOLOv9 are
described below.
Auxiliary Reversible Branch. PGI exploits the properties of reversible ar-
chitecture to solve the information bottleneck problem in deep neural networks.
This is completely different from the general-purpose reversible architecture
which simply maximizes the information to be retained. What PGI uses is
to share the information retained by reversible architecture with the main
branch in the form of auxiliary information. On the premise of retaining the
information required for the target task, retain as much information as possible
from the original data.
Multi-level Auxiliary Information. PGI proposed the concept of multi-
level auxiliary information so that each layer of the main branch features retains
the information required for all task objectives as much as possible. This
can avoid the problem that past methods tend to lose important information
at the shallow level, which in turn leads to the inability to obtain sufficient
information at the deep level.
Generalize to Down-stream Tasks. Because PGI can maximize the
retention of original data information, models trained by PGI achieve more
robust performance in small datasets, transfer learning, multi-task learning,
and adaptation to new downstream tasks.
YOLOv1 to YOLOv10 19

Figure 13: Architecture of YOLOv9. YOLOv9 proposed using auxiliary branches to help
learning.

Generalize to Various Architectures. PGI can also be applied to other

architectures, such as conventional CNN, depth-wise convolutional CNN,
transformer, and different types of computer vision methods, such as anchor-
based, anchor-free, post-processing free, etc. Therefore, PGI has absolutely
superior versatility.

2.18 YOLOv10

The overall architecture of YOLOv10 [102] is shown in Figure 14, which is

similar to YOLOv6 3.0, but the transformer-based module is added to enhance
the extraction of global features. They changed the dual head to one-to-
many and one-to-one matching, respectively. This change allows YOLO to
do without post-processing likes the DETR-based method, and can directly
obtain end-to-end object detection results. Next, we introduce some of the
distinct features of YOLOv10.

Figure 14: Architecture of YOLOv10 [102]. YOLOv10 proposed to use one-to-many

prediction head to guide the learning of one-to-one prediction head.

Dual Label Assignment. Use the label assignment method similar to

DATE [12], and add stop gradient operation to the one-to-one branch.
NMS-free Object Detection. The design of one-to-one matching mechanism
enables the prediction process without relying on NMS for post-processing.
20 Wang and Liao

Rank-guided Block Design. Proposed to use rank to determine which stages

use conventional convolution and which stages use depth-wise convolution.
Partial Self-attention. YOLOv10 combined CSPNet and Transformer and
proposed the self-attention module.

2.19 Comparison of different YOLOs

The architecture of the entire YOLO series is summarized in Table 2. YOLOv1

innovatively proposed the first one-stage object detection architecture and
was also the pioneer of anchor-free object detection. YOLOv3 and YOLOv4
respectively integrated the most advanced technologies at he time, allowing
subsequent YOLO series versions to follow their main architecture. YOLOv3
designed ResNet-based backbone and FPN-based neck, while YOLOv4 designed
CSPNet-based backbone and PAN-based neck. Gaussian YOLOv3 proposed
the earliest and very successful distribution-based prediction head research,
while YOLOX introduced FCOS’s anchor-free head into the YOLO mechanism,
and thus promoted the research of the YOLO series in the anchor-free direc-
tion. PP-YOLOE further used TOOD’s distribution-based prediction head
to enhance the design of anchor-free head, and thus became the subsequent
mainstream prediction head, while YOLOv7 designed ELAN-based backbone
and thus became today’s mainstream architecture. In addition, YOLOv7
also designed the multiple consistent head training method to drive many
related research. As for YOLOv10, it proposed to add the one-to-one matching
mechanism to label assignment, and then designed end-to-end one-stage object
detector.
Although each generation of YOLO versions continues to advance the
accuracy-efficiency trad-off, it is worth noting that in addition to improvements
in architecture, different versions of YOLO also seek improvements in training
methods. For this reason, users need to have a deeper understanding of the
nature of every method in order to choose the method that best suits them.
In addition, YOLOv7 trained several different versions of YOLO systems
using YOLOR’s training strategy and compared them. The purpose of doing
this is to allow users to more objectively observe the performance of different
versions under the same conditions. Since the basic training strategies of
different subsequent YOLO versions have almost not changed since YOLOR,
we present the data of training different YOLOs under the same conditions
in Table 3. From Table 3, we can clear see the differences in data between
different versions of YOLO. However, since the inference time of TensorRT
(TRT) is very sensitive to the layer number of the network, users must still
choose the architecture that is most friendly to them in actual use.
In addition to the architecture, the prediction head and label assignment
algorithms are also important factors that effect the object detection results.
For example, when objects do not obscure each other, it is a better solution to
YOLOv1 to YOLOv10 21

Table 2: Architecture of YOLO series. The bold font indicates changes that significantly
affected subsequent versions.

Model Backbone Neck Head

YOLO (YOLO1) PlainNet-based Fully-Connected Anchor-free
YOLO9000 (YOLOv2) PlainNet-based Passthrough Anchor-based
YOLOv3 ResNet-based FPN-based Anchor-based
Gaussian YOLOv3 ResNet-based FPN-based Anchor-based
YOLOv4 CSPNet-based PAN-based Anchor-based
YOLOv5 ResNet-based FPN-based Anchor-based
Scaled-YOLOv4 CSPNet-based PAN-based Anchor-based
YOLOv5 r1.0 CSPNet-based PAN-based Anchor-based
PP-YOLO ResNet-based FPN-based Anchor-based
YOLOv5 r2.0 CSPNet-based PAN-based Anchor-based
YOLOv5 r3.0 CSPNet-based PAN-based Anchor-based
Scaled-YOLOv4 P6 CSPNet-based PAN-based Anchor-based
YOLOv5 r3.1 CSPNet-based PAN-based Anchor-based
YOLOv5 r4.0 CSPNet-based PAN-based Anchor-based
PP-YOLOv2 ResNet-based PAN-based Anchor-based
YOLOv5 r5.0 CSPNet-based PAN-based Anchor-based
YOLOR CSPNet-based PAN-based Anchor-based
YOLOX CSPNet-based PAN-based Anchor-free
YOLOv5 r6.0 CSPNet-based PAN-based Anchor-based
PP-PicoDet CSPNet-based PAN-based Anchor-free
YOLOv5 r6.1 CSPNet-based PAN-based Anchor-based
PP-YOLOE CSPNet-based PAN-based Anchor-free
YOLOv6 PlainNet-based PAN-based Anchor-free
YOLOv7 ELAN-based PAN-based Anchor-based
YOLOv5 r6.2 CSPNet-based PAN-based Anchor-based
YOLOv6 2.0 CSPNet-based PAN-based Anchor-free
YOLOv7 AF ELAN-based PAN-based Anchor-free
DAMO-YOLO ELAN-based PAN-based Anchor-based
YOLOv5 r7.0 CSPNet-based PAN-based Anchor-based
YOLOv8 ELAN-based PAN-based Anchor-free
YOLOv6 3.0 CSPNet-based PAN-based Anchor-free
YOLOv6 4.0 CSPNet-based PAN-based Anchor-free
YOLO-NAS ELAN-based PAN-based Anchor-free
Gold-YOLO CSPNet-based PAN-based Anchor-free
YOLOv8 r1.0 ELAN-based PAN-based Anchor-free
YOLOv9 ELAN-based PAN-based Anchor-free
YOLOv8 r2.0’ ELAN-based PAN-based Anchor-free
YOLOv10 ELAN-based PAN-based Anchor-free
YOLOv8 r2.0 ELAN-based PAN-based Anchor-free
YOLOv8 r3.0 ELAN-based PAN-based Anchor-free

use label assignment methods such as YOLOv4 that can effectively improve
recall. On the contrary, it is recommended to use the dynamic label assignment
method developed after YOLOX. When the aspect ratio of an object is
relatively fixed, it is more effective to use anchor-based prediction head. When
the aspect ratio of an object is extreme, the anchor-free method is suitable.
22 Wang and Liao

Table 3: The performance of the main architecture of the YOLO series of papers on the
COCO dataset. The bold font indicates Pareto optimal.

Model #Param. (M) FLOPs (G) mAP (%) T4, TRT (ms)
YOLOv4 64.4 142.8 49.7 –
Scaled-YOLOv4 52.9 120.4 50.3 –
YOLOR-CSP 52.9 120.4 50.8 –
YOLOv7 36.9 104.7 51.2 –
YOLOv5-L r6.2 46.5 109.1 49.0 –
YOLOv6-L 2.0 58.5 144.0 51.0 –
YOLOv7-L AF 43.6 130.5 53.0 6.7
YOLOv8-L 43.7 165.2 52.9 8.1
YOLOv6-L 3.0 59.6 150.7 51.8 7.9
YOLOv9-C 25.3 102.1 53.0 6.1
YOLOv9-TR 14.1 67.5 53.1 5.9
YOLOv10-B 19.1 92.0 52.5 5.7
YOLOv10-L 24.4 120.3 53.2 7.2
YOLOv8-L r3.0 25.3 86.9 53.4 6.2

As for YOLOv10 which proposed using one-to-one prediction head, it is easy

to generate duplicated prediction boxes for the same object, so appropriate
post-processing still needed. In Figure 15, we show prediction results and
comparisons for the latest three YOLO versions.

Figure 15: Prediction of latest versions of YOLO.

YOLOv1 to YOLOv10 23

3 Impact of YOLO series

The YOLO series of algorithms have the characteristics of (1) relatively simple
frame and (2) relatively easy deployment. In what follows we will describe
these characteristics in detail.

3.1 Simpler

Simpler Frame. Based on the most forward-looking research on DeepMulti-

Box [19] and OverFeat [88], YOLO proposed a new way of one-stage object
detection, and this new approach influenced many subsequent computer vision
research. Before YOLO series was proposed, the tasks that originally required
deep learning to perform dense prediction mainly included pixel-level tasks
such as semantic segmentation and optical flow estimation. As for object
detection, pose estimation and other instance-level tasks, most of them are
split into multiple sub-tasks and predicted in the cascade way. After YOLO
was proposed, many algorithms that originally needed to use multi-stage and
bottom-up methods were suddenly converted to end-to-end, top-down, and
one-stage methods. Examples of this sort include pose estimation and facial
landmark detection that directly predict the bounding box and the relative
positions of anchor points in the bounding box, as well as multi-object tracking
that simultaneously detects and extracts re-identification features.
Simpler Deployment. YOLO does not use specially structured modules,
so it is very easy to be deployed on a variety of computing devices. However,
there are many special designs such as receptive field module and attention
mechanism etc., which are very helpful in improving the accuracy of object
detection, but it is not easy to design them into universal and simple modules.
YOLO converts these special modules into modules with simple structures
through clever design. For example, YOLOv3 proposed to use max-pooling
with multiple resolutions to sweep the feature map with a stride of 1 to
improve the SPP layer that is originally limited to a fixed input size, and
the ASPP layer that requires dilated convolution. The above approach can
greatly enhance the model’s multi-resolution and global perception capabilities.
As for YOLOv4, it proposed to use a convolution layer to replace the small
network containing various pooling layers and fully connected layers in the
attention module. The entire model of YOLOv4-tiny, YOLOv6 and YOLOv7
was even improved to the point where it only needs to be composed using
1×1 convolution, 3×3 convolution, and max-pooling. In addition, Darknet
developed in C language by Joseph Redmon allows the training and inference
process of YOLO without relying on additional software packages. The above-
mentioned friendly deployment to existing equipment makes the YOLO series
widely used in various practical systems.
24 Wang and Liao

3.2 Better

In addition to being lightweight and easy-to-use as described in the previous

section, the YOLO series models also have some more advanced functions,
such as better training techniques, better model scalability, and better model
generalizability. Below we describe these functions in detail.
Better Training Technique. The training technology proposed by YOLO
series models is not only more advanced but also complementary to the most
advanced training technology currently available. Many studies in the past
have mostly verified the proposed method on a foundation method, such as
ResNet [47] and ViT [18] for image recognition, or faster R-CNN [85] and
DETR [7] for object detection. However, most studies ignore whether the
proposed methods are compatible with the current state-of-the-art methods
and complement each other to promote the overall progress of the field. Since
YOLOv2, the YOLO series has considered compatibility with the most ad-
vanced technologies when designing, and at the same time proposed new
methods that can complement these technologies. In YOLOv3, YOLOv4, and
PP-YOLO series of models, developers also try to analyze technologies that
cannot be compatible with each other. This kind of attitude has an important
guiding role for subsequent developers.
Better Model Scalability. The YOLO series models do not require special
settings when performing model scaling. Scaled-YOLOv4 and YOLOv7 pro-
posed some guidelines for model scaling, while YOLOv5 follows EfficientDet’s
model scaling method. These scaling methods are directly integrated into the
framework, and this kind of design allows users to obtain stable and great
performance no matter how they adjust the hyperparameters of model scaling.
Better Model Generalizability. The methods proposed by the YOLO series
can be applied to many fields. For example, the concept of using prediction and
ground truth to calculate metric proposed by YOLOv1 has been widely used
in various soft label generation methods, while the method of using K-means
as the initial anchor proposed by YOLOv2 has been extended to the pose
estimation field. The WorldTree group softmax method proposed by YOLO
is transformed into dealing with the imbalanced data distribution problem
of long-tailed learning. SimOTA proposed by YOLOX is used as the basic
method for various dynamic label assignment, and the hybrid label assignment
method proposed by YOLOv7 is also widely used. The methods proposed by
YOLO series are also applicable to a variety of architectures, for example, the
CSPNet used by YOLOv4 not only shows excellent results on CNN, but has
also been proven to work well with architectures such as Transformer [100],
Graph Neural Networks [87], Spiking Neural Networks [96], and MAMBA [43].
The subsequent ELAN used by YOLOv7 has also been rapidly applied in
various computer vision fields.
YOLOv1 to YOLOv10 25

3.3 Faster

Faster Architecture. Another feature of the YOLO series is its very fast
inference speed, mainly because its architecture is designed for the actual
inference speed of the hardware. The designers of YOLOv3 found that even a
simple 1×1 convolution and 3×3 convolution combined architecture, although
it has a lower computational load, does not necessarily mean that it has an
advantage in inference speed. Therefore, they designed DarkNet for real-time
object detection. As for the designers of Scaled-YOLOv4, they referred to
research including ShuffleNetv2 [70] and HarDNet [8], and further analyzed
the criteria that need to be considered for high inference speed architecture for
different levels of devices from edge to cloud. To achieve the same purpose, the
developers of scaled-YOLOv4 designed Fully CSPOSANet and CSPDarkNet.
As for the developers of YOLOv6, they used the efficient RepVGG as the
backbone, while the designers of DAMO-YOLO used NAS technology to
directly search for efficient architectures in CSPNet and ELAN.

3.4 Stronger
Stronger Adaptability. The YOLO series has gained great progress and
response in the open source community. The training method integrated by
Darknet and PyTorch YOLOv3 allows YOLO series to train object detectors
without relying on ImageNet’s pre-trained model. Due to the above reasons,
the YOLO series can be easily applied to data in different domains without
relying on a large number of training models corresponding to the domain.
The above advantages enable the YOLO series to be widely used in various
application domains. In addition, the YOLO series can also be easily applied to
different datasets. For example, PyTorch YOLOv3 proposes to use evolutionary
algorithms to automatically search for hyperparameters, which can be applied
to different datasets. In addition, the improved anchor-free-based YOLO from
YOLOX to PP-YOLOE allows YOLO to rely on fewer hyperparameters during
training and can be used more widely in various application domains.
Stronger Capability. The YOLO series has excellent performance in a variety
of computer vision tasks. For example, after being widely used in the field of
real-time object detection, many other computer vision models based on YOLO
have been developed, including YOLACT [3] instance segmentation model,
JDE [113] multiple object tracking, and so on. Taking YOLOR as an example,
it began to combine multiple tasks into the same model for prediction. It can
perform image recognition, object detection, and multi-object tracking at the
same time, and significantly improve the effect of multi-task joint learning. On
the same task, YOLOv5 trains image recognition and object detection models
separately. In addition, YOLOv7 also demonstrated outstanding performance
in a variety of computer vision domains. At that time, it became the most
26 Wang and Liao

advanced method for real-time object detection, instance segmentation, and

pose estimation. On the same issue, YOLOv8 additionally integrates tasks
such as rotating object detection and pose estimation. In addition, YOLOv9
further combines YOLOv7 and YOLOR to extend the multi-task model to
the visual-language domain.
Stronger Versatility. Since object detection is a necessary starting step for
many practical applications, and as a top object detection method, YOLO’s
design is very suitable for matching with various downstream task models. In
this regard, the design of PP-YOLO series is particularly outstanding, and
this series can provide an integrated system for dozens of downstream tasks
including face analysis, license plate recognition, multi-object tracking, traffic
statistics, behavior analysis, etc.

4 YOLO for various computer vision tasks

The YOLO series systems have been widely used in many fields. In this section,
we will introduce YOLO’s representative works in other computer vision fields
and explain the new designs either in architecture or methods completed by
these representative works in order to achieve real-time performance.

4.1 Multiple Object Tracking

In the past, deep learning-based multiple object tracking related algorithms,

such as Deep-SORT [114], needed to crop the detected object area from the
original image after detecting objects, and then capture features through
additional networks for tracking. ROLO [44] proposes objects directly detected
by YOLO, and uses LSTM [50] for single object tracking. They proposed to use
multiple LSTM to design MOLO and then perform multiple object tracking.
JDE [113] proposes to output the re-ID features for object tracking while
detecting objects. However, JDE’s multi-scale dense prediction re-ID feature
requires a large amount of calculations. In addition, a set of re-ID features will
match multiple anchors, making it easy to confuse IDs. CSTrack [60] further
combines JDE and FairMOT [123], and after integrating multi-scale features,
only outputs re-ID features at one scale. This can achieve more accurate
multi-object tracking effects.

4.2 Instance Segmentation

In the past, most instance segmentation prediction was performed separately

for each detected object, so more complex segmentation network is required.
YOLACT [3] and YOLACT++ [4] decompose the instance segmentation pro-
cess into two steps, namely prototypes and coefficients, and only need to predict
YOLOv1 to YOLOv10 27

coefficients to use these prototypes to form the output instance segmentation

results. Using the above method can greatly reduce the amount of operations
required when instance segmentation is executed. YOLACTEdge [64] then
pushes instance segmentation further to the video domain. The concept of
using FeatFlowNet greatly reduces the number of features extracted by the
backbone.
Another way to reduce the computation of instance segmentation prediction
is to express binary mask in other ways, such as expressing mask in the form
of polygon or polar coordinates. Although this expression method will cause
some distortion, it can express the mask of the object in very few dimensions.
Insta-YOLO [74] and Poly YOLO [52] are two examples to use the polygon
form to predict the result of instance segmentation.

4.3 Automated Driving

YOLO series is also widely used in visual perception tasks in self-driving
scenarios. YOLOP [115] and YOLOPv2 [45] respectively use CSPNet and
ELAN as the main architecture for object detection, and therefore can be used
for area detection and lane prediction. HybridNet [101], YOLOPv3 [121], and
YOLOPX [122] are also modified by different versions of YOLO and perform
self-driving tasks.

4.4 Human Pose Estimation

Human pose estimation can be viewed as additional spatial attributes for
predicting object detection targets. Since keypoints do not necessarily fall
in grids, additional decoder design is required. KAPAO [72] divides human
pose estimation into human pose object and keypoint object expressions for
prediction and combination. YOLO-Pose [71] directly predicts the regression
value of the key relative to the center of the grid, and then execute human
pose estimation. The above design can achieve pretty good results.

4.5 3D Object Detection

There are also some studies that generalize YOLO series from 2D to 3D. In
addition to ComplexYOLO [91] which combines images and LIDAR as input,
and Expandable YOLO [94] which uses RGB-D images as input, there is also
YOLO 6D [97] and YOLO 3D [86] which simply use images as input.

4.6 Video Perception

YOLO series, which performs extremely well in real-time object detection

in images, will naturally be applied to the video domain. Among them,
28 Wang and Liao

YOLOV [89] and YOLOV++ [90] can be applied to video object detection.
Alternatively, stream YOLO [118] can be used with streaming perception.

4.7 Face Detection

Face detection is one of the most popular subfields among the various possible
application domains of object detection. The face detection models designed
based on YOLO [11, 80, 120] also performs quite well in this field.

4.8 Image Segmentation

Due to the real-time and high-performance characteristics of YOLO, it has
also begun to be combined with many foundation models and applied to new
computer vision tasks. Fast-SAM [125] combines YOLO with SAM [55] and
applies it to the general image segmentation task. The above combination can
greatly improve the inference speed of the task model.

4.9 Open Vocabulary Object Detection

YOLO is also used in conjunction with visual language foundation models.
Examples of this sort include YOLO-world [13] and Open-YOLO 3D [5], which
combine YOLO and CLIP [81] methods and can be used to perform 2D and
3D open vocabulary object detection respectively.

4.10 Combine With Other Architecture

YOLO also demonstrates compatibility with a variety of deep neural network ar-
chitectures. Architectures of this sort include ViT [75–78, 124], MAMBA [112],
SNN [54, 69], GNN [10, 42], and KAN [15]. They can all be effectively combined
with YOLO.

4.11 Summarization

Observing the development of YOLO series and thier applications in different

fields, we summarize the possible future development trends of YOLO. First,
various services integrated based on YOLO object detection. Examples of this
sort are PP-YOLO and YOLOv8. They used YOLO as a real-time detector
for various objects, and then used multi-object tracking methods to conduct
more in-depth value-added analysis. Second, a multi-modal multi-task model
designed with the YOLO architecture as the main body. Since YOLOR started
working on multi-modal and multi-tasking YOLO, YOLOv7 integrates instance
segmentation and pose estimation into the main YOLO version for the first
time. By YOLOv9, multi-task models including visual-language tasks have
YOLOv1 to YOLOv10 29

also been developed more and more complete. We believe that real-time
multi-modal multi-tasking based on YOLO will have further development in
the future. The third is the combination of YOLO with large language models
and foundation models. The combination of YOLO and SAM in FastSAM
and the combination of YOLO and CLIP in YOLO-World are pioneers of this
type of work. Due to the attention brought by these studies, it is conceivable
that combining YOLO with large language models will be a direction worth
exploring in the future.

5 Conclusions

In this article, we introduce the evolution of the YOLO series over the years,
review these technologies from the perspective of modern object detection
technology, and point out the key contributions they made at each stage. We
analyze YOLO’s influence on the field of modern computer vision from aspects
such as ease of use, accuracy improvement, speed improvement, and versatility
in various fields. Finally, we introduce the YOLO-related models in various
fields. The purpose is that through this review article, readers can not only be
inspired by the development of the YOLO series, but also better understand
how to develop various real-time computer vision methods. We also hope
to provide readers an idea of the different tasks YOLO can be used for and
possible future directions.

References

[1] AlexeyAB, “darknet”, 2019, https://github.com/AlexeyAB/darknet.

[2] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “YOLOv4: Optimal
Speed and Accuracy of Object Detection”, arXiv preprint arXiv:2004.
10934, 2020.
[3] D. Bolya, C. Zhou, F. Xiao, and Y. J. Lee, “YOLACT: Real-time
instance segmentation”, in Proceedings of the IEEE International Con-
ference on Computer Vision (ICCV), 2019, 9157–66.
[4] D. Bolya, C. Zhou, F. Xiao, and Y. J. Lee, “YOLACT++ Better Real-
Time Instance Segmentation”, IEEE Transactions on Pattern Analysis
and Machine Intelligence (TPAMI), 2022.
[5] M. E. A. Boudjoghra, A. Dai, J. Lahoud, H. Cholakkal, R. M. Anwer,
S. Khan, and F. S. Khan, “Open-YOLO 3D: Towards Fast and Ac-
curate Open-Vocabulary 3D Instance Segmentation”, arXiv preprint
arXiv:2406.02548, 2024.
30 Wang and Liao

[6] Z. Cai and N. Vasconcelos, “Cascade R-CNN: Delving into high quality
object detection”, in Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), 2018, 6154–62.
[7] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S.
Zagoruyko, “End-to-end object detection with transformers”, in Pro-
ceedings of the European conference on computer vision (ECCV), 2020,
213–29.
[8] P. Chao, C.-Y. Kao, Y.-S. Ruan, C.-H. Huang, and Y.-L. Lin, “HarDNet:
A low memory traffic network”, in Proceedings of the IEEE International
Conference on Computer Vision (ICCV), 2019, 3552–61.
[9] K. Chen, J. Pang, J. Wang, Y. Xiong, X. Li, S. Sun, W. Feng, Z. Liu, J.
Shi, W. Ouyang, et al., “Hybrid task cascade for instance segmentation”,
in Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), 2019, 4974–83.
[10] P. Chen, Y. Wang, and H. Liu, “GCN-YOLO: YOLO Based on Graph
Convolutional Network for SAR Vehicle Target Detection”, IEEE Geo-
science and Remote Sensing Letters, 2024.
[11] W. Chen, H. Huang, S. Peng, C. Zhou, and C. Zhang, “YOLO-Face: a
real-time face detector”, The Visual Computer, 37, 2021, 805–13.
[12] Y. Chen, Q. Chen, Q. Hu, and J. Cheng, “DATE: Dual assignment
for end-to-end fully convolutional object detection”, arXiv preprint
arXiv:2211.13859, 2022.
[13] T. Cheng, L. Song, Y. Ge, W. Liu, X. Wang, and Y. Shan, “YOLO-
World: Real-time open-vocabulary object detection”, in Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern Recogni-
tion (CVPR), 2024, 16901–11.
[14] J. Choi, D. Chun, H. Kim, and H.-J. Lee, “Gaussian YOLOv3: An
accurate and fast object detector using localization uncertainty for au-
tonomous driving”, in Proceedings of the IEEE International Conference
on Computer Vision (ICCV), 2019, 502–11.
[15] danielsyahputra, “KAN-YOLO”, 2024, https://github.com/danielsyahp
utra/ultralytics.
[16] X. Ding, H. Chen, X. Zhang, K. Huang, J. Han, and G. Ding, “Re-
parameterizing your optimizers rather than architectures”, in The In-
ternational Conference on Learning Representations (ICLR), 2023.
[17] X. Ding, X. Zhang, N. Ma, J. Han, G. Ding, and J. Sun, “RepVGG: Mak-
ing VGG-style ConvNets great again”, in Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 2021,
13733–42.
[18] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T.
Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al.,
“An image is worth 16x16 words: Transformers for image recognition at
YOLOv1 to YOLOv10 31

scale”, in The International Conference on Learning Representations

(ICLR), 2021.
[19] D. Erhan, C. Szegedy, A. Toshev, and D. Anguelov, “Scalable object
detection using deep neural networks”, in Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 2014,
2147–54.
[20] C. Feng, Y. Zhong, Y. Gao, M. R. Scott, and W. Huang, “TOOD: Task-
aligned one-stage object detection”, in Proceedings of the IEEE/CVF
International Conference on Computer Vision (ICCV), 2021, 3490–9.
[21] Z. Ge, S. Liu, Z. Li, O. Yoshie, and J. Sun, “OTA: Optimal transport
assignment for object detection”, in Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR),
2021, 303–12.
[22] Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “YOLOX: Exceeding YOLO
series in 2021”, arXiv preprint arXiv:2107.08430, 2021.
[23] R. Girshick, “Fast R-CNN”, in Proceedings of the IEEE International
Conference on Computer Vision (ICCV), 2015, 1440–8.
[24] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hier-
archies for accurate object detection and semantic segmentation”, in
Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), 2014, 580–7.
[25] J. Glenn, “YOLOv3 PyTorch”, 2019, https://github.com/ultralytics/y
olov3.
[26] J. Glenn, “YOLOv5 (2020.05)”, 2020, https://github.com/ultralytics/y
olov5.
[27] J. Glenn, “YOLOv5 release v1.0”, 2020, https://github.com/ultralytics
/yolov5/releases/tag/v1.0.
[28] J. Glenn, “YOLOv5 release v2.0”, 2020, https://github.com/ultralytics
/yolov5/releases/tag/v2.0.
[29] J. Glenn, “YOLOv5 release v3.0”, 2020, https://github.com/ultralytics
/yolov5/releases/tag/v3.0.
[30] J. Glenn, “YOLOv5 release v3.1”, 2020, https://github.com/ultralytics
/yolov5/releases/tag/v3.1.
[31] J. Glenn, “YOLOv5 release v4.0”, 2021, https://github.com/ultralytics
/yolov5/releases/tag/v4.0.
[32] J. Glenn, “YOLOv5 release v5.0”, 2021, https://github.com/ultralytics
/yolov5/releases/tag/v5.0.
[33] J. Glenn, “YOLOv5 release v6.0”, 2021, https://github.com/ultralytics
/yolov5/releases/tag/v6.0.
[34] J. Glenn, “YOLOv5 release v6.1”, 2022, https://github.com/ultralytics
/yolov5/releases/tag/v6.1.
[35] J. Glenn, “YOLOv5 release v6.2”, 2022, https://github.com/ultralytics
/yolov5/releases/tag/v6.2.
32 Wang and Liao

[36] J. Glenn, “YOLOv5 release v7.0”, 2022, https://github.com/ultralytics

/yolov5/releases/tag/v7.0.
[37] J. Glenn, “YOLOv8 (2023.01)”, 2023, https://github.com/ultralytics/u
ltralytics.
[38] J. Glenn, “YOLOv8 release v8.1.0”, 2024, https://github.com/ultralyti
cs/ultralytics/releases/tag/v8.1.0.
[39] J. Glenn, “YOLOv8 release v8.2.0”, 2024, https://github.com/ultralyti
cs/ultralytics/releases/tag/v8.2.0.
[40] J. Glenn, “YOLOv8 release v8.2.0’”, 2024, https://github.com/ultralyt
ics/ultralytics/releases.
[41] J. Glenn, “YOLOv8 release v8.3.0”, 2024, https://github.com/ultralyti
cs/ultralytics/releases/tag/v8.3.0.
[42] M. Gong, R. Liu, and V. K. Asari, “YOLO-based GNN for multi-
person pose estimation”, in Pattern Recognition and Tracking XXXV,
Vol. 13040, 2024, 124–31.
[43] A. Gu and T. Dao, “MAMBA: Linear-time sequence modeling with
selective state spaces”, arXiv preprint arXiv:2312.00752, 2023.
[44] Guanghan, “ROLO”, 2016, https://github.com/Guanghan/ROLO.
[45] C. Han, Q. Zhao, S. Zhang, Y. Chen, Z. Zhang, and J. Yuan, “YOLOPv2:
Better, faster, stronger for panoptic driving perception”, arXiv preprint
arXiv:2208.11434, 2022.
[46] K. Han, Y. Wang, Q. Tian, J. Guo, C. Xu, and C. Xu, “GhostNet:
More features from cheap operations”, in Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 2020,
1580–9.
[47] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition”, in Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition (CVPR), 2016, 770–8.
[48] K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in
deep convolutional networks for visual recognition”, IEEE Transactions
on Pattern Analysis and Machine Intelligence (TPAMI), 37(9), 2015,
1904–16.
[49] M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf, “Sup-
port vector machines”, IEEE Intelligent Systems and their Applications,
13(4), 1998, 18–28.
[50] S. Hochreiter and J. Schmidhuber, “Long short-term memory”, Neural
Computation, 9(8), 1997, 1735–80.
[51] X. Huang, X. Wang, W. Lv, X. Bai, X. Long, K. Deng, Q. Dang, S.
Han, Q. Liu, X. Hu, et al., “PP-YOLOv2: A practical object detector”,
arXiv preprint arXiv:2104.10419, 2021.
YOLOv1 to YOLOv10 33

[52] P. Hurtik, V. Molek, J. Hula, M. Vajgl, P. Vlasanek, and T. Nejezchleba,

“Poly-YOLO: higher speed, more precise detection and instance seg-
mentation for YOLOv3”, Neural Computing and Applications, 34(10),
2022, 8275–90.
[53] Y. Jiang, Z. Tan, J. Wang, X. Sun, M. Lin, and H. Li, “GiraffeDet:
A heavy-neck paradigm for object detection”, in The International
Conference on Learning Representations (ICLR), 2022.
[54] S. Kim, S. Park, B. Na, and S. Yoon, “Spiking-YOLO: spiking neural
network for energy-efficient object detection”, in Proceedings of the
AAAI Conference on Artificial Intelligence (AAAI), Vol. 34, No. 07,
2020, 11270–7.
[55] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T.
Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo, et al., “Segment anything”,
in Proceedings of the IEEE International Conference on Computer
Vision (ICCV), 2023, 4015–26.
[56] C. Li, L. Li, Y. Geng, H. Jiang, M. Cheng, B. Zhang, Z. Ke, X. Xu,
and X. Chu, “YOLOv6 v3.0: A full-scale reloading”, arXiv preprint
arXiv:2301.05586, 2023.
[57] C. Li, L. Li, H. Jiang, K. Weng, Y. Geng, L. Li, Z. Ke, Q. Li, M. Cheng,
W. Nie, et al., “YOLOv6: A single-stage object detection framework
for industrial applications”, arXiv preprint arXiv:2209.02976, 2022.
[58] C. Li, B. Zhang, L. Li, L. Li, Y. Geng, M. Cheng, X. Xiaoming, X. Chu,
and X. Wei, “YOLOv6: A single-stage object detection framework for
industrial applications”, 2023.
[59] X. Li, W. Wang, L. Wu, S. Chen, X. Hu, J. Li, J. Tang, and J. Yang,
“Generalized focal loss: Learning qualified and distributed bounding
boxes for dense object detection”, Advances in Neural Information
Processing Systems (NeurIPS), 33, 2020, 21002–12.
[60] C. Liang, Z. Zhang, X. Zhou, B. Li, S. Zhu, and W. Hu, “Rethinking
the competition between detection and reid in multiobject tracking”,
IEEE Transactions on Image Processing (TIP), 31, 2022, 3182–96.
[61] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie,
“Feature pyramid networks for object detection”, in Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition
(CVPR), 2017, 2117–25.
[62] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss
for dense object detection”, in Proceedings of the IEEE International
Conference on Computer Vision (ICCV), 2017, 2980–8.
[63] E. Linder-Norén, “YOLOv3 PyTorch”, 2018, https://github.com/erikli
ndernoren/PyTorch-YOLOv3.
[64] H. Liu, R. A. R. Soto, F. Xiao, and Y. J. Lee, “YOLACTEdge: Real-time
instance segmentation on the edge”, in IEEE International Conference
on Robotics and Automation (ICRA), 2021, 9579–85.
34 Wang and Liao

[65] R. Liu, J. Lehman, P. Molino, F. Petroski Such, E. Frank, A. Sergeev,

and J. Yosinski, “An intriguing failing of convolutional neural net-
works and the CoordConv solution”, Advances in Neural Information
Processing Systems (NeurIPS), 31, 2018.
[66] S. Liu, L. Qi, H. Qin, J. Shi, and J. Jia, “Path aggregation network for
instance segmentation”, in Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition (CVPR), 2018, 8759–68.
[67] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and
A. C. Berg, “SSD: Single shot multibox detector”, in Proceedings of the
European Conference on Computer Vision (ECCV), 2016, 21–37.
[68] X. Long, K. Deng, G. Wang, Y. Zhang, Q. Dang, Y. Gao, H. Shen,
J. Ren, S. Han, E. Ding, et al., “PP-YOLO: An effective and efficient
implementation of object detector”, arXiv preprint arXiv:2007.12099,
2020.
[69] X. Luo, M. Yao, Y. Chou, B. Xu, and G. Li, “Integer-Valued Train-
ing and Spike-Driven Inference Spiking Neural Network for High-
performance and Energy-efficient Object Detection”, in Proceedings
of the European Conference on Computer Vision (ECCV), 2024.
[70] N. Ma, X. Zhang, H.-T. Zheng, and J. Sun, “ShuffleNet v2: Practical
guidelines for efficient cnn architecture design”, in Proceedings of the
European conference on computer vision (ECCV), 2018, 116–31.
[71] D. Maji, S. Nagori, M. Mathew, and D. Poddar, “YOLO-Pose: En-
hancing YOLO for multi person pose estimation using object keypoint
similarity loss”, in Proceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition Workshops (CVPRW), 2022,
2637–46.
[72] W. McNally, K. Vats, A. Wong, and J. McPhee, “Rethinking key-
point representations: Modeling keypoints and poses as objects for
multi-person human pose estimation”, in Proceedings of the European
Conference on Computer Vision (ECCV), 2022, 37–54.
[73] meituan, “YOLOv6 (2022.06)”, 2022, https://github.com/meituan
/YOLOv6.
[74] E. Mohamed, A. Shaker, A. El-Sallab, and M. Hadhoud, “Insta-YOLO:
Real-time instance segmentation”, arXiv preprint arXiv:2102.06777,
2021.
[75] H. Ouyang, “DEYO: DETR with YOLO for End-to-End Object Detec-
tion”, arXiv preprint arXiv:2402.16370, 2024.
[76] H. Ouyang, “DEYO: DETR with YOLO for step-by-step object detec-
tion”, arXiv preprint arXiv:2211.06588, 2022.
[77] H. Ouyang, “DEYOv2: Rank feature with greedy matching for end-to-
end object detection”, arXiv preprint arXiv:2306.09165, 2023.
[78] H. Ouyang, “DEYOv3: DETR with YOLO for Real-time Object Detec-
tion”, arXiv preprint arXiv:2309.11851, 2023.
YOLOv1 to YOLOv10 35

[79] pjreddie, “darknet”, 2018, https://github.com/pjreddie/darknet.

[80] D. Qi, W. Tan, Q. Yao, and J. Liu, “YOLO5Face: Why reinventing a
face detector”, in Proceedings of the European Conference on Computer
Vision Workshops (ECCVW), 2022, 228–44.
[81] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal,
G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., “Learning transferable
visual models from natural language supervision”, in International
Conference on Machine Learning (ICML), 2021, 8748–63.
[82] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once:
Unified, real-time object detection”, in Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 2016,
779–88.
[83] J. Redmon and A. Farhadi, “YOLO9000: better, faster, stronger”, in
Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), 2017, 7263–71.
[84] J. Redmon and A. Farhadi, “YOLOv3: An incremental improvement”,
arXiv preprint arXiv:1804.02767, 2018.
[85] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards
real-time object detection with region proposal networks”, Advances in
Neural Information Processing Systems (NeurIPS), 28, 2015.
[86] ruhyadi, “YOLO3D”, 2022, https://github.com/ruhyadi/YOLO3D.
[87] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini,
“The graph neural network model”, IEEE Transactions on Neural
Networks, 20(1), 2008, 61–80.
[88] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y.
LeCun, “Overfeat: Integrated recognition, localization and detection
using convolutional networks”, in The International Conference on
Learning Representations (ICLR), 2014.
[89] Y. Shi, N. Wang, and X. Guo, “YOLOV: Making still image object
detectors great at video object detection”, in Proceedings of the AAAI
Conference on Artificial Intelligence (AAAI), Vol. 37, No. 2, 2023, 2254–
62.
[90] Y. Shi, T. Zhang, and X. Guo, “Practical Video Object Detection via
Feature Selection and Aggregation”, arXiv preprint arXiv:2407.19650,
2024.
[91] M. Simony, S. Milzy, K. Amendey, and H.-M. Gross, “Complex-YOLO:
An Euler-region-proposal for real-time 3D object detection on point
clouds”, in Proceedings of the European Conference on Computer Vision
Workshops (ECCVW), 2018.
[92] Z. Sun, M. Lin, X. Sun, Z. Tan, H. Li, and R. Jin, “MAE-Det: Revis-
iting maximum entropy principle in zero-shot nas for efficient object
detection”, in Proceedings of the International Conference on Machine
Learning (ICML), 2022.
36 Wang and Liao

[93] super-gradients, “YOLO-NAS”, 2023, https://github.com/Deci-AI/sup

er-gradients/blob/master/YOLONAS.md.
[94] M. Takahashi, Y. Ji, K. Umeda, and A. Moro, “Expandable YOLO: 3D
object detection from RGB-D images”, in International Conference on
Research and Education in Mechatronics (REM), 2020, 1–5.
[95] M. Tan, R. Pang, and Q. V. Le, “EfficientDet: Scalable and efficient
object detection”, in Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), 2020, 10781–90.
[96] A. Tavanaei, M. Ghodrati, S. R. Kheradpisheh, T. Masquelier, and
A. Maida, “Deep learning in spiking neural networks”, Neural Networks,
111, 2019, 47–63.
[97] B. Tekin, S. N. Sinha, and P. Fua, “Real-time seamless single shot 6D
object pose prediction”, in Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition (CVPR), 2018, 292–301.
[98] Z. Tian, C. Shen, H. Chen, and T. He, “FCOS: A simple and strong
anchor-free object detector”, IEEE Transactions on Pattern Analysis
and Machine Intelligence (TPAMI), 44(4), 2020, 1922–33.
[99] J. R. Uijlings, K. E. Van De Sande, T. Gevers, and A. W. Smeul-
ders, “Selective search for object recognition”, International Iournal of
Computer Vision (IJCV), 104, 2013, 154–71.
[100] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
Ł. Kaiser, and I. Polosukhin, “Attention is all you need”, Advances in
Neural Information Processing Systems (NeurIPS), 30, 2017.
[101] D. Vu, B. Ngo, and H. Phan, “HybridNets: End-to-end perception
network”, arXiv preprint arXiv:2203.09035, 2022.
[102] A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding,
“YOLOv10: Real-time end-to-end object detection”, Advances in Neural
Information Processing Systems (NeurIPS), 2024.
[103] C. Wang, W. He, Y. Nie, J. Guo, C. Liu, K. Han, and Y. Wang, “Gold-
YOLO: Efficient Object Detector via Gather-and-Distribute Mecha-
nism”, Advances in Neural Information Processing Systems (NeurIPS),
2023.
[104] C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “Scaled-YOLOv4:
Scaling cross stage partial network”, in Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), 2021,
13029–38.
[105] C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “YOLOv7: Trainable
bag-of-freebies sets new state-of-the-art for real-time object detectors”,
in Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), 2023, 7464–75.
YOLOv1 to YOLOv10 37

[106] C.-Y. Wang, H.-Y. M. Liao, Y.-H. Wu, P.-Y. Chen, J.-W. Hsieh, and I.-H.
Yeh, “CSPNet: A new backbone that can enhance learning capability
of CNN”, in Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition Workshops (CVPRW), 2020, 390–1.
[107] C.-Y. Wang, H.-Y. M. Liao, and I.-H. Yeh, “Designing network design
strategies through gradient path analysis”, Journal of Information
Science and Engineering (JISE), 39(4), 2023, 975–95.
[108] C.-Y. Wang, H.-Y. M. Liao, I.-H. Yeh, Y.-Y. Chuang, and Y.-L. Lin,
“Exploring the power of lightweight YOLOv4”, in roceedings of the IEEE
International Conference on Computer Vision Workshops (ICCVW),
2021, 779–88.
[109] C.-Y. Wang, I.-H. Yeh, and H.-Y. M. Liao, “YOLOv9: Learning what
you want to learn using programmable gradient information”, in Pro-
ceedings of the European Conference on Computer Vision (ECCV),
2024.
[110] C.-Y. Wang, I.-H. Yeh, and H.-Y. M. Liao, “You only learn one repre-
sentation: Unified network for multiple tasks”, Journal of Information
Science and Engineering (JISE), 39(2), 2023, 691–709.
[111] X. Wang, R. Zhang, T. Kong, L. Li, and C. Shen, “SOLOv2: Dy-
namic and fast instance segmentation”, Advances in Neural Information
Processing Systems (NeurIPS), 33, 2020, 17721–32.
[112] Z. Wang, C. Li, H. Xu, and X. Zhu, “Mamba YOLO: SSMs-Based
YOLO For Object Detection”, arXiv preprint arXiv:2406.05835, 2024.
[113] Z. Wang, L. Zheng, Y. Liu, Y. Li, and S. Wang, “Towards real-time
multi-object tracking”, in Proceedings of the European conference on
computer vision (ECCV), 2020, 107–22.
[114] N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime track-
ing with a deep association metric”, in IEEE International Conference
on Image Processing (ICIP), 2017, 3645–9.
[115] D. Wu, M.-W. Liao, W.-T. Zhang, X.-G. Wang, X. Bai, W.-Q. Cheng,
and W.-Y. Liu, “YOLOP: You only look once for panoptic driving
perception”, Machine Intelligence Research, 19(6), 2022, 550–62.
[116] S. Xu, X. Wang, W. Lv, Q. Chang, C. Cui, K. Deng, G. Wang, Q. Dang,
S. Wei, Y. Du, et al., “PP-YOLOE: An evolved version of YOLO”,
arXiv preprint arXiv:2203.16250, 2022.
[117] X. Xu, Y. Jiang, W. Chen, Y. Huang, Y. Zhang, and X. Sun, “DAMO-
YOLO: A report on real-time object detection design”, arXiv preprint
arXiv:2211.15444, 2022.
[118] J. Yang, S. Liu, Z. Li, X. Li, and J. Sun, “Real-time object detection
for streaming perception”, in Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition (CVPR), 2022, 5385–95.
38 Wang and Liao

[119] G. Yu, Q. Chang, W. Lv, C. Xu, C. Cui, W. Ji, Q. Dang, K. Deng,

G. Wang, Y. Du, et al., “PP-PicoDet: A better real-time object detector
on mobile devices”, arXiv preprint arXiv:2111.00902, 2021.
[120] Z. Yu, H. Huang, W. Chen, Y. Su, Y. Liu, and X. Wang, “YOLO-Facev2:
A scale and occlusion aware face detector”, Pattern Recognition, 155,
2024, 110714.
[121] J. Zhan, J. Liu, Y. Wu, and C. Guo, “Multi-Task Visual Perception for
Object Detection and Semantic Segmentation in Intelligent Driving”,
Remote Sensing, 16(10), 2024, 1774.
[122] J. Zhan, Y. Luo, C. Guo, Y. Wu, J. Meng, and J. Liu, “YOLOPX:
Anchor-free multi-task learning network for panoptic driving percep-
tion”, Pattern Recognition, 148, 2024, 110152.
[123] Y. Zhang, C. Wang, X. Wang, W. Zeng, and W. Liu, “FairMOT: On the
fairness of detection and re-identification in multiple object tracking”,
International Journal of Computer Vision (IJCV), 129, 2021, 3069–87.
[124] Z. Zhang, X. Lu, G. Cao, Y. Yang, L. Jiao, and F. Liu, “ViT-YOLO:
Transformer-based YOLO for object detection”, in Proceedings of the
IEEE International Conference on Computer Vision Workshops (IC-
CVW), 2021, 2799–808.
[125] X. Zhao, W. Ding, Y. An, Y. Du, T. Yu, M. Li, M. Tang, and J. Wang,
“Fast segment anything”, arXiv preprint arXiv:2306.12156, 2023.