0% found this document useful (0 votes)

20 views

Object_Detection_in_Aerial_Images_A_Large_Scale_Benchmark_and_Challenges

This paper introduces the Dataset of Object Detection in Aerial images (DOTA), which includes 1,793,658 object instances across 18 categories from 11,268 aerial images, addressing the lack of large-scale benchmarks for object detection in aerial images. The authors provide comprehensive baselines for 10 state-of-the-art algorithms, a code library for oriented object detection, and a website for algorithm evaluation, facilitating robust algorithm development in this field. The expanded DOTA dataset and associated resources aim to enhance reproducible research and advance the understanding of object detection challenges in aerial imagery.

Uploaded by

Haytham Kenway

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

20 views

Object_Detection_in_Aerial_Images_A_Large_Scale_Benchmark_and_Challenges

Uploaded by

Haytham Kenway

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 18

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TPAMI.2021.3117983, IEEE Transactions on Pattern Analysis and Machine Intelligence
1

Object Detection in Aerial Images:

A Large-Scale Benchmark and Challenges
Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang,
Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, Liangpei Zhang

Abstract—In the past decade, object detection has achieved significant progress in natural images but not in aerial images, due to the
massive variations in the scale and orientation of objects caused by the bird’s-eye view of aerial images. More importantly, the lack of
large-scale benchmarks has become a major obstacle to the development of object detection in aerial images (ODAI). In this paper,
we present a large-scale Dataset of Object deTection in Aerial images (DOTA) and comprehensive baselines for ODAI. The proposed
DOTA dataset contains 1,793,658 object instances of 18 categories of oriented-bounding-box annotations collected from 11,268 aerial
images. Based on this large-scale and well-annotated dataset, we build baselines covering 10 state-of-the-art algorithms with over 70
configurations, where the speed and accuracy performances of each model have been evaluated. Furthermore, we provide a code
library for ODAI and build a website for evaluating different algorithms. Previous challenges run on DOTA have attracted more than 1300
teams worldwide. We believe that the expanded large-scale DOTA dataset, the extensive baselines, the code library and the challenges
can facilitate the designs of robust algorithms and reproducible research on the problem of object detection in aerial images.

Index Terms—Object detection, remote sensing, aerial images, oriented object detection, benchmark dataset.

F
1 I NTRODUCTION
Currently, Earth vision (also known as Earth observation half meter. Although challenging, developing mathematical
and remote sensing) technologies enable us to observe the tools and numerical algorithms is necessary for interpreting
Earth‘s surface with aerial images1 with a resolution up to a these huge volumes of images, among which object detec-
tion refers to localizing objects of interest (e.g., vehicles and
ships) on the Earth‘s surface and predicting their categories.
• J. Ding and L. Zhang are with the State Key Lab. LIES- Object detection in aerial images (ODAI) has been an es-
MARS, Wuhan University, Wuhan, 430079, China. Email: {jian.ding,
zlp62}@whu.edu.cn. sential step in many real-world applications such as urban
• N. Xue is with the National Engineering Research Center for Multi- management, precision agriculture, emergency rescue and
media Software, School of Computer Science and Institute of Artificial disaster relief [1], [2]. Although extensive studies have been
Intelligence, Wuhan University, Wuhan, 430072, China. Email: xue-
[email protected].
devoted to object detection in aerial images and apprecia-
• G.-S. Xia is with the National Engineering Research Center for Multi- ble breakthroughs have been made [3]–[8], the task has
media Software, School of Computer Science and Institute of Artificial numerous difficulties such as arbitrary orientations, scale
Intelligence, and also the State Key Lab. LIESMARS, Wuhan University, variations, extremely nonuniform object densities and large
Wuhan, 430072, China. Email: [email protected].
• X. Bai is with the School of Electronic Information, Huazhong Uni- aspect ratios (ARs), as shown in Fig. 1.
versity of Science and Technology, Wuhan, 430079, China. Email: Among these difficulties, the arbitrary orientation of
[email protected]. objects caused by the overhead view is the main differ-
• W. Yang is with the School of Electronic Information, and the State Key
Lab. LIESMARS, Wuhan University, Wuhan, 430072, China. Email:
ence between natural images and aerial images, and it
[email protected]. complicates the object detection task in two ways. First,
• M. Y. Yang is with Faculty of Geo-Information Science and Earth rotation-invariant feature representations are preferred in
Observation (ITC), University of Twente, The Netherlands. Email: the detection of arbitrarily orientated objects, but they are
[email protected].
• S. Belongie is with Department of Computer Science, Cornell University often beyond the capability of most of current deep neural
and Cornell Tech. Email: [email protected]. network models. Although the methods such as those de-
• J. Luo is with Department of Computer Science, University of Rochester, signed in [6], [9], [10] use rotation-invariant convolutional
Rochester, NY 14627. Email: [email protected].
• M. Datcu is with Remote Sensing Technology Institute, German
neural networks (CNNs), the problem is far from solved.
Aerospace Center (DLR), 82234, Germany. Email: [email protected]. Second, the horizontal bounding box (HBB) object representa-
• M. Pelillo is with DAIS, Ca’ Foscari University of Venice, Italy. Email: tion used in conventional object detection [11]–[13] cannot
[email protected]. localize the oriented objects precisely, such as ships and
• The studies in this paper have been supported by the NSFC projects under
the contracts No.61922065, No.61771350 and No.41820104006. Dr. Nan large vehicles, as shown in Fig. 1. The oriented bounding box
Xue was also supported by National Post-Doctoral Program for Innovative (OBB) object representation is more appropriate for aerial
Talents under Grant BX20200248. images [4], [14]–[17]. It allows us to distinguish densely
• The corresponding author is Gui-Song Xia ([email protected]).
packed instances (as shown in Fig. 3) and extract rotation-
1. This paper uses the term ”aerial” to refer to any overhead image invariant features [4], [18], [19]. The OBB object representa-
looking approximately straight down onto the Earth, including both tion actually introduces a new object detection task, called
satellite images and airborne images, for simplification unless other-
wise indicated. We use the term “airborne” if we do not want to include oriented object detection. In contrast with horizontal object
the satellite images. detection [8], [20]–[22], oriented object detection is a recently

0162-8828 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more
Authorized licensed use limited to: UNIVERSITY OF TWENTE.. Downloaded on March 10,2022 at 08:20:16 UTC from IEEE Xplore. Restrictions apply.
information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TPAMI.2021.3117983, IEEE Transactions on Pattern Analysis and Machine Intelligence
2

Fig. 1. An example image taken from DOTA. (a) A typical image in DOTA consisting of many instances from multiple categories. (b), (c), (d), (e)
are cropped from the source image. We can see that instances such as small vehicles have arbitrary orientations. There is also a massive scale
variation across different instances. Moreover, the instances are not distributed uniformly. The instances are sparse in most areas but crowded in
some local areas. Large vehicles and ships have large ARs. (f) and (g) exhibit the size and orientation histograms, respectively, for all instances.

emerging research direction and most of the methods for is highly desirable. Nevertheless, current object detection
this new task attempt to transfer successful deep-learning- libraries, e.g., MMDetection [28] and Detectron [29], do not
based object detectors pre-trained on large-scale natural support oriented object detection.
image datasets (e.g., ImageNet [12] and Microsoft Common To address the above-mentioned problems, in this paper
Objects in Context (MS COCO) [13]) to aerial scenes [18], we first extend the preliminary version of DOTA, i.e., DOTA-
[19], [23]–[25] due to the lack of large-scale annotated aerial v1.0 [14], to DOTA-v2.0. Specifically, DOTA-v2.0 collects
image datasets. 11, 268 aerial images from various sensors and platforms
To mitigate the dataset problem, some public datasets and contains approximately 1.8 million object instances
of aerial images have been created, see e.g. [7], [15]–[17], annotated with OBBs in 18 common categories, which, to
[26], but they contain a limited number of instances and our knowledge, is the largest public Earth vision object de-
tend to use images taken under ideal conditions (e.g., clear tection dataset. Then, to facilitate algorithm developments
backgrounds and centered objects), which cannot reflect the and comparisons with DOTA, we provide a well-designed
real-world difficulties of the problem. The recently released code library that supports oriented object detection in aerial
xView [27] dataset provides a wide range of categories and images. Based on the code library, we also build more
contains large quantities of instances in complicated scenes. comprehensive baselines than the preliminary version [14],
However, it annotates the instances with HBBs instead of keeping the hardware, software platform, and settings the
the more precise OBBs. Thus, a large-scale dataset that has same. In total, we evaluate 10 algorithms and over 70 mod-
OBB annotations and reflects the difficulties in real-world els with different configurations. We then provide detailed
applications of aerial images is in high demand. speed and accuracy analyses to explore the module designs
Another issue with ODAI is that the module design and parameter settings in aerial images to guide future
and the hyperparameter setting of conventional object de- research. These experiments verify the large differences in
tectors learned from natural images are not appropriate object detector design between natural and aerial images
for aerial images due to domain differences. Thus, when and provide materials for universal object detection algo-
developing new algorithms, comprehensive baselines and rithms [30].
sufficient ablative analyses of models on aerial images are The main contributions of this paper are three-fold:
required. However, comparing different algorithms is diffi-
cult due to the diversities in hardware, software platforms, • To the best of our knowledge, the expanded DOTA is
detailed settings and so on. These factors influence both the largest dataset for object detection in Earth vision.
speed and accuracy. Therefore, when building the baselines, The OBB annotations of DOTA not only provide a
implementing the algorithms with a unified code library large-scale benchmark for object detection in Earth
and keeping the hardware and software platform the same vision but also pose interesting algorithmic questions

and challenges to generalized object detection in datasets of natural and aerial images. Moreover, the large
computer vision. images include more instances per image and contextual
• We build a code library for object detection in aerial information. Tab. 1 provides the detailed comparisons.
images. This is expected to facilitate the development
and benchmarking of object detection algorithms in TABLE 1
aerial images with both HBB and OBB representa- DOTA vs. general object detection datasets. BBox is short for bounding
tions. box, Avg. BBox quantity indicates the average number of bounding
boxes per image. For PASCAL VOC (07++12), we count the whole
• With the expanded DOTA, we evaluate 10 repre- PASCAL VOC 07 and training and validation (trainval) set of PASCAL
sentative algorithms over 70 model configurations, VOC 12. DOTA has a comparable scale with the large-scale datasets
providing comprehensive analysis that can guide for object detection in natural images. Note that for the average number
of instances per image, DOTA surpasses the other datasets.
the designs of object detection algorithms in aerial
Image Megapixel BBox Avg. BBox
images. Dataset Classes
quantity area quantity quantity
PASCAL VOC
The dataset, code library, and regular evaluation server (07++12)
20 21,503 5,133 52,090 2.42
are available and maintained on the DOTA website2 . It MS COCO
80 123,287 32,639 886,266 7.19
is worth noting that the creation and use of DOTA have (2014 trainval)
ImageNet
advanced object detection in aerial images. For instance, the (2014 train)
200 456,567 82,820 478,807 1.05
regular DOTA evaluation server and two object detection DOTA-v1.0 15 2,806 19,173 188,282 67.10
contests organized at the 2018 International Conference DOTA-v1.5 16 2,806 19,173 402,089 143.73
DOTA-v2.0 18 11,268 126,306 1,793,658 159.18
on Pattern Recognition (ICPR’ 2018 with DOTA-v1.0)3 and
2019 Conference on Computer Vision and Pattern Recogni-
tion (CVPR’2019 with DOTA-v1.5)4 have attracted approx- 2.2 Datasets for Object Detection in Aerial Images
imately 1300 registrations. We believe that our new DOTA
In aerial object detection, a dataset resembling MS COCO
dataset, with a comprehensive code library and an online
and ImageNet both in terms of the image number and de-
evaluation platform, will further promote the reproducible
tailed annotations has been missing, which becomes one of
research in Earth vision.
the main obstacles to research in Earth vision, especially for
developing deep-learning-based algorithms. In Earth vision,
2 R ELATED WORK many aerial image datasets are prepared for actual demands
Well-annotated datasets have played an important role in in a specific category, such as building datasets [7], [36], ve-
data-driven computer vision research [12], [13], [31]–[35] hicle datasets [8], [15], [16], [26], [37]–[39], ship datasets [4],
and have promoted cutting-edge research in a number of [40], and plane datasets [17], [41]. Although some public
tasks such as object detection and classification. In this datasets [17], [42]–[45] have multiple categories, they have
section, we first review object detection datasets of natural only limited number of samples, which are hardly efficient
and aerial images. Then we discuss the recent deep learning for training robust deep models. For example, NWPU [42]
based object detectors in aerial images. Finally, we briefly only contains 800 images, 10 classes and 3, 651 instances.
review the code libraries for object detection. To alleviate this problem, our preliminary work DOTA-
v1.0 [14] presented a dataset with 15 categories and 188, 282
2.1 Datasets for Conventional Object Detection instances, which for the first time enables us to efficiently
train robust deep models for ODAI without the help of
As a pioneer, PASCAL Visual Object Classes (VOC) [11] large-scale datasets of natural images, such as MS COCO
held challenges on object detection from 2005 to 2012. The and ImageNet. Later, iSAID [46] provided an instance seg-
computer vision community has widely adopted PASCAL mentation extension of DOTA-v1.0 [14]. A notable dataset
VOC datasets and their evaluation metrics. Specifically, is xView [27], which contains 1, 413 images, 16 main cate-
the PASCAL VOC Challenge 2012 dataset contains 11, 530 gories, 60 fine-grained categories, and 1 million instances.
images, 20 classes, and 27, 450 annotated bounding boxes. Another dataset DIOR [47] provided a comparable number
Later, the ImageNet dataset [12] was developed and is an of instances as DOTA-v1.0 [14]. However, the instances in
order of magnitude larger than PASCAL VOC, containing xView and DIOR are both annotated by HBBs, which are
200 classes and approximately 500, 000 annotated bounding not suitable for precisely detecting objects that are arbitrarily
boxes. However, non-iconic views are not addressed. Then oriented in aerial images. In addition, VisDrone [48] is also
MS COCO [13] was released, containing a total of 328K a large-scale dataset for drone images but focuses more on
images, 91 categories, and 2.5 million labeled segmented video object detection and tracking. The image subset in
objects. MS COCO has on average more instances and cate- VisDrone for object detection is not very large. Furthermore,
gories per image and contains more contextual information most of the previous datasets are heavily imbalanced in
than PASCAL VOC and ImageNet. It is worth noticing that, favor of positive samples, whose negative samples are not
in Earth vision, the image size could be extremely large (e.g., sufficient to represent the real-world distribution.
20, 000 × 20, 000 pixels), so the number of images cannot As we stated previously [14], a good dataset for aerial
reflect the scale of a dataset. In this case, the pixel area would
image object detection should have the following proper-
be more reasonable when comparing the scale between the
ties: 1) substantial annotated data to facilitate data-driven,
2. https://captain-whu.github.io/DOTA/ especially deep-learning-based methods; 2) large images to
3. https://captain-whu.github.io/ODAI/results.html contain more contextual information; 3) OBB annotation to
4. https://captain-whu.github.io/DOAI2019/challenge.html describe the precise location of objects; and 4) balance in

TABLE 2
DOTA vs. object detection datasets in aerial images. HBB is horizontal bounding box, and OBB is oriented bounding box . CP is center point.
# of main Total # of Image
Dataset Source Annotation # of instances # of images Year
categories categories width
TAS [26] satellite HBB 1 1 1,319 30 792 2008
SZTAKI-INRIA [7] multi source OBB 1 1 665 9 ∼800 2012
NWPU VHR-10 [42] multi source HBB 10 10 3,651 800 ∼1000 2014
VEDAI [15] satellite OBB 3 9 2,950 1,268 512, 1024 2015
DLR 3k [16] aerial OBB 2 8 14,235 20 5616 2015
UCAS-AOD [17] Google Earth OBB 2 2 14,596 1,510 ∼1000 2015
COWC [37] aerial CP 1 1 32,716 53 2000−19,000 2016
HRSC2016 [4] Google Earth OBB 1 26 2,976 1,061 ∼1100 2016
RSOD [43] Google Earth HBB 4 4 6,950 976 ∼1000 2017
CARPPK [8] drone HBB 1 1 89,777 1,448 1280 2017
ITCVD [38] aerial HBB 1 1 228 23,543 5616 2018
LEVIR [44] Google Earth HBB 3 3 11,000 22,000 800−600 2018
xView [27] satellite HBB 16 60 1,000,000 1,413 ∼3000 2018
VisDrone [48] drone HBB 10 10 54,200 10,209 2000 2018
SpaceNet MVOI [36] satellite polygon 1 1 126,747 60,000 900 2019
HRRSD [45] multi source HBB 13 13 55,740 21,761 152−10569 2019
DIOR [47] Google Earth HBB 20 20 190,288 23,463 800 2019
iSAID [46] multi source polygon 14 15 655,451 2,806 800−13,000 2019
FGSD [40] Google Earth OBB 1 43 5,634 2,612 930 2020
RarePlanes [41] satellite polygon 1 110 644,258 50,253 1080 2020
DOTA-v1.0 [14] multi source OBB 14 15 188,282 2,806 800−13,000 2018
DOTA-v1.5 multi source OBB 15 16 402,089 2,806 800−13,000 2019
DOTA-v2.0 multi source OBB 17 18 1,793,658 11,268 800−20,000 2021

image sources, as pointed in [49]. DOTA is built considering feature pyramids [19], [58] and image pyramids [24], [25]
these principles (unless otherwise specified, DOTA refers to are widely used to extract scale-invariant features in aerial
DOTA-v2.0). Detailed comparisons of these existing datasets images. We evaluate the geometric transformation network
and DOTA are shown in Tab. 2. Compared to other aerial modules and geometric data augmentations in Sec. 6.1.
datasets, as we shall see in Sec. 4, DOTA is challenging due Crowded instances represented by HBBs are difficult to
to its large number of object instances, arbitrary orientations, distinguish (see Fig. 3). Traditional HBB-based non maxi-
various categories, density distribution, and diverse aerial mum suppression (NMS) will fail in such cases. Therefore,
scenes from various image sources. These properties make these methods [18], [24], [25] use rotated NMS (R-NMS),
DOTA helpful for real-world applications. which requires precise detections to address this problem.
Similar to text and face detection in natural scenes, e.g. [23],
[59]–[61], precise ODAI can also be modeled as an oriented
2.3 Deep Models for Object Detection in Aerial Images object detection task. Most of the previous works [14], [23]–
Object detection in aerial images is a longstanding problem. [25], [62] consider it as a regression problem and regress
Recently, with the development of deep learning, many the offsets of the OBB ground truth relative to anchors
researchers in Earth vision have adapted deep object detec- (or proposals). However, the definition of an OBB is am-
tors [50]–[54] developed for natural images to aerial images. biguous. For example, there are four permutations of the
However, the challenges caused by the domain shift need to corner points in a quadrilateral. The Faster R-CNN OBB [14]
be addressed. Here, we highlight some notable works. solves it by using a defined rule to determine the order
Objects in aerial images are often arbitrarily oriented due of points in OBBs. Work in [63] further uses the gliding
to the bird’s-eye view, and the scale variations are larger offset and obliquity factor to eliminate the ambiguity. The
than those in natural images. To handle rotation variations, circular smooth label (CSL) [64] transforms the regression of
a simple model [9] plugs an additional rotation-invariant angle as a classification problem to avoid the problem. Mask
layer into R-CNN [52] relying on rotation data augmenta- OBB [65] and CenterMap [66] consider object detection as a
tion. The oriented response network (ORN) introduces ac- pixel-level classification problem to avoid ambiguity. Mask-
tive rotating filters (ARF) to produce the rotation-invariant based methods converge more easily but have more floating
feature without using data augmentation, which is adopted point operations per second (FLOPS) than regression-based
by the rotation-sensitive regression detector (RRD) [23]. The methods. We will give a more detailed comparison between
deformable modules [55] designed for general object defor- them in one unified code library in Sec. 6.1.1.
mation are also widely used in aerial images. The methods The final challenge is detecting objects in large images.
mentioned above do not fully utilize the OBB annotations. Aerial images are usually extremely large (over 20k × 20k
When OBB annotations are available, a rotation R-CNN (RR- pixels). Current GPU memory capacity is insufficient to
CNN) [56] uses rotation region-of-interest (RRoI) pooling process large images. Downsampling a large image to a
to extract rotation-invariant region features. However, RR- small size would lose the detailed information. To solve this
CNNs [56] generate proposals by hand-crafted way. Then problem [14], [16], the large images can be simply split into
the RoI Transformer [18] tries to use the supervision of small patches. After obtaining the results on these patches,
OBBs to learn RoI-wise spatial transformation. The later the results are integrated into large images. To speed up
S2 A-Net [57] extracts spatially invariant features in one- inference on large images, these methods [20]–[22], [67] first
stage detectors. To solve the challenges of scale variations, find regions that are likely to contain instances in the large

Fig. 2. Number of instances per image among DOTA and general object detection datasets. For PASCAL, ImageNet and MS COCO, we count
the statistics of 10,000 random images. As the images in DOTA are very large (20, 000 × 20, 000), for a fair comparison, we count the statistics of
10,1000 image patches with the size of 1024 × 1024, which is also the size used for the baselines in Sec. 5.2. DOTA-v2.0 has a wider range of the
number of instances per image.

images and then detect objects in the regions. In this paper,

we simply follow the naive solutions [14], [16] to build
baselines.

2.4 Code Libraries for Object Detection

The development of object detection algorithms is a so-
phisticated process. In addition, there are too many design
choices and hyperparameter settings, which make compar-
isons between different methods difficult. Therefore, object Fig. 3. Comparisons between HBB and OBB representations for
detection code libraries such as the Tensorflow Object Detec- objects. (a) OBB representation. (b) HBB representation. The HBB
tion API [68], Detectron [29], MaskRCNN-Benchmark [69], representation cannot distinguish rotated dense objects.
Detectron2 [70], MMDetection [28] and SimpleDet [71] are
developed to facilitate the comparisons of object detection 3.2 Category Selection
algorithms. These code libraries primarily use a modular We choose eighteen categories, plane, ship, storage tank, base-
design, which makes it easy to develop new algorithms. ball diamond, tennis court, swimming pool, ground track field,
The current widely used settings, such as the training sched- harbor, bridge, large vehicle, small vehicle, helicopter, roundabout,
ule, are from Detectron [29]. However, these code libraries soccer ball field, basketball court, container crane, airport and he-
mainly focus on horizontal object detection. Only Detec- lipad. We select these categories according to their frequency
tron2 [70] has limited support for oriented object detection. of occurrence and value for real-world applications. The first
In our work, we enriched MMDetection [28] with several 10 categories are common in the existing datasets, e.g., [16],
crucial operators for oriented object detection and evaluated [17], [37], [42]. Other categories are added considering their
10 algorithms for object detection in aerial images. value in real-world applications. For example, we selected
“helicopter” as moving objects are of significant importance
3 C ONSTRUCTION OF DOTA in aerial images, and “roundabout” as it plays an essential
role in roadway analyses. It is worth discussing whether to
3.1 Image Collection take “stuff” categories into account. There are usually no
In aerial images, the resolution and a variety of sensors clear definitions for the ”stuff” categories (e.g.harbor, airport,
are the factors that produce dataset biases [72]. To elimi- parking lot), as shown in the Scene UNderstanding (SUN)
nate these biases, we collect images from various sensors dataset [74]. However, their contextual information may be
and platforms with multiple resolutions, including Google helpful for object detection. Based on this idea, we select
Earth, the Gaofen-2 (GF-2) Satellite, Jilin-1 (JL-1) Satellite, the harbor and airport categories because their borders are
and airborne images (taken by CycloMedia [73] in Rotter- relatively easy to define and there are abundant harbor and
dam). To obtain the DOTA images, we first collected the airport instances in our image sources.
coordinates of areas of interest (e.g., airports or harbors)
from all over the world. Then, according to the coordinates,
images are collected from Google Earth, GF-2 and JL-1 3.3 Oriented Object Annotation
(GF&JL) satellites. The airborne images taken by CycloMe- In computer vision, many visual concepts, such as region
dia [73] were obtained from five perspectives in Rotterdam, descriptions, objects, attributes, and relationships, are often
which include both oblique views and nadir views. The tilt represented with bounding boxes, as shown in [75]. A
angle of the oblique view was approximately 45◦ . common representation of the bounding box is (xc , yc , w, h),
For the Google Earth images, we collect the images that where (xc , yc ) is the center location and w, h are the width
contain instances of interest with sizes from 800 × 800 to and height, respectively, of the bounding box. We call this
4000 × 4000 pixels. However, for the GF&JL satellite and type of bounding box an HBB. The HBB can describe objects
airborne images, we maintained their original sizes. Large well in most cases. However, it cannot accurately outline
images can approach real-world distributions, and also pose oriented instances such as text and objects in aerial images.
a challenge for finding small instances [20]. In DOTA-v2.0, As shown in Fig. 3, the HBB cannot differentiate densely-
the sizes of newly collected GF-2 satellite images and Cy- distributed oriented objects. The conventional NMS algo-
cloMedia airborne images are usually 29, 200 × 27, 620 and rithm fails in such cases. On the other hand, the regional
7, 360 × 4, 912 pixels, respectively. features extracted from HBBs are not rotation invariant.

To address these problems, we represent the objects with TABLE 3

OBBs. In detail, an OBB is denoted by {(xi , yi )|i = 1, 2, 3, 4}, The statistics for the annotated objects across different data sources in
DOTA-v1.5 and DOTA-v2.0. The total image areas, the area of objects
where (xi , yi ) denotes the position of the OBB’s vertex in the and the ratio of foreground pixels of the annotated objects to the image
image. The vertices are arranged in clockwise order. areas are reported.
The most straightforward way to annotate an OBB is to DOTA-v1.5
draw an HBB and then adjust the angle. However, since Google Earth GF&JL Aerial All
there is no reference for HBBs, several adjustments in the # of images 2375 431 / 2806
center, height, width and angle are usually needed to fit Images Area (106 ) 11,873 7,301 / 19,173
Objects Area (106 ) 784 20 / 804
an arbitrarily oriented object well. Clicking on physical Foreground Ratio 0.066 0.003 / 0.042
points lying on the object [76] could make crowd-sourced DOTA-v2.0
annotations more efficient for HBBs, as these points are easy # of images 10186 516 566 11268
to find. Inspired by this idea, we allow the annotators to Images Area (106 ) 29,991 75,854 20,462 126,306
Objects Area (106 ) 1,111 243 673 2,027
click four corners of the OBBs. For most categories, the Foreground Ratio 0.037 0.003 0.033 0.016
corners of the OBBs (e.g., tennis court and basketball court)
lie on or close to the objects (vehicles), however, there are However, He et al. [79] only studied the inaccurate HBB
still some categories whose shapes are very different from annotation. The modeling of inaccurate annotation of the
OBBs. For these categories, we annotate four key points OBBs in DOTA can be researched in the future.
lying on the object. For example, we annotate the planes
with 4 key points, representing the head, two wingtips, and
tail. Then we transfer the 4 key points to an OBB. 4 P ROPERTIES OF DOTA
However, when using OBBs to represent objects, we 4.1 Image Sources
could obtain four different representations for the same The images in DOTA-v2.0 are from three different sources,
object by changing the order of the points. For ex- i.e., Google Earth images, GF-2 and JL-1 (GF&JL) satellite
ample, assume that (x1 , y1 , x2 , y2 , x3 , y3 , x4 , y4 ) represents images, and the CycloMedia [73] airborne images. Tab. 3
an object, but we could represent the same object by shows the statistics of three image sources in terms of the
(x2 , y2 , x3 , y3 , x4 , y4 , x1 , y1 ). For categories having differ- images area, objects area, and foreground ratio. We can
ences between the head and tail (e.g., helicopter, large vehicle, see that the carefully selected Google Earth images contain
small vehicle, harbor), we carefully select the first point to the majority of positive samples. Nevertheless, the nega-
imply the “head” of the object. For other categories (e.g., tive samples are also important to avoid positive sample
soccer-ball field, swimming pool and bridge) that do not have bias [49]. The object distributions in the collected GF&JL
visual clues to determine the first point, we choose the top- satellite images and CycloMedia airborne images are close
left point as the starting point. to those in real-world applications and provide enough
The detailed pipeline is described in the following. First, background area. It is worthwhile to notice that DOTA-
we developed a customized annotation tool. Based on this, v2.0 contains both RGB images and grayscale images. More
we asked experts in the interpretation of aerial images to precisely, the images collected from Google Earth and Cy-
annotate some examples for each category. The annotations cloMedia are often RGB-rendered versions of original aerial
by the experts are used to train the volunteers for the images, and the images from GF-2, JL-1 are 8-bit per pixel
large-scale annotation. After the training, we evaluate the optimally converted from their original panchromatic band
annotation ability of volunteers to separate them into the in 10-bit. However, during those spectral rendering and bit-
plain and senior groups. The volunteers in the plain group length optimization processes, the structure and appearance
are asked to yield the initial annotations which are doubly- information of the image content are always consistent and
checked by senior volunteers and the authors. The images the images are feasible for recognition-oriented tasks [34].
that do not pass the checking were sent back to volunteers to The acquisition dates are available for all the images
improve the annotation quality. The volunteers were mainly from GF-2, JL-1, and CycloMedia, and for 27% of the images
recruited from Wuhan University, with a background in collected from Google Earth. As the main goal of our task
remote sensing image interpretation. Some examples of is to recognize objects in aerial images by relying on visual
annotated patches are shown in Fig. 4. cues, for which the geolocation of an image is insignificant
Discussion. There are two types of possible errors in for the process, DOTA-v2.0 does not provide the geolocation
the object annotations: 1) missing annotations; 2) inaccurate of its images.
bounding boxes annotations. Missing annotations is mainly
caused by the difficulty in identifying tiny objects. The 4.2 Spatial Resolution Information
proportion of missed objects can be ignored and does not The ground sample distance (GSD), which indicates the
influence the training and evaluation of object detectors in distance between pixel centers measured on Earth, has
DOTA-v2.0. However, if the researchers want to study this potential usages. For example, it allows us to calculate the
problem and further improve the performance, we recom- actual sizes of objects, which can be used to filter mislabeled
mend researchers refer to these prior works [77], [78]. The or misclassified outliers since the object sizes of the same
inaccurate bounding boxes annotations exist in all object category are usually limited to a small range. The GSD can
detection datasets since there exists ambiguity to define the also be directly incorporated into object detectors [80] to
boundary of objects sometimes (e.g., the occlusion). Devel- improve the classification accuracy of categories that have
oping algorithms that modeling the inaccurate bounding less physical size variation. Furthermore, we can conduct
boxes annotation has been studied in [79] for natural images. scale normalization [81] based on the priors of the object size

Fig. 4. Examples of annotated images in DOTA. We show three examples per category.

103

102
# Images

101

100
0.0 0.5 1.0 1.5 2.0
Fig. 5. Typical examples of images collected from Google Earth (a), GSD (m/pixel)
GF&JL satellite (b) and CycloMedia (c). Fig. 6. The statistics of the GSD in 30% of the images in DOTA-v2.0.

and GSD. In DOTA-v2.0, the GSDs of the images from GF-2,

to gravity. The unique angle distributions of DOTA make
JL-1, and CycloMedia are 0.81, 0.72, and 0.1 meters per pixel,
it a good dataset for research on rotation-invariant feature
respectively. While the GSDs of the images from Google
extraction and oriented object detection.
Earth range from 0.1m to 4.5m per pixel. The statistical
distribution of GSDs is shown in Fig. 6. It is noted that only
30% of the images in DOTA-v2.0 have the GSD information. 4.4 Various Instances Pixel Sizes
However, the missing of GSDs will not have a big impact on
the applications that require GSDs, since a learning-based Following the convention in [84], we use the height of an
method can be used to estimate the GSD [82]. HBB to measure the pixel size of the instance. We divide
all the instances in our dataset into three splits according
to their heights of HBBs: small, with range from 10 to
4.3 Various Instance Orientations 50, medium, with range from 50 to 300, and large, with
Objects in the overhead view images have a high diversity range above 300. Tab. 4 illustrates the percentages of these
of orientations without the restriction of gravity. As shown three instance splits in different datasets. It is clear that the
in Fig. 1 (g), the objects have equal probabilities of arbitrary PASCAL VOC dataset, NWPU VHR-10 dataset and DLR 3K
angles in [−π, π]. It is worthwhile to note that although Munich Vehicle dataset are dominated by medium instances
objects in scene text detection [83] and face detection [60] or small instances.
also have many orientation variations, the angles of most MS COCO and DOTA-v1.0 have a good balance between
objects lie within a narrow range (e.g., [−π/2, π/2]) due small instances and medium instances. DOTA-v2.0 has more

1
TABLE 4 0.9
Comparison of the instance size distributions of aerial and natural 0.8
images in some datasets. 0.7
0.6
Dataset 10-50 pixels 50-300 pixels >300 pixels 0.5
PASCAL VOC [11] 0.14 0.61 0.25 0.4
0.3
MS COCO [13] 0.43 0.49 0.08 0.2
NWPU VHR-10 [42] 0.15 0.83 0.02 0.1
DLR 3K [16] 0.93 0.07 0 0

DOTA-v1.0 [14] 0.57 0.41 0.02

DOTA-v1.5 0.79 0.2 0.01
DOTA-v2.0 0.77 0.22 0.01
dense[0, 10] normal[10, 50] sparse[50, ∞]

Fig. 9. Densities of the different categories. The density is measured by

calculating the distance to the closest instance.

image in DOTA varies more widely than in natural image

datasets.
Different categories also have different density distribu-
tions. To give a quantitative analysis, for each instance, we
first measure the distance to the closest instance in the same
category. We then bin the distances into three parts, dense
[0, 10), normal [10, 50) and sparse [50, ∞) (see Fig. 9). Fig. 9
Fig. 7. Size variations for each category in DOTA. The sizes of different shows that the storage tank, ship and small vehicle are top-3
categories vary in different ranges. dense categories.
105
105
#Instances

#Instances

104 4.7 DOTA Versions

104
103 It is important to notice the significant improvements from
0 2 4 6 0 2 4 DOTA-v1.0 to DOTA-v2.0. In DOTA-v1.0, tiny objects (be-
AR of oriented bounding box AR of horizontal bounding box low 10 pixels) have not been annotated, and images are
Fig. 8. AR distributions of the instances in DOTA. (a) The ARs of the mainly from a single domain, i.e., Google Earth images.
OBBs. (b) The ARs of the HBBs.
Moreover, the images from DOTA-v1.0 are usually selected
areas that contain many objects from large-size images. Al-
small instances than DOTA-v1.0. In DOTA-v2.0, some in-
though, in the past years, promising progress has been
stances that are approximately 10 pixels are annotated.
reported in oriented object detection with DOTA-v1.0, fol-
In Fig. 7, we also show the distribution of instances’ pixel
lowing challenging aspects can not been fully addressed by
sizes for different categories in DOTA. This figure indicates
using DOTA-v1.0:
that the scales vary greatly both within and between cate-
gories. These large-scale variations among instances make
- to benchmark detection models for oriented objects
the detection task more challenging.
both in tiny size and normal size;
- to address the object detection problem in large-scale
4.5 Various Instance Aspect Ratios (ARs) images, e.g., images with size larger than 20, 000 ×
The AR is essential for anchor-based models, such as Faster 20, 000 pixels, that only contain a few objects;
R-CNN [54] and You Only Look Once (YOLOv2) [51]. We - to develop robust oriented object detection mod-
use two kinds of ARs for all the instances in our dataset to els with strong generalization capability for multi-
guide the model design namely, 1) the ARs of the original source overhead images.
OBBs and 2) the AR of HBBs, which are generated by
calculating the axis-aligned bounding boxes over the OBBs. To address these problems, DOTA-v1.5 added the annota-
Fig. 8 illustrates the distributions of these two types of tion of the tiny objects in DOTA-v1.0. DOTA-v2.0 further
aspect ratios in DOTA. We can see that instances vary collected many more large-size GF-2 and airborne images,
significantly in aspect ratio. Moreover, many instances have which have a lower foreground ratio, approaching the object
a large aspect ratio in our dataset. distribution in real-world applications, as shown in Tab. 3.
The number of objects for each category and the dataset split
for three versions of DOTA are summarized in Tab. 5.
4.6 Various Instance Densities of the Images
The number of instances per image is an important property
4.7.1 DOTA-v1.0
for object detection datasets and varies largely in DOTA. It
can be very dense (up to 1000 instances per image patch), DOTA-v1.0 contains 15 common categories, 2,806 images
or very sparse (only one instance per image patch). We and 188, 282 instances. The proportions of the training set,
compare this property among DOTA and the general object validation set, and testing set in DOTA-v1.0 are 1/2, 1/6,
detection datasets in Fig. 1. The number of instances per and 1/3, respectively.

TABLE 5 (HBB and OBB) in our paper. The HBB is a rectangle

Comparisons of the three versions of DOTA. We count the number of (x, y, w, h), and the OBB is a quadrilateral {(xi , yi )|i =
instances for each category and dataset split.
1, 2, 3, 4}. Then, there are two tasks, detection with HBB
DOTA-v1.0 DOTA-v1.5 DOTA-v2.0 and detection with OBB. To be more specific, we evaluate
Plane 14,085 14,978 23,930
BD 1,130 1,127 3,834 these methods on two kinds of ground truths: HBB and OBB
Bridge 3,760 3,804 21,433 ground truths. For the two tasks, each detected bounding
GTF 678 689 4,933 box has a corresponding confidence score. We adopt the
SV 48,891 242,276 1,235,658
LV 31,613 39,249 89,353
PASCAL VOC 07 metric [11] for the calculation of the mean
Ship 52,516 62,258 251,883 average precision (mAP). Average Precision (AP) computes
TC 4,654 4,716 9,396 the average precision value for recall value over 0 to 1 (i.e.,
BC 954 988 3,556 the area under precision/recall curve). The mean Average
ST 11,794 12,249 79,497
SBF 720 727 2,404 Precision (mAP) is the average of AP over all classes. The
RA 871 929 6,809 detailed computation of the precision and recall can refer
Harbor 12,287 12,377 29,581 to [11]. The intersection over union (IoU) is crucial in
SP 3,507 4,652 20,095
determining true positives and false positives, which are
HC 822 833 893
CC 0 237 3,887 required to compute precision and recall. It is worthwhile
Airport 0 0 5,905 to note that for the OBB task, the intersection over union
Helipad 0 0 611 (IoU) is calculated between OBBs, as shown in Fig. 10. The
Total 188,282 402,089 1,793,658
Training 98,990 210,631 268,627
two T OBBs (Bp and Bgt ), and the intersection between OBBs
Validation 28,853 69,565 81,048 (Bp Bgt ) are all convex polygons, whose area can be easily
6
Test/Test-dev 60,439 121,893 353,346 computed S . The union area of two T OBBs can be calculated
Test-challenge 0 0 1,090,637 as |Bp Bgt | = |Bp | + |Bgt | − |Bp Bgt |. The code for the
mAP and IoU computation between OBBs can be found in
our development kit.

Bp Bgt 5.2 Implementation Details

In the previous benchmarks [14], the algorithms were im-
plemented with different codes and settings, which makes
Fig. 10. The computation of the IoU between two OBBs. these algorithms hard to compare in DOTA. To this end, we
implement and evaluate all the algorithms in one unified
4.7.2 DOTA-v1.5
code library modified from MMDetection [28].
DOTA-v1.5 uses the same images as DOTA-v1.0, but ex- Since large images cannot be directly fed to CNN-based
tremely small instances (less than 10 pixels) are also an- detectors due to the memory limitations, we crop a series of
notated. Moreover, a new category, “container crane” con- 1, 024×1, 024 patches from the original images with a stride
taining 402,089 instances in total is added. The number of set to 824 (different from the previous stride of 512 [14]).
images and dataset splits are the same as those in DOTA- During inference, we first send the patches (same settings
v1.0. as training) to obtain temporary results. Then we map the
detected results from the patch coordinates to the original
4.7.3 DOTA-v2.0 image coordinates. Finally, we apply NMS on these results
There are 18 common categories, 11,268 images and in the original image coordinates. We set the threshold of
1,793,658 instances in DOTA-v2.0. Compared to DOTA- NMS to 0.3 for the HBB experiments and 0.1 for the OBB
v1.5, it further adds the new categories of “airport” and experiments. For multi-scale training and testing, we first
“helipad”. DOTA-v2.0 are split into training, validation, test- scale the original images to [0.5, 1.0, 1.5] and then crop the
dev, and test-challenge subsets. To avoid the problem of over- images into patches of size 1, 024×1, 024 and a stride of 824.
fitting, the proportion of the training and validation sets We use 4 GPUs for training with a total batch size of 8 (2
is smaller than that of the test set. Furthermore, we have images per GPU). The learning rate is set to 0.01. Except for
two test subsets, namely test-dev and test-challenge, which RetinaNet [85], which adopts the ”2×” schedule, the other
are similar to the MS COCO dataset [13]. algorithms adopt the ”1×” [29] training schedule. We set the
For the test-dev subset and the test-challenge subset, we number of proposals and maximum number of predictions
release the images without annotations. For evaluation, one per image patch to 2,000 for all the experiments except when
can submit the results to the evaluation server5 . All the otherwise mentioned. The other hyperparameters follow
DOTA-v2.0 experiments in this paper are evaluated on test- those of Detecron [29].
dev.
5.2.1 Baselines with HBBs
5 B ENCHMARKS We use two ways to build baselines for the HBB task.
5.1 Evaluation Tasks and Metrics The first way directly predicts the HBB results, while the
second way first predicts the OBB results and then converts
The task of object detection is to locate and classify the
instances in images. We use two location representations 6. A polygon can be decomposed into a group of non-overlapping
triangles. The area of the convex polygon is equal to the sum of all the
5. https://captain-whu.github.io/DOTA/evaluation.html triangular areas.

OBBs to HBBs. To directly predict the HBB results, we large number of instances per image, as illustrated in Fig. 2,
use RetinaNet [85], Mask R-CNN, Cascade Mask R-CNN, we apply a mask head to all the HBBs after NMS. In this
Hybrid Task Cascade and Faster R-CNN [54] as baselines. way, we evaluate Mask R-CNN [86], Cascade Mask R-CNN
For the OBB predictions, we will introduce the methods in and Hybrid Task Cascade [88].
the following section.
5.3 Codebase and Development Kit
5.2.2 Baselines with OBBs
Most of the state-of-the-art object detection methods are not We also provide an aerial object detection code library7
designed for oriented objects. To enable these methods to and a development kit8 for using DOTA. To construct the
predict OBBs, we build the baselines in two ways. The first comprehensive baselines, we select MMDetection as the
is to change HBB head to OBB Head, which regresses the fundamental code library since it contains rich object de-
offsets of OBBs relative to the HBBs. The second is Mask tection algorithms and has the feature of modular design.
Head, which considers the OBBs to a coarse mask and However, the original MMDetection [28] lacks the modules
predicts the pixel-level classification from each RoI. to support oriented object detection. Therefore, we enriched
OBB Head To predict OBB, the previous Faster R- MMDetection with OBB Head as described in Sec. 5.2.2
CNN OBB [14] and Textboxes++ [62] modified RoI Head to enable OBB predictions. We also implemented modules
of Faster R-CNN and the Anchor Head of the single- such as rotated RoI Align and rotated position-sensitive
shot detector (SSD), respectively, to regress quadrangles. RoI Align for rotated region feature extraction, which are
In this paper, we use the representation (x, y, w, h, θ) in- crucial components in algorithms such as rotated region
stead of {(xi , yi )|i = 1, 2, 3, 4} for OBB regression. More proposal network (RRPN) [89] and RoI Transformer [18].
precisely, rectangular RoIs (anchors) can be written as B = These new modules are compatible with the modularly
(xmin , ymin , xmax , ymax ). We can also consider it a special designed MMDetection, so we can easily create new al-
OBB and rewrite it as R = (x, y,w, h, θ). For matching, gorithms for oriented object detection not restricted to the
IoUs are calculated between the horizontal RoIs (anchors) baseline methods in this paper. We also provide a develop-
and HBBs of the ground truths for computational simplicity. ment kit containing necessary functions for object detection
Each OBB, it has four forms: G = {gti |i = 1, 2, 3, 4}, where in DOTA, including:
gt1 = (xg , yg , wg , hg , θg ), gt2 = (xg , yg , wg , hg , θg + π), - Loading and visualizing the ground truths.
gt3 = (xg , yg , hg , wg , θg ), and gt4 = (xg , yg , hg , wg , θg + π). - Calculating the IoU between OBBs, which is imple-
Before calculating the targets, we choose the best matched mented in a mixture of Python/C program. We pro-
ground-truth form. The index of the best matched form is vide both the CPU and GPU versions.
calculated by ξ = arg min D(R, gti ), where D is a distance - Evaluating the results. The evaluation metrics are de-
i
function, which could be Euclidean distance or another scribed in Sec. 5.1.
distance function. We denote the best matched form by - Cropping and merging images. When using the large-
gtξ = (xb , yb , wb , hb , θb ). Then the learning target is calcu- size images in DOTA, one can utilize this tool kit to
lated as split an original image into patches. After testing on
the patches, one can use the tools to map the results
tx = (xb − x)/w, ty = (yb − y)/h,
of patches back to the original image coordinates and
tw = log(wb /w), th = log(hb /h), (1) apply NMS.
tθ = θb − θ
We then simply replace the HBB RoI Head of Faster R- 6 R ESULTS
CNN and anchor head of RetinaNet with OBB Head and 6.1 Benchmark Results and Analyses
obtain two models, called Faster R-CNN OBB and RetinaNet
In this section, we conduct a comprehensive evaluation
OBB. We also modify the Faster R-CNN to predict both
of over 70 experiments and analyze the results. First, we
the HBB and OBB in parallel, which is similar to Mask
demonstrate the baseline results of 10 algorithms on DOTA-
R-CNN [86]. We call this model Faster R-CNN H-OBB. We
v.10, DOTA-v1.5 and DOTA-v2.0. The baselines cover both
further evaluate the deformable RoI pooling (Dpool) and
two-stage and one-stage algorithms. For most algorithms,
RoI Transformer by replacing the RoI Align in Faster R-CNN
we report the mAPs of HBB and OBB predictions, respec-
OBB. Then we have two models: Faster R-CNN OBB + Dpool
tively, except for RetinaNet and Faster R-CNN, since they
and Faster R-CNN OBB + RoI Transformer. Note that the RoI
do not support oriented object detection. For algorithms
Transformer used here is slightly different from the original
with only OBB heads (RetinaNet OBB, Faster R-CNN OBB,
one. The original RoI Transformer uses the Light Head R-
Faster R-CNN OBB +DPool, Faster R-CNN OBB +RoI Trans-
CNN [87] as the base detector while we use Faster R-CNN.
former), we obtain their HBB results by transferring from
Mask Head Mask R-CNN [86] was originally used for in-
OBB as described in Sec. 5.2.1. For algorithms with both
stance segmentation. Although DOTA does not have pixel-
HBB and OBB heads (Mask R-CNN, Cascade Mask R-CNN,
level annotation for each instance, the OBB annotations
Hybrid Task Cascade*, and Faster R-CNN H-OBB), the HBB
can be considered coarse pixel-level annotations, so we can
mAP is the maximum of the predicted HBB mAP and the
apply Mask R-CNN [86] to DOTA. During inference, we
transferred HBB mAP. It can be seen that the OBB mAP
calculate the minimum OBBs that contain the predicted
masks. The original Mask R-CNN [86] only applies a mask 7. https://github.com/dingjiansw101/AerialDetection
head to the top 100 HBBs in terms of the score. Due to the 8. https://github.com/CAPTAIN-WHU/DOTA devkit

TABLE 6
Baseline results on DOTA. For the evaluation of DOTA-v2.0, we use the DOTA-v2.0 test-dev set. The implementation details are described in
Sec. 5.2. All the algorithms in this table adopt the ResNet-50 with an FPN as backbone. The speed refers to the inference speed, which is reported
for a single NVIDIA Tesla V100 GPU on DOTA-v2.0 test-dev. The image size is 1, 024 × 1, 024. Hybrid Task Cascade* means that the semantic
branch is not used since there are no semantic annotations in DOTA.
DOTA-v1.0 DOTA-v1.5 DOTA-v2.0
method speed (fps)
HBB mAP OBB mAP HBB mAP OBB mAP HBB mAP OBB mAP
RetinaNet [85] 16.7 67.45 - 61.64 - 49.31 -
RetinaNet OBB [85] 12.1 69.05 66.28 62.49 59.16 49.26 46.68
Mask R-CNN [86] 9.7 71.61 70.71 64.54 62.67 51.16 49.47
Cascade Mask R-CNN [88] 7.2 71.36 70.96 64.31 63.41 50.98 50.04
Hybrid Task Cascade* [88] 7.9 72.49 71.21 64.47 63.40 50.88 50.34
Faster R-CNN [54] 14.3 70.76 - 64.16 - 50.71 -
Faster R-CNN OBB [14] 14.1 71.91 69.36 63.85 62.00 49.37 47.31
Faster R-CNN OBB + Dpool [55] 12.1 71.83 70.14 63.67 62.20 50.48 48.77
Faster R-CNN H-OBB [14] 13.7 70.37 70.11 64.43 62.57 50.38 48.90
Faster R-CNN OBB + RoI Transformer [18] 12.4 74.59 73.76 66.09 65.03 53.37 52.81

TABLE 7
Baseline results of class-wise AP on DOTA-v1.0. The abbreviations of algorithms are: Mask R-CNN (MR), CMR-Cascade Mask R-CNN (CMR),
Hybrid Task Cascade without a semantic branch (HTC*), Faster R-CNN (FR), Deformable RoI Pooling (Dp) and RoI Transformer (RT). The short
names for categories are defined as: BD–Baseball diamond, GTF–Ground field track, SV–Small vehicle, LV–Large vehicle, TC–Tennis court,
BC–Basketball court, SC–Storage tank, SBF–Soccer-ball field, RA–Roundabout, SP–Swimming pool, HC–Helicopter.
OBB Results
method Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC mAP
RetinaNet [85] 86.54 77.45 42.8 64.87 71.06 58.5 73.53 90.72 80.97 66.67 52.42 62.16 60.79 64.84 40.83 66.28
MR [86] 88.7 74.13 50.75 63.66 73.64 73.98 83.68 89.74 78.92 80.26 47.43 65.09 64.79 66.09 59.79 70.71
CMR [88] 88.93 75.21 51.55 64.9 74.39 75.37 84.74 90.23 77.48 81.51 46.57 63.49 65.39 67.63 56.96 70.96
HTC* [88] 89.17 75.05 51.95 64.5 74.19 76.3 86.05 90.55 79.51 77.18 50.3 61.23 65.89 68.29 58.01 71.21
FR OBB [14] 88.42 74.24 45.31 61.49 73.53 70.03 77.76 90.87 81.8 82.64 48.75 60.14 63.42 67.65 54.31 69.36
FR OBB + Dp [55] 88.82 74.12 45.44 63.07 73.13 73.59 84.39 90.71 82.28 83.59 42.76 58.49 63.52 68.25 59.89 70.14
FR H-OBB [14] 88.41 79.35 45.39 63.16 73.91 72.11 83.86 90.25 77.34 81.04 48.5 60.53 63.88 66.43 57.53 70.11
FR OBB + RT [18] 88.34 77.07 51.63 69.62 77.45 77.15 87.11 90.75 84.9 83.14 52.95 63.75 74.45 68.82 59.24 73.76
HBB Results
method Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC mAP
RetinaNet [85] 88.28 77.76 47.47 59.07 73.83 63.49 77.69 90.43 78.57 65.87 48.67 61.82 68.92 71.59 38.22 67.45
RetinaNet OBB [85] 88.51 78.21 47.9 64.46 75.17 72.9 78.5 90.72 82.22 67.11 51.33 62.16 69.47 69.79 37.3 69.05
MR [86] 88.79 79.06 53.04 63.14 78.22 65.23 77.94 89.6 81.99 81.06 47.17 65.19 72.72 68.67 62.27 71.61
CMR [88] 88.93 81.5 52.85 64.01 79.18 65.85 78.12 90.08 77.48 81.87 46.45 62.88 72.99 69.95 58.32 71.36
HTC* [88] 89.11 79.76 53.87 64.4 79.06 76.23 86.57 90.5 79.51 81.66 50.44 63.75 73.33 70.61 48.59 72.49
FR [54] 89.02 75.8 53.47 60.8 78.02 65.56 78.01 90.1 77.3 81.7 47.12 61.48 72.59 70.89 59.54 70.76
FR OBB [14] 88.37 75.45 52.11 59.98 78.08 73.42 85.65 90.81 83.22 83.18 51.45 60.01 72.44 71.09 53.44 71.91
FR OBB + Dp [55] 88.86 77.4 51.6 63.12 77.62 74.73 86.2 90.72 82.73 83.75 44.35 58.81 71.97 70.98 54.68 71.83
FR H-OBB [14] 88.67 78.95 52.63 57.34 78.55 65.22 78.08 90.69 82.01 82.06 46.08 61.46 72.43 70.34 51.12 70.37
FR OBB + RT [18] 88.47 81.0 54.1 69.19 78.42 81.16 87.35 90.75 84.9 83.55 52.63 62.97 75.89 71.31 57.22 74.59
TABLE 8
Baseline results of class-wise AP on DOTA-v1.5. The abbreviations of algorithms are: Mask R-CNN (MR), CMR-Cascade Mask R-CNN (CMR),
Hybrid Task Cascade without a semantic branch (HTC*), Faster R-CNN (FR), Deformable RoI Pooling (Dp) and RoI Transformer (RT). The short
names for categories are defined as: BD–Baseball diamond, GTF–Ground field track, SV–Small vehicle, LV–Large vehicle, TC–Tennis court,
BC–Basketball court, SC–Storage tank, SBF–Soccer-ball field, RA–Roundabout, SP–Swimming pool, HC–Helicopter, CC-Container Crane.
OBB Results
method Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC CC mAP
RetinaNet [85] 71.43 77.64 42.12 64.65 44.53 56.79 73.31 90.84 76.02 59.96 46.95 69.24 59.65 64.52 48.06 0.83 59.16
MR [86] 76.84 73.51 49.9 57.8 51.31 71.34 79.75 90.46 74.21 66.07 46.21 70.61 63.07 64.46 57.81 9.42 62.67
CMR [88] 77.77 74.62 51.09 63.44 51.64 72.9 79.99 90.35 74.9 67.58 49.54 72.85 64.19 64.88 55.87 3.02 63.41
HTC* [88] 77.8 73.67 51.4 63.99 51.54 73.31 80.31 90.48 75.12 67.34 48.51 70.63 64.84 64.48 55.87 5.15 63.4
FR OBB [14] 71.89 74.47 44.45 59.87 51.28 68.98 79.37 90.78 77.38 67.5 47.75 69.72 61.22 65.28 60.47 1.54 62.0
FR OBB + Dp [55] 71.79 73.61 44.76 61.99 51.34 70.04 79.67 90.78 76.58 67.73 44.58 70.51 61.8 65.49 64.35 0.15 62.2
FR H-OBB [14] 71.57 74.71 46.39 63.4 51.54 70.11 79.09 90.63 76.81 67.4 48.66 70.9 63.1 65.67 56.66 4.55 62.57
FR OBB + RT [18] 71.92 76.07 51.87 69.24 52.05 75.18 80.72 90.53 78.58 68.26 49.18 71.74 67.51 65.53 62.16 9.99 65.03
HBB Results
method Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC CC mAP
RetinaNet [85] 74.05 77.75 48.75 59.94 49.23 61.43 77.31 90.38 75.46 61.19 47.29 69.99 67.99 74.15 40.88 10.4 61.64
RetinaNet OBB [85] 71.66 77.22 48.71 65.16 49.48 69.64 79.21 90.84 77.21 61.03 47.3 68.69 67.22 74.48 46.16 5.78 62.49
MR [86] 78.36 77.41 53.36 56.94 52.17 63.6 79.74 90.31 74.28 66.41 45.49 71.32 70.77 73.87 61.49 17.11 64.54
CMR [88] 78.61 75.43 54.0 63.76 52.55 63.93 79.88 90.06 75.05 67.83 45.76 72.48 72.1 74.11 53.9 9.59 64.31
HTC* [88] 78.41 74.41 53.41 63.17 52.45 63.56 79.89 90.34 75.17 67.64 48.44 69.94 72.13 74.02 56.42 12.14 64.47
FR [54] 71.88 74.06 52.69 62.35 52.08 63.22 79.69 90.55 76.91 66.86 47.84 70.72 70.61 68.55 63.34 15.26 64.16
FR OBB [14] 71.91 71.6 50.58 61.95 51.99 71.05 80.16 90.78 77.16 67.66 47.93 69.35 69.51 74.4 60.33 5.17 63.85
FR OBB + Dp [55] 71.9 72.8 50.84 61.99 51.97 72.15 80.13 90.74 76.53 67.95 45.09 69.93 70.43 74.72 58.28 3.19 63.67
FR H-OBB [14] 71.6 74.14 52.69 62.85 52.13 71.32 79.75 90.64 76.29 67.52 49.31 71.27 72.11 73.91 55.11 10.26 64.43
FR OBB + RT [18] 71.92 75.21 54.09 68.1 52.54 74.87 80.79 90.46 78.58 68.41 51.57 71.48 74.91 74.84 56.66 13.01 66.09

TABLE 9
Baseline results of class-wise AP on DOTA-v2.0. The abbreviations of algorithms are: Mask R-CNN (MR), CMR-Cascade Mask R-CNN (CMR),
Hybrid Task Cascade without a semantic branch (HTC*), Faster R-CNN (FR), Deformable RoI Pooling (Dp) and RoI Transformer (RT). The short
names for categories are defined as: BD–Baseball diamond, GTF–Ground field track, SV–Small vehicle, LV–Large vehicle, TC–Tennis court,
BC–Basketball court, SC–Storage tank, SBF–Soccer-ball field, RA–Roundabout, SP–Swimming pool, HC–Helicopter, CC–Container Crane,
Air–Airport, Heli–Helipad.
OBB Results
method Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC CC Air Heli mAP
RetinaNet [85] 70.63 47.26 39.12 55.02 38.1 40.52 47.16 77.74 56.86 52.12 37.22 51.75 44.15 53.19 51.06 6.58 64.28 7.45 46.68
MR [86] 76.2 49.91 41.61 60.0 41.08 50.77 56.24 78.01 55.85 57.48 36.62 51.67 47.39 55.79 59.06 3.64 60.26 8.95 49.47
CMR [88] 77.01 47.54 41.79 58.02 41.58 51.74 57.86 78.2 56.75 58.5 37.89 51.23 49.38 55.98 54.59 12.31 67.33 3.01 50.04
HTC* [88] 77.69 47.25 41.15 60.71 41.77 52.79 58.87 78.74 55.22 58.49 38.57 52.48 49.58 56.18 54.09 4.2 66.38 11.92 50.34
FR OBB [14] 71.61 47.2 39.28 58.7 35.55 48.88 51.51 78.97 58.36 58.55 36.11 51.73 43.57 55.33 57.07 3.51 52.94 2.79 47.31
FR OBB + Dp [55] 71.55 49.74 40.34 60.4 40.74 50.67 56.58 79.03 58.22 58.24 34.73 51.95 44.33 55.1 53.14 7.21 59.53 6.38 48.77
FR H-OBB [14] 71.39 47.59 39.82 59.01 41.51 49.88 57.17 78.36 56.87 58.24 37.66 51.86 44.61 55.49 54.74 7.56 61.88 6.6 48.9
FR OBB + RT [18] 71.81 48.39 45.88 64.02 42.09 54.39 59.92 82.7 63.29 58.71 41.04 52.82 53.32 56.18 57.94 25.71 63.72 8.7 52.81
HBB Results
method Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC CC Air Heli mAP
RetinaNet [85] 71.86 48.69 42.2 53.12 41.16 45.64 55.9 77.74 56.14 52.0 37.68 51.46 53.27 57.51 46.76 15.66 67.76 12.97 49.31
RetinaNet OBB [85] 70.99 46.77 43.76 55.08 41.55 51.06 58.01 77.78 57.72 53.5 37.66 51.72 53.56 57.32 49.18 12.01 64.53 4.48 49.26
MR [86] 77.61 51.35 44.89 60.12 42.51 48.1 57.93 77.84 57.55 57.88 36.53 51.71 54.79 58.93 60.01 14.42 60.32 8.43 51.16
CMR [88] 78.12 47.89 46.43 57.8 42.97 48.23 59.11 78.19 57.17 58.88 37.42 51.32 53.66 58.07 55.5 17.67 67.37 1.81 50.98
HTC* [88] 78.28 47.95 45.7 59.95 42.97 48.7 59.14 78.58 55.91 58.77 37.75 52.46 53.34 58.64 55.53 9.78 66.67 5.74 50.88
FR [54] 76.14 49.93 44.97 57.8 42.4 47.86 57.76 77.7 56.57 58.65 39.24 52.6 54.94 58.92 56.62 12.88 61.64 6.24 50.71
FR OBB [14] 71.68 45.8 45.56 58.7 42.18 51.28 59.28 79.01 58.74 58.75 37.26 51.93 52.36 58.08 54.12 8.48 53.01 2.4 49.37
FR OBB + Dp [55] 71.58 47.68 46.16 60.48 42.34 52.55 59.48 79.07 59.61 58.46 35.35 53.73 53.12 58.33 52.06 12.56 59.76 6.38 50.48
FR H-OBB [14] 77.14 50.54 45.6 57.53 42.27 48.09 57.6 78.4 59.78 57.8 36.64 52.13 52.51 58.42 48.91 14.99 60.0 8.47 50.38
FR OBB + RT [18] 71.84 48.2 47.84 63.94 42.97 54.79 60.74 82.88 63.51 58.89 40.63 52.83 55.7 58.87 57.94 27.04 64.27 7.68 53.37

��

�� DOTA-v2.0 test-dev set as an example, Mask R-CNN still
��
outperforms Faster R-CNN H-OBB by 0.57 points in OBB
��
��
mAP. Nevertheless, Mask R-CNN is slower than Faster R-
��

��

�� CNN H-OBB by 4 fps. Note that the process of transferring

�� the mask to the OBB is not considered here. Otherwise,
Mask R-CNN should be slower.
��
��
�� 6.1.2 RoI Transformer vs. Deformable RoI Pooling
� � ��
Geometric variations are still challenging in object detection.
Fig. 11. Results of using different backbones. The algorithms are In this part, we evaluate RoI Transformer and Dpool by
tested on DOTA-2.0 test-dev. For each algorithm, we choose 3 different replacing RoI Align in Faster R-CNN OBB. We call these two
backbones, i.e., ResNet-50 with an FPN, ResNet-101 with an FPN, and
64×4d ResNeXt-101 with an FPN. Faster R-CNN-O is the Faster R- models Faster R-CNN OBB + RoI Transformer and Faster R-
CNN OBB in this work. RetinaNet-O stands for RetinaNet OBB. Dpool CNN OBB + Dpool. Tab. 6 and Fig. 11 show that Dpool
means the Deformable RoI Pooling, and RT means the RoI Transformer. improves the performance of Faster R-CNN OBB at most
The speeds are tested on a single Tesla V100.
times, while RoI Transformer performs better than Dpool.
is usually slightly lower than the HBB mAP for the same This finding verifies that carefully designed geometry trans-
algorithm since the OBB task needs a more precise location formation modules such as RoI Transformer are better than
than the HBB task. general geometry transformation modules such as Dpool for
Tab. 6 shows that the performance on DOTA-v1.0, aerial images.
DOTA-v1.5 and DOTA-v2.0 are declining, indicating the
increased difficulty of the datasets. The class-wise AP are 6.1.3 Excluding Small Instances
reported in Tab. 7, Tab. 8 and Tab. 9. To give more de- During the training on DOTA-v1.5 and DOTA-v2.0, many
tailed comparisons of speed vs. accuracy, we evaluate all extremely small instances will cause numerical instability.
algorithms using different backbones (as shown in Fig. 11). For the experiments in DOTA-v1.5 and DOTA-v2.0, we
From the speed-accuracy curve, the Faster R-CNN OBB + set a threshold to exclude too small instances. We try to
RoI Transformer outperforms the other methods. To explore explore the influence of different thresholds on DOTA-v2.0.
the properties of DOTA and provide guidelines for future We filter the small instances by two rules: 1) the area of
research, we evaluate the module design and the hyper- instance is below a certain threshold, and 2) max(w, h) is
parameter setting. Then, we analyze the influence of data below a threshold, where the w and h are the width and
augmentation in detail. Finally, we visualize the results to height, respectively, of the corresponding HBB. The results
show the difficulties of ODAI. in Tab. 10 show that small instances have little influence on
the results.
6.1.1 Mask Head vs. OBB Head
The OBB head tackles oriented object detection as a regres- 6.1.4 Number of Proposals
sion problem, while the mask head tackles oriented object The number of proposals is an important hyperparameter
detection as a pixel-level classification problem. The mask in modern detectors. As mentioned before, the possible
head more easily converges and achieves better results but number of instances in aerial images is quite different from
is more computationally expensive. Taking the results on the that in natural images. In DOTA, one 1, 024 × 1, 024 image

TABLE 10 6.1.6 Class-Wise Results

Results after excluding extremely small instances by different
thresholds in DOTA-v2.0. There are 642,601 instances before filtering.
The baseline results of class-wise AP on DOTA-v1.0, DOTA-
v1.5, and DOTA-v2.0 are reported in Tab. 7, Tab. 8 and
# of filtered Instance Filtering strategy HBB mAP
99,317 area ≤ 50 and max(w, h) ≤ 10 51.08 Tab. 9. In contrast with DOTA-v1.0, DOTA-v1.5 additionally
157,287 area ≤ 80 and max(w, h) ≤ 10 51.35 annotated the tiny objects (most of them are small vehicles
158,629 area ≤ 80 and max(w, h) ≤ 12 50.71 below 10 pixels). Therefore, by comparing the AP of small
vehicles of the same detector on DOTA-v1.0 and DOTA-
may contain more than 1,000 instances. There is no doubt v1.5, we can see the challenges in the detection of tiny
that the parameters that perform well for natural images are objects. Taking Faster R-CNN OBB with RoI Transformer
not optimal for aerial images. Here we explore the optimal as an example, the AP on small vehicles decrease by 25.4
settings for aerial images. As shown in Tab. 11, the number points from 77.45 to 52.05. The challenge on tiny object
of proposals with the highest performance for Faster R-CNN detection can also be checked in the last row of Fig. 12.
OBB + RoI Transformer is 8,000. For Faster R-CNN OBB, the The advantage of OBBs over HBBs in detecting the densely
increase in the mAP slows at approximately 8,000 proposals. packed objects can be demonstrated by comparing the mAP
Furthermore, from 1,000 to 10,000 proposals, the improve- of OBB detectors and HBB detectors. For example, Faster
ments in Faster R-CNN + RoI Transformer and Faster R- R-CNN OBB outperforms Faster R-CNN by 8 points in
CNN OBB are 2.2 and 1.39 points in mAP, respectively. AP on large vehicles in DOTA-v1.0. Some examples of the
However, the increased number of proposals bring more detection of densely packed objects are shown in the first
computation. Therefore, for the other experiments in this row of Fig. 12.
paper, we choose 2,000 proposals. The optimal number of
proposals in DOTA is quite a bit larger than that in PASCAL 6.1.7 Visualization of the Results
VOC, where 300 is the optimal number. This finding con- We show the performance of Faster R-CNN [54], Faster R-
firms that the difference between aerial and natural images CNN OBB, RetinaNet OBB, Mask R-CNN and Faster R-
is again significant. CNN OBB + RoI Transformer on difficult scenes in Fig. 12:
1) The first row demonstrates densely packed large vehicles.
6.1.5 Data Augmentation
Faster R-CNN misses many instances due to the high over-
In this section, we explore the influence of data augmenta- laps between neighboring large vehicles in the HBBs. Those
tion in detail. We followed the multi-scale training, testing, large vehicles are suppressed through NMS. Faster R-CNN
and rotation training strategies in [25] and further con- OBB, Mask R-CNN and Faster R-CNN OBB + RT perform
duct rotation testing. Note that as data augmentation often well, while RetinaNet OBB has lower location precision due
produces a huge number of patches and will dramatically to feature misalignment. 2) The second and third rows show
increase the time complexity of experiments, we conduct long shape instances with a large ARs. These instances are
our ablation study of data augmentation on DOTA-v1.5, self-similar, which means that each part of the instance
which is similar to DOTA-v2.0 in data distribution while has a similar feature as the whole instance. For example,
much smaller. The model we select is Faster R-CNN OBB the second row shows that all methods have at least two
+ RoI Transformer. We choose R-50-FPN as the backbone predictions on a single bridge. The third row also reveals
and adopt five data augmentation strategies. The first is the this problem. There exist several predictions on a single
high patch overlap. We change the overlap between patches ship. 3) The second and third rows also indicate that several
from 200 to 512 since the large instances may be cut off at the different categories have very similar features. Bridges are
edge. The second and third are multi-scale training and test- easily classified as airports and harbors while the ships are
ing, respectively. We resize the original images by factors of easily to be classified as harbors and bridges. 4) The last row
[0.5, 1.0, 1.5] and then crop the original images into patches shows the difficulty of detecting extremely small instances
of size 1, 024 × 1, 024. The fourth is the rotation training. For (less than or approximately 10 pixels). The recall of the
images with roundabouts and storage tanks, we rotate the extremely small instances is very low.
patches randomly by four angles [π/2, π, −π/2, −π]. For im-
ages with the other categories, we rotate the angle randomly 6.2 State-of-the-Art Results on DOTA-v1.0
from [−π, π] continuously during training. We also rotate In this section, we compare the performance of Faster R-
the images at four angles ([0, π/2, π, 3π/2]) during testing. CNN OBB + RoI Transformer with the state-of-the-art al-
When performing test time augmentation, the results from gorithms on DOTA-v1.0 [14]. As shown in Tab. 13, Faster R-
images at different angles and scales are merged through the CNN OBB + RoI Transformer achieves an OBB mAP of 73.76
Non-Maximum Suppression (NMS) process in the original for DOTA-v1.0, and it outperforms all the previous state-
image coordinates. As shown in Tab. 12, both scale and of-the-art methods except that proposed by Li et al. [25].
rotation data augmentations improve the performance of Note that the method of Li et al. [25], SCRDet [24] and the
object detection by a large margin, which is consistent image cascade network (ICN) [19] all use multiple scales
with the large scale and orientation variations in DOTA. for training and testing to achieve high performance. The
Furthermore, this baseline model already used a feature method in [25] further used rotation data augmentation
pyramid network (FPN) and RoI Transformer. This indicates during training as described in Sec. 6.1.5. When using the
that the FPN and RoI Transformer do not completely solve same data augmentation, we achieve an mAP of 79.82. It
the problem of scale and rotation variations, and geometric outperforms the method in [25] by 3.46 points in OBB mAP
modeling with CNNs is still an open problem. and 1.96 points in HBB mAP. In addition, there is a signif-
icant improvement in densely packed small instances. (e.g.,

TABLE 11
Results using different number of proposals on DOTA-v2.0 test-dev. The speeds are tested on a single Tesla V100 GPU. The other settings are the
same with those in Tab. 6.
Method # of proposals 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000 10,000
OBB mAP (%) 51.72 52.81 52.81 53.24 53.29 53.51 53.70 53.94 53.93 53.92
Faster R-CNN OBB
HBB* mAP 52.56 53.37 53.37 54.63 54.86 55.07 55.08 55.09 55.08 55.06
+ RoI Transformer
speed (fps) 14.4 12.4 12.2 9.1 8.7 7.8 7.5 6.5 6 5.7
OBB mAP (%) 47.10 47.31 48.03 48.09 48.32 48.35 48.48 48.49 48.49 48.49
Faster R-CNN OBB HBB* mAP 48.44 49.37 49.46 49.71 49.74 50.09 50.37 50.39 50.38 50.47
speed (fps) 15.8 14.1 12.5 11.9 10.9 9.9 9.3 9.1 8.4 7.8

Fig. 12. Visualization of the results on DOTA-v2.0 test-dev. The five models are from the DOTA-v2.0 models in Tab. 6. The predictions with
scores above 0.1 are shown. The results illustrate the performance in cases of orientation variations, density variations, large ARs and small ARs.

TABLE 12 Task, and 22 teams submitted valid results on the HBB Task.
Data augmentation experiments on DOTA-v1.5. Each column in this The detailed leaderboards for the two tasks can be found on
table indicates an experiment configuration. The first column
represents our baseline method without additional data augmentations, the DOAI Challenge-2019 website10 , and the top 3 results
while the other columns gradually add augmentation. We use Faster are listed in Tab. 14. Notice that most of these results have
R-CNN OBB + RoI Transformer as the baseline. High overlap means been achieved by using an ensemble of detection models,
an overlap of 512 between patches instead of 200 as in Tab. 6. except [25] which used a single model and reported 74.9
Baseline Data augmentation and 77.9 in terms of mAP on the OBB and HBB tasks,
High overlap X X X X X
Multi scale Train X X X X
respectively. Both in the training and testing phase, multi-
Multi scale Test X X X scaling and rotation strategies were used for data augmenta-
Rotation Train X X tions. With the same settings, our single model [18] achieved
Rotation Test X 76.43 and 77.24 in terms of mAP on the OBB and HBB tasks
OBB mAP 65.03 67.57 69.44 73.62 76.43 77.60
HBB mAP 66.09 67.94 70.63 74.63 77.24 78.88 respectively, as shown in Tab. 12, which was the best results
reported on the OBB task.
the small vehicles, large vehicles, and ships). For example,
the detection performance for the large vehicle category 7 C ONCLUSION
gains an improvement of 12.18 points compared to the ODAI is challenging. To advance future research, we in-
previous results. troduce a large-scale dataset, DOTA, containing 1,793,658
instances annotated by OBBs. The DOTA statistics show that
6.3 DOAI 2019 Challenge Results it can well represent the real world well. Then, we build
DOTA-v1.5 was first used to hold the DOAI Challenge-2019 a code library for both oriented and horizontal ODAI to
in conjunction with CVPR 20199 . There were 173 registra- conduct a comprehensive evaluation. We hope these exper-
tions in total, 13 teams submitted valid results on the OBB iments can act as benchmarks for fair comparisons between
9. https://captain-whu.github.io/DOAI2019/ 10. https://captain-whu.github.io/DOAI2019/results.html

TABLE 13
State-of-the-art results on DOTA-v1.0 [14]. The short names for categories are defined as: BD–Baseball diamond, GTF–Ground field track,
SV–Small vehicle, LV–Large vehicle, TC–Tennis court, BC–Basketball court, SC–Storage tank, SBF–Soccer-ball field, RA–Roundabout,
SP–Swimming pool, and HC–Helicopter. FR-O indicates the Faster R-CNN OBB detector, which is the previous official baseline provided by
DOTA-v1.0 [14]. ICN [19] is the image cascade network. The LR-O + RT means Light Head R-CNN + RoI Transformer. DR-101-FPN means
deformable ResNet-101 with an FPN. SCRDet means small, cluttered and rotated object detector. R-101-SF-MDA means ResNet-101 with
sampling fusion network (SF-Net) and multi-dimensional attention network (MDA-Net). RT means RoI Transformer. Aug. means the data
augmentations in Sec. 6.1.5. FR-O* means the re-implemented Faster R-CNN OBB detector, which is slightly different from the original FR-O [14].
OBB Results
method backbone Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC mAP
FR-O [14] R-101 79.42 77.13 17.70 64.05 35.30 38.02 37.16 89.41 69.64 59.28 50.30 52.91 47.89 47.40 46.30 54.13
ICN [19] DR-101-FPN 81.36 74.30 47.70 70.32 64.89 67.82 69.98 90.76 79.06 78.20 53.64 62.90 67.02 64.17 50.23 68.16
LR-O + RT [18] R-101-FPN 88.64 78.52 43.44 75.92 68.81 73.68 83.59 90.74 77.27 81.46 58.39 53.54 62.83 58.93 47.67 69.56
SCRDet [24] R-101-SF-MDA 89.98 80.65 52.09 68.36 68.36 60.32 72.41 90.85 87.94 86.86 65.02 66.68 66.25 68.24 65.21 72.61
DRN [90] H-104 89.71 82.34 47.22 64.10 76.22 74.43 85.84 90.57 86.18 84.89 57.65 61.93 69.30 69.63 58.48 73.23
Gliding Vertex [63] R-101-FPN 89.64 85.00 52.26 77.34 73.01 73.14 86.82 90.74 79.02 86.81 59.55 70.91 72.94 70.86 57.32 75.02
CenterMap [66] R-101-FPN 89.83 84.41 54.60 70.25 77.66 78.32 87.19 90.66 84.89 85.27 56.46 69.23 74.13 71.56 66.06 76.03
CSL [91] R-152-FPN 90.25 85.53 54.64 75.31 70.44 73.51 77.62 90.84 86.15 86.69 69.60 68.04 73.83 71.10 68.93 76.17
Li et al. [25] R-101-FPN 90.41 85.21 55.00 78.27 76.19 72.19 82.14 90.70 87.22 86.87 66.62 68.43 75.43 72.70 57.99 76.36
2
S A-Net [57] R-50-FPN 88.89 83.60 57.74 81.95 79.94 83.19 89.11 90.78 84.87 87.81 70.30 68.25 78.30 77.01 69.58 79.42
FR-O* + RT [18] R-50-FPN 88.34 77.07 51.63 69.62 77.45 77.15 87.11 90.75 84.90 83.14 52.95 63.75 74.45 68.82 59.24 73.76
FR-O* + RT (Aug.) [18] R-50-FPN 87.89 85.01 57.83 78.55 75.22 84.37 88.04 90.88 87.28 85.79 71.04 69.67 79.00 83.29 73.43 79.82
HBB Results
method backbone Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC mAP
ICN [19] DR-101-FPN 89.97 77.71 53.38 73.26 73.46 65.02 78.22 90.79 79.05 84.81 57.20 62.11 73.45 70.22 58.08 72.45
SCRDet [24] R-101-SF-MDA 90.18 81.88 55.30 73.29 72.09 77.65 78.06 90.91 82.44 86.39 64.53 63.45 75.77 78.21 60.11 75.35
CenterMap [91] R-101-FPN 89.70 84.92 59.72 67.96 79.16 80.66 86.61 90.47 84.47 86.19 56.42 69.00 79.33 80.53 64.81 77.33
Li et al. [25] ResNet101 90.41 85.77 61.94 78.18 77.00 79.94 84.03 90.88 87.30 86.92 67.78 68.76 82.10 80.44 60.43 78.79
FR-O* + RT [18] R-50-FPN 88.47 81.00 54.10 69.19 78.42 81.16 87.35 90.75 84.90 83.55 52.63 62.97 75.89 71.31 57.22 74.59
FR-O* + RT (Aug.) [18] R-50-FPN 87.91 85.11 62.65 77.73 75.83 85.03 88.18 90.88 87.28 86.18 71.49 70.37 84.94 84.11 73.61 80.75

TABLE 14
DOAI 2019 Challenge Results. CC is the container crane for short. The other abbreviations for categories are the same as those in Tab. 13. The
USTC-NELSLIP, pca lab and czh, AICyber are the top 3 participants in the OBB and HBB Tasks. The FR-O means Faster R-CNN OBB. RT
means the RoI Transformer. Aug. means the data augmentation method described in Sec. 6.1.5. Note that FR-O + RT and FR-O + RT (Aug.) are
single models, while others are ensembles of multiple models.
OBB results
team (method) Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC CC mAP
USTC-NELSLIP [92] 89.19 85.32 57.27 80.86 73.87 81.26 89.5 90.84 85.94 85.62 69.5 76.73 76.34 76 77.84 57.33 78.34
pca lab [25] 89.11 83.83 59.55 82.8 66.93 82.51 89.78 90.88 85.36 84.22 71.95 77.89 78.47 74.27 74.77 53.22 77.84
czh 89 83.22 54.47 73.79 72.61 80.28 89.32 90.83 84.36 85 68.68 75.3 74.22 74.41 73.45 42.13 75.69
FR-O + RT [18] 71.92 76.07 51.87 69.24 52.05 75.18 80.72 90.53 78.58 68.26 49.18 71.74 67.51 65.53 62.16 9.99 65.03
FR-O + RT (Aug.) [18] 87.54 84.34 62.22 79.77 67.29 83.16 89.93 90.86 83.85 77.74 73.91 75.31 78.61 77.07 75.20 54.77 77.60
HBB results
team (method) Plane BD Bridge GTF SV LV Ship TC BC ST SBF RA Harbor SP HC CC mAP
pca lab [25] 88.26 86.55 65.68 79.83 74.59 79.35 88.12 90.86 85.45 84.15 73.9 77.44 84.1 81.07 76.07 57.07 79.53
USTC-NELSLIP [92] 89.26 85.6 59.61 80.86 75.2 81.13 89.58 90.84 85.94 85.71 69.5 76.34 81.7 81.84 76.53 57.09 79.17
AICyber 89.2 85.56 64.44 74.07 77.45 81.5 89.65 90.83 85.72 86.03 69.82 76.34 82.89 82.95 74.64 44.02 78.44
FR-O + RT [18] 71.92 75.21 54.09 68.10 52.54 74.87 80.79 90.46 78.58 68.41 51.57 71.48 74.91 74.84 56.66 13.01 66.09
FR-O + RT (Aug.) [18] 87.79 84.33 63.75 79.13 72.92 83.08 90.04 90.86 83.85 77.80 73.30 75.66 84.84 82.16 75.20 57.39 78.88

ODAI algorithms. The results show that hyperparameter advice on running Faster R-CNN, and Jinwang Wang for
selection and module design of the algorithms (e.g., number valuable discussions in details of parameter settings. The
of proposals) for aerial images are very different from those numerical calculations in this paper have been done on the
for natural images. It indicates that DOTA can be used as supercomputing system in the Supercomputing Center of
a supplement to natural scene images to facilitate universal Wuhan University.
object detection.
In the future, we will continue to extend the dataset,
host more challenges, and integrate more algorithms for
R EFERENCES
oriented object detection into our code library. We believe [1] E. J. Sadgrove, G. Falzon, D. Miron, and D. W. Lamb, “Real-time
that DOTA, challenges and code library will not only pro- object detection in agricultural/remote environments using the
multiple-expert colour feature extreme learning machine (mec-
mote the development of object detection in Earth vision elm),” Computers in Industry, vol. 98, pp. 183–191, 2018.
but also pose interesting algorithmic questions for general [2] V. Reilly, H. Idrees, and M. Shah, “Detection and tracking of large
object detection in computer vision. number of targets in wide area surveillance,” in ECCV. Springer,
2010, pp. 186–199.
[3] J. Porway, Q. Wang, and S. C. Zhu, “A hierarchical and contextual
ACKNOWLEDGMENT model for aerial image parsing,” IJCV, vol. 88, no. 2, pp. 254–283,
2010.
We thank the the support of CycloMedia B.V. for providing [4] Z. Liu, H. Wang, L. Weng, and Y. Yang, “Ship rotated bounding
the airborne images in DOTA-v2.0. We thank Huan Yi, box space for ship extraction from high-resolution optical satellite
Zhipeng Lin, Fan Hu, Pu Jin, Xinyi Tong, Xuan Hu, Zhipeng images with complex backgrounds,” IEEE Geosci. Remote Sensing
Dong, Liang Wu, Jun Tang, Linyan Cui, Duoyou Zhou, Lett., vol. 13, no. 8, pp. 1074–1078, 2016.
[5] T. Moranduzzo and F. Melgani, “Detecting cars in uav images
Tengteng Huang, and all the others who involved in the with a catalog-based approach,” IEEE Trans. Geosci. Remote Sens.,
annotations of DOTA. We also thank Zhen Zhu for his vol. 52, no. 10, pp. 6356–6367, 2014.

[6] G. Cheng, J. Han, P. Zhou, and D. Xu, “Learning rotation-invariant [32] B. Zhou, À. Lapedriza, J. Xiao, A. Torralba, and A. Oliva, “Learn-
and fisher discriminative convolutional neural networks for object ing deep features for scene recognition using places database,” in
detection,” IEEE TIP, vol. 28, no. 1, pp. 265–278, 2018. NIPS, 2014, pp. 487–495.
[7] C. Benedek, X. Descombes, and J. Zerubia, “Building development [33] Q. You, J. Luo, H. Jin, and J. Yang, “Building a large scale dataset
monitoring in multitemporal remotely sensed image pairs with for image emotion recognition: The fine print and the benchmark,”
stochastic birth-death dynamics,” IEEE TPAMI, vol. 34, no. 1, pp. in AAAI, 2016, pp. 308–314.
33–50, 2012. [34] G. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, L. Zhang, and X. Lu,
[8] M.-R. Hsieh, Y.-L. Lin, and W. H. Hsu, “Drone-based object count- “AID: A benchmark data set for performance evaluation of aerial
ing by spatially regularized regional proposal network,” in ICCV, scene classification,” IEEE Trans. Geosci. Remote Sens., vol. 55, no. 7,
2017, pp. 4145–4153. pp. 3965–3981, 2017.
[9] G. Cheng, P. Zhou, and J. Han, “Learning rotation-invariant con- [35] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-
volutional neural networks for object detection in VHR optical Tuset, S. Kamali, S. Popov, M. Malloci, T. Duerig et al., “The open
remote sensing images,” IEEE Trans. Geosci. Remote Sens., vol. 54, images dataset v4: Unified image classification, object detection,
no. 12, pp. 7405–7415, 2016. and visual relationship detection at scale,” arXiv:1811.00982, 2018.
[10] Y. Zhou, Q. Ye, Q. Qiu, and J. Jiao, “Oriented response networks,” [36] N. Weir, D. Lindenbaum, A. Bastidas, A. V. Etten, S. McPherson,
in CVPR. IEEE, 2017, pp. 4961–4970. J. Shermeyer, V. Kumar, and H. Tang, “Spacenet mvoi: A multi-
[11] M. Everingham, L. V. Gool, C. Williams, J. Winn, and A. Zisser- view overhead imagery dataset,” in ICCV, October 2019.
man, “The pascal visual object classes (VOC) challenge,” IJCV, [37] T. N. Mundhenk, G. Konjevod, W. A. Sakla, and K. Boakye, “A
vol. 88, no. 2, pp. 303–338, 2010. large contextual dataset for classification, detection and counting
[12] J. Deng, W. Dong, R. Socher, L. Li, K. Li, and F. Li, “Imagenet: A of cars with deep learning,” in ECCV, 2016, pp. 785–800.
large-scale hierarchical image database,” in CVPR, 2009, pp. 248– [38] M. Y. Yang, W. Liao, X. Li, and B. Rosenhahn, “Deep learning for
255. vehicle detection in aerial images,” in ICIP. IEEE, 2018, pp. 3079–
[13] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, 3083.
P. Dollár, and C. L. Zitnick, “Microsoft COCO: common objects in [39] M. Y. Yang, W. Liao, X. Li, Y. Cao, and B. Rosenhahn, “Vehicle
context,” in ECCV, 2014, pp. 740–755. detection in aerial images,” Photogrammetric engineering and remote
[14] G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, sensing : PE&RS, vol. 85, no. 4, pp. 297–304, 2019.
M. Pelillo, and L. Zhang, “Dota: A large-scale dataset for object [40] K. Chen, M. Wu, J. Liu, and C. Zhang, “Fgsd: A dataset for
detection in aerial images,” in CVPR, 2018. fine-grained ship detection in high resolution satellite images,”
[15] S. Razakarivony and F. Jurie, “Vehicle detection in aerial imagery: arXiv:2003.06832, 2020.
A small target detection benchmark,” J Vis. Commun. Image R., [41] J. Shermeyer, T. Hossler, A. Van Etten, D. Hogan, R. Lewis,
vol. 34, pp. 187–203, 2016. and D. Kim, “Rareplanes: Synthetic data takes flight,”
[16] K. Liu and G. Máttyus, “Fast multiclass vehicle detection on aerial arXiv:2006.02963, 2020.
images,” IEEE Geosci. Remote Sensing Lett., vol. 12, no. 9, pp. 1938– [42] G. Cheng, J. Han, P. Zhou, and L. Guo, “Multi-class geospatial
1942, 2015. object detection and geographic image classification based on
collection of part detectors,” ISPRS Journal of Photogrammetry and
[17] H. Zhu, X. Chen, W. Dai, K. Fu, Q. Ye, and J. Jiao, “Orientation
Remote Sensing, vol. 98, pp. 119–132, 2014.
robust object detection in aerial images using deep convolutional
neural network,” in ICIP, 2015, pp. 3735–3739. [43] Y. Long, Y. Gong, Z. Xiao, and Q. Liu, “Accurate object local-
ization in remote sensing images based on convolutional neural
[18] J. Ding, N. Xue, Y. Long, G.-S. Xia, and Q. Lu, “Learning roi
networks,” IEEE Trans. Geosci. Remote Sens., vol. 55, no. 5, pp. 2486–
transformer for oriented object detection in aerial images,” in
2498, 2017.
CVPR, 2019, pp. 2849–2858.
[44] Z. Zou and Z. Shi, “Random access memories: A new paradigm for
[19] S. M. Azimi, E. Vig, R. Bahmanyar, M. Körner, and P. Reinartz, target detection in high resolution aerial remote sensing images,”
“Towards multi-class object detection in unconstrained remote IEEE TIP, vol. 27, no. 3, pp. 1100–1111, 2017.
sensing imagery,” in ACCV. Springer, 2018, pp. 150–165.
[45] Y. Zhang, Y. Yuan, Y. Feng, and X. Lu, “Hierarchical and robust
[20] J. Pang, C. Li, J. Shi, Z. Xu, and H. Feng, “R2 -cnn: Fast tiny object convolutional neural network for very high-resolution remote
detection in large-scale remote sensing images,” IEEE Trans. Geosci. sensing object detection,” IEEE Trans. Geosci. Remote Sens., 2019.
Remote Sens., 2019. [46] S. Waqas Zamir, A. Arora, A. Gupta, S. Khan, G. Sun, F. Shah-
[21] R. LaLonde, D. Zhang, and M. Shah, “Clusternet: Detecting small baz Khan, F. Zhu, L. Shao, G.-S. Xia, and X. Bai, “isaid: A large-
objects in large scenes by exploiting spatio-temporal information,” scale dataset for instance segmentation in aerial images,” in CVPR
in CVPR, June 2018. Workshops, 2019, pp. 28–37.
[22] F. Yang, H. Fan, P. Chu, E. Blasch, and H. Ling, “Clustered object [47] K. Li, G. Wan, G. Cheng, L. Meng, and J. Han, “Object detection in
detection in aerial images,” in CVPR, 2019, pp. 8311–8320. optical remote sensing images: A survey and a new benchmark,”
[23] M. Liao, Z. Zhu, B. Shi, G.-s. Xia, and X. Bai, “Rotation-sensitive ISPRS Journal of Photogrammetry and Remote Sensing, vol. 159, pp.
regression for oriented scene text detection,” in CVPR, 2018, pp. 296–307, 2020.
5909–5918. [48] P. Zhu, L. Wen, X. Bian, H. Ling, and Q. Hu, “Vision meets drones:
[24] X. Yang, J. Yang, J. Yan, Y. Zhang, T. Zhang, Z. Guo, X. Sun, and A challenge,” arXiv:1804.07437, 2018.
K. Fu, “Scrdet: Towards more robust detection for small, cluttered [49] A. Torralba and A. A. Efros, “Unbiased look at dataset bias,” in
and rotated objects,” in ICCV, 2019, pp. 8232–8241. CVPR, 2011, pp. 1521–1528.
[25] C. Li, C. Xu, Z. Cui, D. Wang, Z. Jie, T. Zhang, and J. Yang, [50] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed, C. Fu, and
“Learning object-wise semantic representation for detection in A. C. Berg, “SSD: single shot multibox detector,” in ECCV, 2016,
remote sensing imagery,” in CVPR Workshops, 2019, pp. 20–27. pp. 21–37.
[26] G. Heitz and D. Koller, “Learning spatial context: Using stuff to [51] J. Redmon and A. Farhadi, “Yolo9000: better, faster, stronger,” in
find things,” in ECCV, 2008, pp. 30–43. CVPR, 2017, pp. 7263–7271.
[27] D. Lam, R. Kuzma, K. McGee, S. Dooley, M. Laielli, M. Klaric, [52] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hier-
Y. Bulatov, and B. McCord, “xview: Objects in context in overhead archies for accurate object detection and semantic segmentation,”
imagery,” arXiv:1802.07856, 2018. in CVPR, 2014, pp. 580–587.
[28] K. Chen, J. Wang, J. Pang, Y. Cao, Y. Xiong, X. Li, S. Sun, W. Feng, [53] R. Girshick, “Fast r-cnn,” in CVPR, 2015, pp. 1440–1448.
Z. Liu, J. Xu et al., “Mmdetection: Open mmlab detection toolbox [54] S. Ren, K. He, R. B. Girshick, and J. Sun, “Faster R-CNN: towards
and benchmark,” arXiv:1906.07155, 2019. real-time object detection with region proposal networks,” IEEE
[29] R. Girshick, I. Radosavovic, G. Gkioxari, P. Dollár, and K. He, “De- TPAMI, vol. 39, no. 6, pp. 1137–1149, 2017.
tectron,” https://github.com/facebookresearch/detectron, 2018. [55] J. Dai, H. Qi, Y. Xiong, Y. Li, G. Zhang, H. Hu, and Y. Wei, “De-
[30] X. Wang, Z. Cai, D. Gao, and N. Vasconcelos, “Towards universal formable convolutional networks,” CoRR, abs/1703.06211, vol. 1,
object detection by domain attention,” in CVPR, 2019, pp. 7289– no. 2, p. 3, 2017.
7298. [56] Z. Liu, J. Hu, L. Weng, and Y. Yang, “Rotated region based cnn for
[31] B. Yao, X. Yang, and S.-C. Zhu, “Introduction to a large-scale ship detection,” in ICIP. IEEE, 2017, pp. 900–904.
general purpose ground truth database: Methodology, annotation [57] J. Han, J. Ding, J. Li, and G.-S. Xia, “Align deep features for
tool and benchmarks,” in EMMCVPR, 2007, pp. 169–183. oriented object detection,” arXiv:2008.09397, 2020.

[58] T.-Y. Lin, P. Dollár, R. B. Girshick, K. He, B. Hariharan, and S. J. drasekhar, S. Lu, F. Shafait, S. Uchida, and E. Valveny, “ICDAR
Belongie, “Feature pyramid networks for object detection.” in 2015 competition on robust reading,” in ICDAR, 2015.
CVPR, vol. 1, no. 2, 2017, p. 3. [84] S. Yang, P. Luo, C. C. Loy, and X. Tang, “WIDER FACE: A face
[59] F. Wang, L. Zhao, X. Li, X. Wang, and D. Tao, “Geometry-aware detection benchmark,” in CVPR, 2016, pp. 5525–5533.
scene text detection with instance transformation network,” in [85] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for
CVPR, 2018, pp. 1381–1389. dense object detection,” in ICCV, 2017, pp. 2980–2988.
[60] C. Huang, H. Ai, Y. Li, and S. Lao, “High-performance rotation [86] K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask r-cnn,” in
invariant multiview face detection,” IEEE TPAMI, vol. 29, no. 4, ICCV. IEEE, 2017, pp. 2980–2988.
pp. 671–686, 2007. [87] Z. Li, C. Peng, G. Yu, X. Zhang, Y. Deng, and J. Sun, “Light-head
[61] X. Shi, S. Shan, M. Kan, S. Wu, and X. Chen, “Real-time rotation- r-cnn: In defense of two-stage object detector,” arXiv:1711.07264,
invariant face detection with progressive calibration networks,” in 2017.
CVPR, 2018, pp. 2295–2303. [88] K. Chen, J. Pang, J. Wang, Y. Xiong, X. Li, S. Sun, W. Feng,
[62] M. Liao, B. Shi, and X. Bai, “Textboxes++: A single-shot oriented Z. Liu, J. Shi, W. Ouyang et al., “Hybrid task cascade for instance
scene text detector,” IEEE TIP, vol. 27, no. 8, pp. 3676–3690, 2018. segmentation,” in CVPR, 2019, pp. 4974–4983.
[63] Y. Xu, M. Fu, Q. Wang, Y. Wang, K. Chen, G.-S. Xia, and X. Bai, [89] J. Ma, W. Shao, H. Ye, L. Wang, H. Wang, Y. Zheng, and X. Xue,
“Gliding vertex on the horizontal bounding box for multi-oriented “Arbitrary-oriented scene text detection via rotation proposals,”
object detection,” IEEE TPAMI, 2020. TMM, 2018.
[64] X. Yang and J. Yan, “Arbitrary-oriented object detection with [90] X. Pan, Y. Ren, K. Sheng, W. Dong, H. Yuan, X. Guo, C. Ma, and
circular smooth label,” ECCV, 2020. C. Xu, “Dynamic refinement network for oriented and densely
[65] J. Wang, J. Ding, H. Guo, W. Cheng, T. Pan, and W. Yang, “Mask packed object detection,” in CVPR, 2020, pp. 11 207–11 216.
obb: A semantic attention-based mask oriented bounding box rep- [91] X. Yang and J. Yan, “Arbitrary-oriented object detection with
resentation for multi-category object detection in aerial images,” circular smooth label,” arXiv:2003.05597, 2020.
Remote Sensing, vol. 11, no. 24, p. 2930, 2019. [92] Y. Zhu, X. Wu, and J. Du, “Adaptive period embedding for
[66] J. Wang, W. Yang, H.-C. Li, H. Zhang, and G.-S. Xia, “Learning representing oriented objects in aerial images,” arXiv:1906.09447,
center probability map for detecting objects in aerial images,” IEEE 2019.
Trans. Geosci. Remote Sens., 2020. Jian Ding is currently pursuing his Ph.D de-
[67] B. Uzkent, C. Yeh, and S. Ermon, “Efficient object detection in gree at the State Key Laboratory of Informa-
large images using deep reinforcement learning,” in WACV, 2020, tion Engineering in Surveying, Mapping and Re-
pp. 1824–1833. mote Sensing, Wuhan University. He received
[68] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara, the B.S. degree in Aircraft Design and Engineer-
A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama et al., ing from Northwestern Polytechnical University,
“Speed/accuracy trade-offs for modern convolutional object de- Xian, China in 2017. His research interests in-
tectors,” in CVPR, 2017, pp. 7310–7311. clude object detection, instance segmentation
[69] F. Massa and R. Girshick, “maskrcnn-benchmark: Fast, mod- and remote sensing.
ular reference implementation of Instance Segmentation and
Object Detection algorithms in PyTorch,” https://github.com/
Nan Xue is currently a Research Associate
facebookresearch/maskrcnn-benchmark, 2018.
Professor in the School of Computer Science,
[70] Y. Wu, A. Kirillov, F. Massa, W.-Y. Lo, and R. Girshick, “Detec-
Wuhan University. He received the B.S., and
tron2,” https://github.com/facebookresearch/detectron2, 2019.
Ph.D. degrees from Wuhan University in 2014
[71] Y. Chen, C. Han, Y. Li, Z. Huang, Y. Jiang, N. Wang, and Z. Zhang, and 2020 respectively. He was a visiting scholar
“Simpledet: A simple and versatile distributed framework for at North Carolina State University from Sep.
object detection and instance recognition,” JMLR, vol. 20, no. 156, 2018 to June 2020. His research interests in-
pp. 1–8, 2019. clude geometric structure analysis in computer
[72] A.-M. de Oca, R. Bahmanyar, N. Nistor, and M. Datcu, “Earth vision.
observation image semantic bias: A collaborative user annotation
approach,” J-STARS, 2017.
[73] Z. Zhang, G. Vosselman, M. Gerke, C. Persello, D. Tuia, and Gui-Song Xia received his Ph.D. degree in im-
M. Y. Yang, “Detecting building changes between airborne laser age processing and computer vision from CNRS
scanning and photogrammetric data,” Remote sensing, vol. 11, LTCI, Télécom ParisTech, Paris, France, in 2011.
no. 20, p. 2417, 2019. From 2011 to 2012, he has been a Post-Doctoral
[74] J. Xiao, J. Hays, K. Ehinger, A. Oliva, and A. Torralba, “SUN Researcher with the Centre de Recherche en
database: Large-scale scene recognition from abbey to zoo,” in Mathématiques de la Decision, CNRS, Paris-
CVPR, 2010, pp. 3485–3492. Dauphine University, Paris, for one and a half
[75] R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz, years. He is currently working as a full professor
S. Chen, Y. Kalantidis, L.-J. Li, D. A. Shamma et al., “Visual at Wuhan University. He has also been work-
genome: Connecting language and vision using crowdsourced ing as Visiting Scholar at DMA, École Normale
dense image annotations,” IJCV, vol. 123, no. 1, pp. 32–73, 2017. Supérieure (ENS-Paris) for two months in 2018.
[76] D. P. Papadopoulos, J. R. Uijlings, F. Keller, and V. Ferrari, “Ex- His current research interests include mathematical modeling of images
treme clicking for efficient object annotation,” in ICCV, 2017, pp. and videos, structure from motion, perceptual grouping, and remote
4930–4939. sensing image understanding. He serves on the Editorial Boards of
[77] Z. Wu, N. Bodla, B. Singh, M. Najibi, R. Chellappa, and L. S. Davis, several journals, including Pattern Recognition, Signal Processing: Im-
“Soft sampling for robust object detection,” arXiv:1806.06986, 2018. age Communications, EURASIP Journal on Image & Video Processing,
[78] H. Zhang, F. Chen, Z. Shen, Q. Hao, C. Zhu, and M. Savvides, Journal of Remote Sensing, and Frontiers in Computer Science: Com-
“Solving missing-annotation object detection with background puter Vision.
recalibration loss,” in ICASSP. IEEE, 2020, pp. 1888–1892.
Xiang Bai received the B.S., M.S., and Ph.D.
[79] Y. He, C. Zhu, J. Wang, M. Savvides, and X. Zhang, “Bounding
degrees from the Huazhong University of Sci-
box regression with uncertainty for accurate object detection,” in
ence and Technology (HUST), Wuhan, China, in
CVPR, 2019, pp. 2888–2897.
2003, 2005, and 2009, respectively, all in elec-
[80] W. Li, W. Wei, and L. Zhang, “Gsdet: Object detection in aerial
tronics and information engineering. He is cur-
images based on scale reasoning,” IEEE TIP, vol. 30, pp. 4599–
rently a Professor with the School of Electronic
4609, 2021.
Information and Communications and the Vice-
[81] B. Singh and L. S. Davis, “An analysis of scale invariance in object Director of the National Center of AntiCounter-
detection snip,” in CVPR, 2018, pp. 3578–3587. feiting Technology, HUST. His research interests
[82] wiki, “gsd,” https://en.wikipedia.org/wiki/Ground sample include object recognition, shape analysis, and
distance. scene text recognition.
[83] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. K. Ghosh, A. D.
Bagdanov, M. Iwamura, J. Matas, L. Neumann, V. R. Chan-

Wen Yang received the Ph.D. degree in com- Mihai Datcu received the M.S. and Ph.D.
munication and information system from Wuhan degrees in electronics and telecommunications
University, Wuhan, China, in 2004. From 2008 to from the University Politehnica of Bucharest
2009, he was a visiting Scholar with the Labora- UPB, Bucharest, Romania, in 1978 and
toire Jean Kuntzmann (LJK), Grenoble, France. 1986, and the title Habilitation a diriger des
From 2010 to 2013, he was a Post-Doctoral recherches from Universit Louis Pasteur,
Researcher with the State Key Laboratory of In- Strasbourg, France. He holds a Professorship
formation Engineering, Surveying, Mapping and in electronics and telecommunications with
Remote Sensing (LIESMARS), Wuhan Univer- UPB since 1981. Since 1993, he has been a
sity. Since then, he has been a Full Professor Scientist with the German Aerospace Center
with the School of Electronic Information, Wuhan (DLR), Oberpfaffenhofen, Germany. He is
University. His research interests include object detection and recogni- currently developing algorithms for model-based information retrieval
tion, semantic segmentation, and change detection. from high-complexity signals, methods for scene understanding from
SAR and interferometric SAR data, and he is engaged in research
in information theoretical aspects and semantic representations in
Michael Ying Yang is currently Assistant Pro- advanced communication systems. His research interests are in
fessor in the Department of Earth Observation Bayesian inference, information and complexity theory, stochastic
Science at ITC - Faculty of Geo-Information processes, model-based scene understanding, image information
Science and Earth Observation, University of mining, for applications in information retrieval and understanding of
Twente, The Netherlands, heading a group work- high-resolution SAR and optical observations.
ing on scene understanding. He received the Marcello Pellilo is a Full Professor of Com-
PhD degree (summa cum laude) from University puter Science at Ca Foscari University, Venice,
of Bonn (Germany) in 2011. He received the where he leads the Computer Vision and Pattern
venia legendi in Computer Science from Leibniz Recognition Lab. He has been the Director of the
University Hannover in 2016. His research inter- European Centre for Living Technology (ECLT)
ests are in the fields of computer vision and pho- and has held visiting research/teaching positions
togrammetry with specialization on scene understanding and semantic in several institutions including Yale University
interpretation from imagery. He serves as Associate Editor of ISPRS (USA), University College London (UK), McGill
Journal of Photogrammetry and Remote Sensing, Co-chair of ISPRS University (Canada), University of Vienna (Aus-
working group II/5 Dynamic Scene Analysis, Program Chair of ISPRS tria), York University (UK), NICTA (Australia),
Geospatial Week 2019, and recipient of ISPRS President’s Honorary Wuhan University (China), Huazhong University
Citation (2016), Best Science Paper Award at BMVC (2016), and The of Science and Technology (China), and South China University of
Willem Schermerhorn Award (2020). Technology (China). He is also an external affiliate of the Computer
Serge Belongie received a B.S. (with honor) in Science Department at Drexel University (USA). His research interests
EE from Caltech in 1995 and a Ph.D. in EECS are in the areas of computer vision, machine learning and pattern
from Berkeley in 2000. While at Berkeley, his recognition where he has published more than 200 technical papers
research was supported by an NSF Graduate in refereed journals, handbooks, and conference proceedings. He has
Research Fellowship. From 2001-2013 he was been General Chair for ICCV 2017, Program Chair for ICPR 2020,
a professor in the Department of Computer Sci- and has been Track or Area Chair for several conferences in his area.
ence and Engineering at University of Califor- He is the Specialty Chief Editor of Frontiers in Computer Vision and
nia, San Diego. He is currently a professor at serves, or has served, on the Editorial Boards of several journals, includ-
Cornell Tech and the Department of Computer ing IEEE Transactions on Pattern Analysis and Machine Intelligence,
Science at Cornell University. His research inter- Pattern Recognition, IET Computer Vision, and Brain Informatics. He
ests include Computer Vision, Machine Learn- also serves on the Advisory Board of Springers International Journal of
ing, Crowdsourcing and Human-in-the-Loop Computing. He is also a co- Machine Learning and Cybernetics. Prof. Pelillo has been elected Fellow
founder of several companies including Digital Persona, Anchovi Labs of the IEEE and Fellow of the IAPR, and is an IEEE SMC Distinguished
and Orpix. He is a recipient of the NSF CAREER Award, the Alfred Lecturer.
P. Sloan Research Fellowship, the MIT Technology Review Innovators Liangpei Zhang received the B.S. degree in
Under 35 Award and the Helmholtz Prize for fundamental contributions physics from Hunan Normal University, Chang-
in Computer Vision. sha, China, in 1982, the M.S. degree in op-
Jiebo Luo joined the University of Rochester tics from the Xian Institute of Optics and Preci-
in Fall 2011 after over fifteen prolific years at sion Mechanics, Chinese Academy of Sciences,
Kodak Research Laboratories, where he was a Xian, China, in 1988, and the Ph.D. degree
Senior Principal Scientist leading research and in photogrammetry and remote sensing from
advanced development. He has been involved in Wuhan University, Wuhan, China, in 1998. He is
numerous technical conferences, including serv- currentlya Chang-Jiang Scholar Chair Professor
ing as the program co-chair of ACM Multimedia with Wuhan University, appointed by the Min-
2010, IEEE CVPR 2012 and IEEE ICIP 2017. He istry of Education of China. He has authored
has served on the editorial boards of the IEEE or coauthored over 500 research papers and ve books. He holds 15
Transactions on Pattern Analysis and Machine patents. His research interests include hyper spectral remote sensing,
Intelligence, IEEE Transactions on Multimedia, high resolution remote sensing, image processing, and articial intelli-
IEEE Transactions on Circuits and Systems for Video Technology, ACM gence. Dr. Zhang was a recipient of the 2010 Best Paper Boeing Award
Transactions on Intelligent Systems and Technology, Pattern Recogni- and the 2013 Best Paper ERDAS Award from the American Society of
tion, Machine Vision and Applications, and Journal of Electronic Imag- Photogrammetry and Remote Sensing. He serves as a Co-Chair for
ing. Dr. Luo will serve as the Editor-in-Chief of the IEEE Transactions on the series SPIE Conferences on Multispectral Image Processing and
Multimedia for the 2020-2022 term. Dr. Luo is a Fellow of SPIE, IAPR, Pattern Recognition, the Conference on Asia Remote Sensing, and
IEEE, ACM, and AAAI. In addition, he is a Board Member of the Greater many other conferences. He serves as an Associate Editor for the IEEE
Rochester Data Science Industry Consortium. Transactions on Geoscience and Remote Sensing. He is a fellow of
IEEE.

ITAA_Gateway_AdminManual
No ratings yet
ITAA_Gateway_AdminManual
36 pages
CCS370 UI and UX Design Lecture Notes 2
No ratings yet
CCS370 UI and UX Design Lecture Notes 2
237 pages
MODICOM 3 - 1 Board Overview - LJ Create
No ratings yet
MODICOM 3 - 1 Board Overview - LJ Create
6 pages
Object Detection in Aerial Images: A Large-Scale Benchmark and Challenges
No ratings yet
Object Detection in Aerial Images: A Large-Scale Benchmark and Challenges
18 pages
附件5
No ratings yet
附件5
29 pages
Application of Machine Learning For Object Detecti
No ratings yet
Application of Machine Learning For Object Detecti
7 pages
Attention_and_Feature_Fusion_SSD_for_Remote_Sensing_Object_Detection
No ratings yet
Attention_and_Feature_Fusion_SSD_for_Remote_Sensing_Object_Detection
9 pages
Boundary Viên Thuốc
No ratings yet
Boundary Viên Thuốc
12 pages
DOTA: A Large-Scale Dataset For Object Detection in Aerial Images
No ratings yet
DOTA: A Large-Scale Dataset For Object Detection in Aerial Images
17 pages
OPODet_Toward_Open_World_Potential_Oriented_Object_Detection_in_Remote_Sensing_Images
No ratings yet
OPODet_Toward_Open_World_Potential_Oriented_Object_Detection_in_Remote_Sensing_Images
13 pages
chen1
No ratings yet
chen1
10 pages
Seminar Paper by Roquia Salam
No ratings yet
Seminar Paper by Roquia Salam
29 pages
SSRN 4542996
No ratings yet
SSRN 4542996
17 pages
s00530-025-01688-7
No ratings yet
s00530-025-01688-7
19 pages
Detection of Multiclass Objects in Optical Remote Sensing Images
No ratings yet
Detection of Multiclass Objects in Optical Remote Sensing Images
5 pages
Doloriel Improving The Detection of Small Oriented Objects in Aerial Images WACVW 2023 Paper
No ratings yet
Doloriel Improving The Detection of Small Oriented Objects in Aerial Images WACVW 2023 Paper
10 pages
drones-07-00526
No ratings yet
drones-07-00526
23 pages
Object Detection in Satellite Images by Faster R-CNN Incorporated With Enhanced ROI Pooling (FrRNet-ERoI) Framework
No ratings yet
Object Detection in Satellite Images by Faster R-CNN Incorporated With Enhanced ROI Pooling (FrRNet-ERoI) Framework
18 pages
Orientedformer: An End-To-End Transformer-Based Oriented Object Detector in Remote Sensing Images
No ratings yet
Orientedformer: An End-To-End Transformer-Based Oriented Object Detector in Remote Sensing Images
16 pages
06 - 2022 - Liao - GLE-Net - A Global and Local Ensemble Network For Aerial Object Detection
No ratings yet
06 - 2022 - Liao - GLE-Net - A Global and Local Ensemble Network For Aerial Object Detection
12 pages
Applsci 13 12977
No ratings yet
Applsci 13 12977
21 pages
YOLOv4 With Deformable-Embedding-Transformer Featu
No ratings yet
YOLOv4 With Deformable-Embedding-Transformer Featu
23 pages
Remote Sensing: Improved YOLO Network For Free-Angle Remote Sensing Target Detection
No ratings yet
Remote Sensing: Improved YOLO Network For Free-Angle Remote Sensing Target Detection
20 pages
Object Detection Using Adaptive Mask RCNN
No ratings yet
Object Detection Using Adaptive Mask RCNN
12 pages
Sensors 23 08118
No ratings yet
Sensors 23 08118
21 pages
Focus-And-Detect A Small Object Detection Framework For Aerial Images
No ratings yet
Focus-And-Detect A Small Object Detection Framework For Aerial Images
9 pages
remotesensing-14-02385-v2
No ratings yet
remotesensing-14-02385-v2
41 pages
mSODANet
No ratings yet
mSODANet
10 pages
44 Review Remote Sensing For Object Detection
No ratings yet
44 Review Remote Sensing For Object Detection
29 pages
Improved YOLOv7-Tiny for Object Detection Based On
No ratings yet
Improved YOLOv7-Tiny for Object Detection Based On
23 pages
1-s2.0-S2667393224000139-main
No ratings yet
1-s2.0-S2667393224000139-main
8 pages
Sensors 23 06423
No ratings yet
Sensors 23 06423
23 pages
Multi-Scale Vehicle Detection in High-Resolution Aerial Images With Context Information
No ratings yet
Multi-Scale Vehicle Detection in High-Resolution Aerial Images With Context Information
15 pages
ref19
No ratings yet
ref19
6 pages
remotesensing-13-00965-v2
No ratings yet
remotesensing-13-00965-v2
18 pages
An Improved Rotation Invariant CNN-based Detector With Rotatable Bounding Boxes For Aerial Image Detection
No ratings yet
An Improved Rotation Invariant CNN-based Detector With Rotatable Bounding Boxes For Aerial Image Detection
5 pages
EBSCO FullText 2024 09 02
No ratings yet
EBSCO FullText 2024 09 02
24 pages
Multiple Objects and Road Detection in Unmanned Aerial Vehicle
No ratings yet
Multiple Objects and Road Detection in Unmanned Aerial Vehicle
19 pages
Remotesensing 14 00984 v2
No ratings yet
Remotesensing 14 00984 v2
21 pages
Remotesensing 16 00327
No ratings yet
Remotesensing 16 00327
28 pages
LWUAVDet_A_Lightweight_UAV_Object_Detection_Network_on_Edge_Devices
No ratings yet
LWUAVDet_A_Lightweight_UAV_Object_Detection_Network_on_Edge_Devices
11 pages
Comparative analysis of feature descriptors and classifiers for real-time object detection
No ratings yet
Comparative analysis of feature descriptors and classifiers for real-time object detection
11 pages
Multiscale Object Detection in Remote Sensing Images Using 1qh06jan
No ratings yet
Multiscale Object Detection in Remote Sensing Images Using 1qh06jan
10 pages
Drones 07 00682
No ratings yet
Drones 07 00682
18 pages
2209.13351v2-SuperYlo
No ratings yet
2209.13351v2-SuperYlo
14 pages
A_Relation-Augmented_Embedded_Graph_Attention_Network_for_Remote_Sensing_Object_
No ratings yet
A_Relation-Augmented_Embedded_Graph_Attention_Network_for_Remote_Sensing_Object_
18 pages
J12-TGRS2023-SuperYOLO
No ratings yet
J12-TGRS2023-SuperYOLO
15 pages
A_Survey_of_Deep_Learning-Based_Object_Detection_M
No ratings yet
A_Survey_of_Deep_Learning-Based_Object_Detection_M
17 pages
remotesensing-17-01027
No ratings yet
remotesensing-17-01027
17 pages
Convolutional SVM Networks For Object Detection in UAV Imagery
No ratings yet
Convolutional SVM Networks For Object Detection in UAV Imagery
12 pages
Remotesensing 15 03265
No ratings yet
Remotesensing 15 03265
29 pages
Ymer 230109
100% (1)
Ymer 230109
11 pages
drones-07-00188
No ratings yet
drones-07-00188
18 pages
Walter Avila Cordova 2020 J. Phys. Conf. Ser. 1642 012003
No ratings yet
Walter Avila Cordova 2020 J. Phys. Conf. Ser. 1642 012003
11 pages
Large Selective Kernel Network For Remote Sensing Object Detection
No ratings yet
Large Selective Kernel Network For Remote Sensing Object Detection
16 pages
sensors-23-07190
No ratings yet
sensors-23-07190
27 pages
Drones 07 00304 v2
No ratings yet
Drones 07 00304 v2
26 pages
Drones 05 00066 v3
No ratings yet
Drones 05 00066 v3
24 pages
Remotesensing 13 00516 v3
No ratings yet
Remotesensing 13 00516 v3
19 pages
A novel model to detect and categorize objects from images by using a hybrid machine learning model
No ratings yet
A novel model to detect and categorize objects from images by using a hybrid machine learning model
13 pages
2023 - FPGA - Accelerator - For - Meta-Recognition - Anomaly - Detection - Case - of - Burned - Area - Detection
No ratings yet
2023 - FPGA - Accelerator - For - Meta-Recognition - Anomaly - Detection - Case - of - Burned - Area - Detection
13 pages
5-Jul-11093 paper
No ratings yet
5-Jul-11093 paper
5 pages
Emerging Technologies in Space Exploration
From Everand
Emerging Technologies in Space Exploration
Matthew N. O. Sadiku
No ratings yet
Green University of Bangladesh: Faculty of Science and Engineering
No ratings yet
Green University of Bangladesh: Faculty of Science and Engineering
4 pages
Cross Entropy Loss Intro, Applications
No ratings yet
Cross Entropy Loss Intro, Applications
21 pages
4040-PRO User Manual V1.0 EN DE JP 202210
No ratings yet
4040-PRO User Manual V1.0 EN DE JP 202210
106 pages
Green University of Bangladesh: Faculty of Science and Engineering
No ratings yet
Green University of Bangladesh: Faculty of Science and Engineering
4 pages
Determination of Transformer Equivalent Circuit Parameters of A
No ratings yet
Determination of Transformer Equivalent Circuit Parameters of A
18 pages
Green University of Bangladesh: Faculty of Science and Engineering
No ratings yet
Green University of Bangladesh: Faculty of Science and Engineering
3 pages
The MODICOM 3 - 2 PCM Receiver Overview - LJ Create
No ratings yet
The MODICOM 3 - 2 PCM Receiver Overview - LJ Create
6 pages
Green University of Bangladesh: Faculty of Science and Engineering
No ratings yet
Green University of Bangladesh: Faculty of Science and Engineering
4 pages
Determination of Transformer Equivalent Circuit Parameters of A Single
No ratings yet
Determination of Transformer Equivalent Circuit Parameters of A Single
16 pages
Experiment: 04 Name of The Experiment: No Load and Loading Characteristics of Three-Phase Alternator. Objectives
No ratings yet
Experiment: 04 Name of The Experiment: No Load and Loading Characteristics of Three-Phase Alternator. Objectives
4 pages
Formulating Questions Based On Bloom's Taxonomy 8 Nov 2020
No ratings yet
Formulating Questions Based On Bloom's Taxonomy 8 Nov 2020
38 pages
Experiment No 6 Difference Amplifier
No ratings yet
Experiment No 6 Difference Amplifier
2 pages
Learner - Centered Syllabus Revised On 17 December 2020
No ratings yet
Learner - Centered Syllabus Revised On 17 December 2020
37 pages
Determination of Transformer Equivalent Circuit Parameters of A Single
No ratings yet
Determination of Transformer Equivalent Circuit Parameters of A Single
16 pages
Determination of Transformer Equivalent Circuit Parameters of A Single
0% (1)
Determination of Transformer Equivalent Circuit Parameters of A Single
18 pages
Determination of Transformer Equivalent Circuit Parameters of A Single
No ratings yet
Determination of Transformer Equivalent Circuit Parameters of A Single
18 pages
Determination of Transformer Equivalent Circuit Parameters of A Single
No ratings yet
Determination of Transformer Equivalent Circuit Parameters of A Single
18 pages
Command Line for Beginners – How to Use the Terminal Like a Pro [Full Handbook]
No ratings yet
Command Line for Beginners – How to Use the Terminal Like a Pro [Full Handbook]
87 pages
Apics CSCP Dictionary 14th Edition
No ratings yet
Apics CSCP Dictionary 14th Edition
9 pages
Mega65 Book
No ratings yet
Mega65 Book
1,279 pages
Class 12 Sumita Arora C++ ch10 Linked Lists Stacks and Queues
100% (1)
Class 12 Sumita Arora C++ ch10 Linked Lists Stacks and Queues
24 pages
ACA Project Report Final
No ratings yet
ACA Project Report Final
12 pages
UGRD-ITE6100-2213T Introduction To Computing Prelim Q1
No ratings yet
UGRD-ITE6100-2213T Introduction To Computing Prelim Q1
5 pages
Epicor Tax Connect Configuration Epicor ERP 10
No ratings yet
Epicor Tax Connect Configuration Epicor ERP 10
14 pages
DMS Acknowledgement
No ratings yet
DMS Acknowledgement
3 pages
Py
No ratings yet
Py
10 pages
Modify A Metz 45ct1 Flash To Make It Full Manual
No ratings yet
Modify A Metz 45ct1 Flash To Make It Full Manual
7 pages
Chap 16
No ratings yet
Chap 16
12 pages
Test Bank With Answers of Accounting Information System by Turner Chapter 04
100% (3)
Test Bank With Answers of Accounting Information System by Turner Chapter 04
30 pages
MCSA Assaignment
No ratings yet
MCSA Assaignment
4 pages
Bypassing Software Update Package Encryption - Extracting The Lexmark MC3224i Printer Rmware (Part 1)
No ratings yet
Bypassing Software Update Package Encryption - Extracting The Lexmark MC3224i Printer Rmware (Part 1)
38 pages
Chapter 5, Additional Exercises
0% (1)
Chapter 5, Additional Exercises
7 pages
CPE Black Book
No ratings yet
CPE Black Book
99 pages
AC Project Report
No ratings yet
AC Project Report
31 pages
PHD Coursework Syllabus For Computer Science
100% (2)
PHD Coursework Syllabus For Computer Science
4 pages
XGL C42B
No ratings yet
XGL C42B
144 pages
Talib Report12
No ratings yet
Talib Report12
70 pages
Gartner - Magic Quadrant For Multiexperience Development Platforms-2020Q3
No ratings yet
Gartner - Magic Quadrant For Multiexperience Development Platforms-2020Q3
30 pages
.Protection of Crops From The Rain Water and Generation of Solar Energy With An Embedded System Using GSM Technology
No ratings yet
.Protection of Crops From The Rain Water and Generation of Solar Energy With An Embedded System Using GSM Technology
4 pages
Dell Perc h310 Spec Sheet
No ratings yet
Dell Perc h310 Spec Sheet
2 pages
ECA2 _Tests_Vocabulary Check 4B_2018
No ratings yet
ECA2 _Tests_Vocabulary Check 4B_2018
4 pages
6000 Diagnostic LED Numbers and Codes
No ratings yet
6000 Diagnostic LED Numbers and Codes
15 pages
BATCH- (703)
No ratings yet
BATCH- (703)
2 pages
PPRS Training Manual
No ratings yet
PPRS Training Manual
145 pages
SAP2000 Lab 01-Insruction sheet
No ratings yet
SAP2000 Lab 01-Insruction sheet
3 pages

Object_Detection_in_Aerial_Images_A_Large_Scale_Benchmark_and_Challenges

Uploaded by

Object_Detection_in_Aerial_Images_A_Large_Scale_Benchmark_and_Challenges

Uploaded by

This article has been accepted for publication in a future issue of this journal, but has not been

Object Detection in Aerial Images:

images and then detect objects in the regions. In this paper,

2.4 Code Libraries for Object Detection

To address these problems, we represent the objects with TABLE 3

and GSD. In DOTA-v2.0, the GSDs of the images from GF-2,

DOTA-v1.0 [14] 0.57 0.41 0.02

Fig. 9. Densities of the different categories. The density is measured by

image in DOTA varies more widely than in natural image

104 4.7 DOTA Versions

TABLE 5 (HBB and OBB) in our paper. The HBB is a rectangle

Bp Bgt 5.2 Implementation Details

��������� ����������� ������������������� ������������������� ���������

�� CNN H-OBB by 4 fps. Note that the process of transferring

TABLE 10 6.1.6 Class-Wise Results

You might also like

��