Vehicle Detection and Classiﬁcation in Aerial Images using Convolutional

Neural Networks

Chih-Yi Li

and Huei-Yung Lin

Department of Electrical Engineering, National Chung Cheng University, Chiayi 621, Taiwan

Department of Electrical Engineering and Advanced Institute of Manufacturing with High-tech Innovations,

National Chung Cheng University, Chiayi 621, Taiwan

Keywords:

Aerial Image, Convolutional Neural Network, Vehicle Detection.

Abstract:

Due to the popularity of unmanned aerial vehicles, the acquisition of aerial images has become widely avail-

able. The aerial images have been used in many applications such as the investigation of roads, buildings,

agriculture distribution, and land utilization, etc. In this paper, we propose a technique for vehicle detection

and classiﬁcation from aerial images based on the modiﬁcation of Faster R-CNN framework. A new dataset for

vehicle detection, VAID (Vehicle Aerial Imaging from Drone), is also introduced for public use. The images

in the dataset are annotated with 7 common vehicle categories, including sedan, minibus, truck, pickup truck,

bus, cement truck and trailer, for network training and testing. We compare the results of vehicle detection in

aerial images with widely used network architectures and training datasets. The experiments demonstrate that

the proposed method and dataset can achieve high vehicle detection and classiﬁcation rates under various road

and trafﬁc conditions.

1 INTRODUCTION

In recent years, due to the popularity of unmanned

aerial vehicles (UAVs), the acquisition of aerial im-

ages has become more convenient. A large volume of

aerial images can be obtained very quickly. The use

of big data is a trend of the future research related

to aerial image analysis. The techniques for aerial

images has been adopted in many applications such

as the investigation of roads, buildings, agriculture

distribution, and land utilization, etc. One speciﬁc

importance is the detection of vehicles from aerial

imaging. This type of vehicle detection is suited for

transportation related work including trafﬁc monitor-

ing, vehicle identiﬁcation and tracking, parking anal-

ysis and planning. At the same time, the technol-

ogy maturity of UAV with the characteristics of being

lightweight, inexpensive and ﬂexible makes the aerial

photography easily apply to the trafﬁc data collection

and emergency response. Thus, it will be helpful and

efﬁcient in the applications if the analysis of aerial

images can be accelerated.

The content of aerial images generally covers a

large number of objects, including trees, lands, roads

and buildings, etc. In the early research, satellite im-

agery was used to analyze the landscape, the distri-

bution of forest, land usage, river and road areas. It

has also been used for the detection of special build-

ings or large venues in the past few decades, espe-

cially for the military purposes such as aircraft and

runway detection. Due to the advances of deep learn-

ing techniques in recent years, object detection can be

achieved under complex backgrounds and a variety of

application scenarios. Thus, it becomes more feasible

to use aerial images for the detection and classiﬁca-

tion of vehicles.

The methods for vehicle detection in aerial im-

ages are generally divided into two categories, the

traditional approaches and the machine learning tech-

niques (Cheng and Han, 2016). In traditional ap-

proaches, the feature extraction is one important step

in object detection, which consists of the use of tex-

ture, shape, color, and spatial information. The work

presented by Kembhavi et al. relies on three features,

histogram of oriented gradient (HoG), color probabil-

ity maps (CPM) and pairs of pixels (PoP), to solve the

regression problem using partial least squares (PLS)

(Kembhavi et al., 2011). Lenhart et al. use the dif-

ference in color channels to detect vehicles with more

signiﬁcant color features and the grayscale images to

extract the blob-like spots (Lenhart et al., 2008). In

(Shao et al., 2012), a vehicle detection framework

which combines different features including HoG and

local binary patterns (LBP) is proposed. Furthermore,

Li, C. and Lin, H.

Vehicle Detection and Classiﬁcation in Aerial Images using Convolutional Neural Networks.

DOI: 10.5220/0008941707750782

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages

775-782

ISBN: 978-989-758-402-2; ISSN: 2184-4321

775

many traditional techniques such as frame difference

and optical ﬂow (Yalcin et al., 2005) are used to detect

moving vehicles.

In recent years, the convolutional neural networks

(CNNs) have achieved good results in target detec-

tion and classiﬁcation. Among them, Tang et al. use

DLR Vehicle Aerial dataset and Vehicle Detection

in Aerial Imagery (VEDAI) dataset to train the Sin-

gle Shot MultiBox Detector (SSD) (Liu et al., 2016),

and adopt the model as the backbone to perform net-

work tuning (Tang et al., 2017). Sommer et al. eval-

uate the performance comparison between Fast R-

CNN (Girshick, 2015) and Faster R-CNN (Ren et al.,

2017) network architectures using public aerial im-

age datasets (Sommer et al., 2017). Deng et al. pro-

pose a fast vehicle detection system based on R-CNN.

It combines AVPN (Accurate Vehicle Proposal Net-

work) and VALN (Vehicle Attributes Learning Net-

work), and provides the results superior to Fast R-

CNN (Deng et al., 2017). Lu et al. analyze the differ-

ences among YOLO (Redmon et al., 2016), YOLOv2

(Redmon and Farhadi, 2017) and YOLOv3 (Redmon

and Farhadi, 2018) networks for vehicle detection us-

ing VEDAI, COWC and DOTA datasets (Xia et al.,

2018) for training and testing (Lu et al., 2018). Sim-

ilarly, Benjdira et al. present a small public vehicle

dataset to compare the pros and cons of Faster R-CNN

and YOLOv3 for vehicle detection (Benjdira et al.,

2019).

Compared to the general object detection and

recognition research and applications, the datasets for

vehicle detection in aerial imagery are fairly limited.

The VEDAI dataset used in this work is made avail-

able by Razakarivony and Jurie (Razakarivony and

Jurie, 2016), and originated from the public Utah

AGRC database. VEDAI contains a total of 1,250 im-

ages, and is manually annotated with nine classes of

objects (plane, boat, camping car, car, pickup truck,

tractor, truck, van, and others) and a total of 2,950

samples. The annotation of each sample includes

the sample class, the center point coordinates, di-

rection and the four corner point coordinates of the

groundtruth. However, the targets in VEDAI are

relatively easy to identify. Most of the vehicles in

the images are sparsely distributed with simple back-

grounds, and the vehicles in the densely distributed

places such as parking lots are excluded.

Compared to the object detection in ground view

images, vehicle detection in aerial images has several

challenges. The targets are usually much smaller with

monotonic appearance, and easily affected by the il-

lumination changes. In this paper, we propose a tech-

nique for vehicle detection and classiﬁcation in aerial

images based on a modiﬁed Faster R-CNN frame-

Figure 1: A schematic diagram of the proposed method for

vehicle detection and classiﬁcation in aerial images. It con-

sists of the creation of our own aerial image dataset and the

development of network architecture for vehicle detection.

work. We compare the advantages, disadvantages

and results of vehicle detection in aerial images with

widely used network architectures. A new aerial im-

age dataset for vehicle detection is introduced with the

annotation of 7 common vehicle categories including

sedan, minibus, truck, pickup truck, bus, cement truck

and trailer. Figure 1 shows the schematic diagram of

the proposed method for vehicle detection and classi-

ﬁcation in aerial images. It consists of the creation of

our own aerial image dataset and the development of

network architecture for vehicle detection.

2 VAID DATASET

As mentioned previously, there are not many public

datasets available speciﬁcally for vehicle detection in

aerial images. Even for the existing datasets, only

a very limited number of places and trafﬁc scenes

are covered. Several popular datasets for vehicle

detection in aerial images include VEDIA, COWC,

DLR-MVDA and KIT AIS. The description of these

datasets are shown in Table 1. This paper introduces a

new vehicle detection dataset, VAID (Vehicle Aerial

Imaging from Drone), with the aerial images captured

by a drone.

We collect about 6,000 aerial images

under different illumination conditions, viewing an-

gles from different places in Taiwan. The images are

taken with the resolution of 1137 × 640 pixels in JPG

format. Our VAID dataset contains seven classes of

vehicles, namely ‘sedan’, ‘minibus’, ‘truck’, ‘pickup

truck’, ‘bus’, ‘cement truck’ and ‘trailer’. Figure 2

VAID Dataset: http://vision.ee.ccu.edu.tw/aerialimage/

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

776

Table 1: Summary of the existing datasets for vehicle detection in aerial images.

Database Image Image Size Resolution Vehicle Size

VEDIA 1250

512 × 512

1024 × 1024

25cm

12.5cm

10 × 20

20 × 40

COWC 53 2000 × 2000, 19000 × 19000 15cm 24 × 48

DLR-MVDA 20 5616 × 3744 13cm 20 × 40

KIT AIS Dataset 241 300-1800 12.5cm-18cm 15 × 25, 20 × 40

Figure 2: The common vehicles are classiﬁed to 7 categories, namely (a) sedan, (b) minibus, (c) truck, (d) pickup truck, (e)

bus, (f) cement truck and (g) trailer, from the left to the right. The sample images are shown in the ﬁgure.

shows several cropped vehicle images from different

categories.

Although the vehicles are divided into the above

seven categories according to the popularity in Tai-

wan’s road scenes, it is sometimes very tricky to anno-

tate. The characteristics of small sedans viewing from

the above are less obvious, and the types are more di-

verse, including two-door and four-door sedans, ﬁve-

door hatchbacks, recreational vehicles and nine-seat

vans. There are a few differences in the deﬁnition of a

truck and a pickup truck for annotation. A truck is de-

ﬁned as a vehicle with a shelter in the cargo area or a

vehicle with its own cargo area as a container, and the

body and the front of the vehicle are completely dis-

connected. However, a pickup truck is not covered by

the canopy. A minibus is a 21-seat medium size bus,

while a bus includes passenger and big buses. The

trailer category includes tank trucks, gravel trucks,

tow trucks, container trucks with detachable tailgates.

The images in the dataset are annotated using the la-

beling tool LabelImg in the format of PASCAL VOC,

including the names of the classes and the bounding

box coordinates.

The images in the dataset are taken by a drone

(DJI’s Mavic Pro). To keep the sizes of the vehicles

consistent in all images, the altitude of the drone is

maintained at about 90 – 95 meters from the ground

during video recording. The output resolution is

2720 × 1530 at 2.7K and the frame rate is about 23.98

fps. For an average sedan with the length of 5 meters

and the width of 2.6 meters, the apparent size in the

image is about 110 × 45 pixels. In the VAID dataset,

the images are scaled to the resolution of 1137× 640,

and a sedan in the images is about the size of 40 × 20

pixels.

The dataset covers ten geographic locations in

southern Taiwan, and contains various trafﬁc and road

conditions. The images are taken on the sunny days

when the light is sufﬁcient, the interference caused

by the shadow of the house in the afternoon, and the

darker imaging condition in the evening. Figure 3

shows some of the dataset images with various road

and trafﬁc scenes.

3 PROPOSED FRAMEWORK

An overview of the proposed framework is illustrated

in Figure 4. It is modiﬁed based on Faster R-CNN

(Ren et al., 2017) and uses ResNet as the backbone

structure for feature learning. This provides high efﬁ-

ciency, robustness and effectiveness during training.

3.1 Feature Learning

In the original Faster R-CNN architecture, the authors

use VGG16 and ZF Net as the feature extraction net-

works. There exist many network architectures, such

as AlexNet, ResNet, and Inception, which also pro-

vide good results in feature extraction. ResNet is

one popular feature extraction network. On the eval-

uation using PASCAL VOC 2007, the mAP is in-

creased from 73.2% to 76.4% if VGG16 is replaced

by ResNet101. The mAP is also increased from

70.4% to 73.8% on PASCAL VOC 2012. The ex-

periments in (Ren et al., 2018) show that, among the

feature extraction networks for Faster R-CNN, the re-

sults obtained by ResNet50 are better than VGG and

Inception.

3.2 Region Proposal Network

In the feature extraction process of Faster R-CNN, the

RPN shares the convolutional layers, which can re-

duce a large amount of cost compared to the selective

search method. When using the sliding window, the

RPN generates multiple anchor boxes for matching,

Vehicle Detection and Classiﬁcation in Aerial Images using Convolutional Neural Networks

777

Figure 3: Some aerial images in our VAID dataset captured using a drone. It consists of different road and trafﬁc scenes.

Figure 4: The proposed modiﬁed Faster R-CNN architecture for vehicle detection and classiﬁcation.

which are used to effectively predict the aspect ratio

and scale settings of larger objects in PASCAL VOC.

Compared to the general object detection, the vehicles

in aerial images are clearly smaller than the common

objects of interest. Thus, if the original magniﬁca-

tion of the anchor box is used, the target might be too

small to be detected.

In the original implementation, the anchor boxes

are built on multiple scales and aspect ratios. The

three aspect ratios used are 0.5, 1 and 2, and the three

bounding box scales are 128

, 256

and 512

. The as-

pect ratios are changed to [0.2, 0.5, 1, 1.2, 2] to handle

to the small target size problem while maintaining the

recognition rate of large vehicles.

In the early work, the most commonly used acti-

vation function for DNN is the sigmoid function. To

solve the deﬁciencies of the sigmoid function, Recti-

ﬁed Linear Unit (ReLU) is increasingly used recently.

It is able to converge faster, and the computation is

less. The vanishing gradient problem can be effec-

tively solved with the observation whether the input

is greater than zero. ReLU gives part of the output

zero, which makes the neural network sparse and re-

duces the overﬁtting problem. Faster R-CNN has two

activation functions in the RPN, one for the classiﬁ-

cation and the other for the bounding box prediction.

In the experiments, it is found that the accuracy can

be improved if the activation function of the classiﬁ-

cation prediction is changed.

4 EXPERIMENTAL RESULTS

In the experiments, we introduce the VAID dataset

presented in this paper, and compare it with the exist-

ing aerial image datasets on the processes and results

of network training for vehicle detection and classiﬁ-

cation.

4.1 Dataset Comparison

We evaluate our method on the public VEDAI dataset

and our VAID dataset, and test with the aerial images

acquired from different places. The VEDAI dataset

is the most common dataset for vehicle detection in

aerial image. To focus on the targets for vehicle de-

tection, the training data of the VEDAI dataset are

processed as follows. (a) The objects in the cate-

gories with the labels ‘Boat’, ‘Plane’ and ‘Others’

are removed. (b) The remaining vehicles are labeled

by their types to two categories. The ﬁrst category

contains the general cars, which is labeled as ‘Car’

in the original dataset. The second category consists

of the combination of other types of vehicles which

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

778

(a) Scene A: A sedan is about 20 × 40 pixels. (b) Scene B: A sedan is about 10 × 20 pixels.

Figure 5: The image data used for testing. (a) Scene A consists of the images recorded from two locations in a city. (b) Scene

B contains a YouTube video recorded with a highway. (c) Scene C contains a YouTube video recorded with an expressway.

(d) Scene D is a YouTube video recorded with a crossroad in Belarus.

includes ‘Camping Car’, ‘Pickup Truck’, ‘Tractor’,

‘Truck’ and ‘Van’.

To compare with the VEDAI dataset, the origi-

nal seven vehicle categories in our dataset are also

adjusted to two categories. The ﬁrst category is the

general car, which is the vehicle classiﬁcation of the

original label ‘Sedan’, and the second category is the

other six original categories of vehicles labeled as

‘Minibus’, ‘Truck’, ‘Pickup Truck’, ‘Bus’, ‘Cement

truck’ and ‘Trailer’. Our testing data are selected

from four different scenes and image acquisition sce-

narios. Scene A consists of the self-recorded aerial

images from two locations in a city. Scenes B – D

are the videos obtained from YouTube, which record

two highways and one expressway in Taiwan, and a

crossroad in Belarus. The details of the testing data

are shown in Figure 5 and tabulated in Table 2.

Table 3 and Figure 6 show the comparison results

of using VEDAI and VAID as the training datasets. In

the four different scenarios, training with the VEDAI

dataset is worse than using our VAID dataset, espe-

cially for Scenes A and D in the testing environment.

In VEDAI images, the vehicle only occupies a small

region, so the classiﬁcation error is high when the ve-

hicle size is relatively large in the testing images such

as Scene D. Scenes A – C are the road data acquired

in Taiwan, and the vehicles such as trucks and trail-

ers are rare in the VEDAI dataset. This also causes

the classiﬁcation problem for certain types of vehi-

cles. Using our VAID dataset for training, high accu-

racy results are obtained for Scenes A, C and D. Es-

pecially, note that Scene D is the road data acquired

in Belarus with the captured vehicle size larger than

those in our training set. The low precision results of

Scene B are mainly caused by the vehicle size in the

scene (about 20 × 10) much smaller compared to the

size of about 40 × 20 used for training.

4.2 Detection and Classiﬁcation

Comparison

In our VAID dataset, there are totally 5,937 aerial im-

ages with vehicles classiﬁed into 7 categories. It in-

cludes 4,456 images for training, 502 images for ver-

iﬁcation, and 979 images for testing. The detailed in-

formation of each class and the number of images for

training, validation and testing are shown in Table 4.

It can be seen that the number of training samples is

not balanced for different classes, so it is very impor-

tant to achieve better results with fewer samples when

training the classiﬁcation network.

To test the effect of different activation func-

tions, we evaluate the network using softplus, ELU

and ReLU, and the results are tabulated in Table

5. Among these activation functions, softplus is the

worst, while ELU and ReLU converge faster than the

original architecture and the mAP is also improved by

about 0.4%. Thus, ReLU is adopted as the activation

function of RPN for Faster R-CNN classiﬁcation.

In this paper, Faster R-CNN with ResNet101 as

Vehicle Detection and Classiﬁcation in Aerial Images using Convolutional Neural Networks

779

Table 2: The description of the training and testing sets.

Training Set Testing Set

Number of image Image size Number of image Image size

VEDAI Dataset 1211 1024 × 1024

A: 99 1137×640

B. 17 1280 ×720

VAID Dataset 4456 1137×640

C: 31 1280×720

D: 35 1920×1080

Table 3: The result comparison (mAP) of using the VEDAI dataset and our VAID dataset for network training.

Training Data VEDAI Dataset VAID Dataset

Class Sedan Others mAP Sedan Others mAP

A 18.8% 0% 9.5% 90.4% 62.6% 76.5%

B 59.4% 35.8% 45.4% 73% 55.2% 64.1%

C 31.1% 23.3% 27.2% 89.1% 86.8% 87.9%

D 12.6% 0% 6.7% 89.1% 90.7% 89.9%

(a) Scene A (b) Scene B

Figure 6: The results of Scenes A – D using the proposed vehicle detection approach. Left: Training with the VEDAI dataset.

Right: Training with our VAID dataset.

the feature extraction network is modiﬁed. It includes

the data enhancement of the training images, the ad-

justment of the anchor box size, the change of the

test image input size, and the modiﬁcation of the

ResNet101 activation function. Table 6 shows the

results of the original Faster R-CNN, our modiﬁed

Faster R-CNN, and the modiﬁcation with the ReLU

activation function for RPN. Using the same train-

ing data, the improvement with the pre-adjustment is

about 1.5% in mAP, and the result of each category

is also slightly improved. Figure 7 shows the vehicle

detection and classiﬁcation results in various common

road scenes in Taiwan. The bounding boxes with dif-

ferent colors are used to represent different types of

vehicles. Successful detection and classiﬁcation with

all kinds of vehicles are shown in the ﬁgures.

5 CONCLUSIONS

In this paper, we propose a technique for vehicle de-

tection and classiﬁcation from aerial images based on

the modiﬁcation of Faster R-CNN framework. A new

dataset with images collected from a drone is further

introduced for vehicle detection and available for pub-

lic access. We compare the results of vehicle detec-

tion in aerial images with widely used network ar-

chitectures and training datasets. The experiments

demonstrate that the proposed method and dataset can

achieve high vehicle detection and classiﬁcation rates

under various road and trafﬁc conditions. In the future

work, the vehicle detection will be incorporated with

trafﬁc management system for parking lot and trafﬁc

ﬂow monitoring.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

780

Table 4: The detailed information of each class and the number of images used for training, validation and testing.

Class Name Sedan Minibus Truck Pickup Truck Bus Cement Truck Trailer

Training 29774 392 2382 2411 447 144 595

Validation 3483 39 267 266 49 21 62

Testing 6774 70 531 322 84 22 56

Table 5: Different activation functions are used in Faster R-CNN for comparison (mAP).

Modiﬁed

Faster R-CNN

Modiﬁed

Faster R-CNN

(softplus)

Modiﬁed

Faster R-CNN

(ELU)

Modiﬁed

Faster R-CNN

(ReLU)

Sedan 90.2% 90.2% 90.2% 90.2%

Minibus 97.9% 92.2% 96.6% 95.6%

Truck 84.9% 87.5% 87.1% 86.8%

Pickup Truck 78.7% 79.3% 78.6% 79.6%

Bus 89.4% 89.6% 89.8% 90.3%

Cement Truck 97.6% 93.8% 97.6% 98.1%

Trailer 83.8% 81.5% 84.7% 84.5%

Avg. 88.9% 87.7% 89.2% 89.3%

Table 6: The results from our modiﬁed Faster R-CNN (mAP).

Original Faster R-CNN Modiﬁed Faster R-CNN

Modiﬁed Faster R-CNN

(ReLU)

Sedan 90.0% 90.2% 90.2%

Minibus 95.0% 97.9% 95.6%

Truck 83.4% 84.9% 86.8%

Pickup Truck 76.8% 78.7% 79.6%

Bus 88.9% 89.4% 90.3%

Cement Truck 94.2% 97.6% 98.1%

Trailer 85.6% 83.8% 84.5%

Avg. 87.7% 88.9% 89.3%

Figure 7: The results from our dataset and the modiﬁed Faster R-CNN with ReLU.

Vehicle Detection and Classiﬁcation in Aerial Images using Convolutional Neural Networks

781

ACKNOWLEDGMENTS

The support of this work in part by the Ministry

of Science and Technology of Taiwan under Grant

MOST 106-2221-E-194-004 and the Advanced In-

stitute of Manufacturing with High-tech Innovations

(AIM-HI) from The Featured Areas Research Cen-

ter Program within the framework of the Higher Ed-

ucation Sprout Project by the Ministry of Education

(MOE) in Taiwan is gratefully acknowledged.

REFERENCES

Benjdira, B., Khursheed, T., Koubaa, A., Ammar, A., and

Ouni, K. (2019). Car Detection using Unmanned

Aerial Vehicles: Comparison between Faster R-CNN

and YOLOv3Benjdira, B., Khursheed, T., Koubaa,

A., Ammar, A., &Ouni, K. (2019). Car Detection

using Unmanned Aerial Vehicles: Comparison be-

tween Faster R-CNN and YOLOv3. 2019 1s. In 2019

1st International Conference on Unmanned Vehicle

Systems-Oman, UVS 2019.

Cheng, G. and Han, J. (2016). A survey on object detection

in optical remote sensing images.

Deng, Z., Sun, H., Zhou, S., Zhao, J., and Zou, H. (2017).

Toward Fast and Accurate Vehicle Detection in Aerial

Images Using Coupled Region-Based Convolutional

Neural Networks. IEEE Journal of Selected Topics in

Applied Earth Observations and Remote Sensing.

Girshick, R. (2015). Fast R-CNN. In Proceedings of the

IEEE International Conference on Computer Vision.

Kembhavi, A., Harwood, D., and Davis, L. S. (2011). Vehi-

cle Detection using Partial Least Squares. IEEE Trans.

Pattern Analysis and Machine Intelligence (PAMI),

33(6):1250–1265.

Lenhart, D., Hinz, S., Leitloff, J., and Stilla, U. (2008). Au-

tomatic trafﬁc monitoring based on aerial image se-

quences. Pattern Recognition and Image Analysis.

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.,

Fu, C. Y., and Berg, A. C. (2016). SSD: Single shot

multibox detector. In Lecture Notes in Computer Sci-

ence (including subseries Lecture Notes in Artiﬁcial

Intelligence and Lecture Notes in Bioinformatics).

Lu, J., Ma, C., Li, L., Xing, X., Zhang, Y., Wang, Z., and

Xu, J. (2018). A Vehicle Detection Method for Aerial

Image Based on YOLO. Journal of Computer and

Communications.

Razakarivony, S. and Jurie, F. (2016). Vehicle detection in

aerial imagery: A small target detection benchmark.

Journal of Visual Communication and Image Repre-

sentation.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time ob-

ject detection. In Proceedings of the IEEE Computer

Society Conference on Computer Vision and Pattern

Recognition.

Redmon, J. and Farhadi, A. (2017). YOLO9000: Better,

faster, stronger. In Proceedings - 30th IEEE Con-

ference on Computer Vision and Pattern Recognition,

CVPR 2017.

Redmon, J. and Farhadi, A. (2018). YOLO v.3. Tech report.

Ren, S., He, K., Girshick, R., and Sun, J. (2017). Faster R-

CNN: Towards Real-Time Object Detection with Re-

gion Proposal Networks. IEEE Transactions on Pat-

tern Analysis and Machine Intelligence.

Ren, Y., Zhu, C., and Xiao, S. (2018). Object Detection

Based on Fast/Faster RCNN Employing Fully Con-

volutional Architectures. Mathematical Problems in

Engineering.

Shao, W., Yang, W., Liu, G., and Liu, J. (2012). Car detec-

tion from high-resolution aerial imagery using multi-

ple features. In International Geoscience and Remote

Sensing Symposium (IGARSS).

Sommer, L. W., Schuchert, T., and Beyerer, J. (2017). Fast

deep vehicle detection in aerial images. In Proceed-

ings - 2017 IEEE Winter Conference on Applications

of Computer Vision, WACV 2017.

Tang, T., Zhou, S., Deng, Z., Lei, L., and Zou, H. (2017).

Arbitrary-oriented vehicle detection in aerial imagery

with single convolutional neural networks. Remote

Sensing.

Xia, G. S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J.,

Datcu, M., Pelillo, M., and Zhang, L. (2018). DOTA:

A Large-Scale Dataset for Object Detection in Aerial

Images. In Proceedings of the IEEE Computer Society

Conference on Computer Vision and Pattern Recogni-

tion.

Yalcin, H., Hebert, M., Collins, R., and Black, M. J. (2005).

A ﬂow-based approach to vehicle detection and back-

ground mosaicking in airborne video. In Proceedings

- 2005 IEEE Computer Society Conference on Com-

puter Vision and Pattern Recognition, CVPR 2005.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

782