DOTA: A Large-scale Dataset for Object Detection in Aerial Images∗

Gui-Song Xia1†, Xiang Bai2†, Jian Ding1 , Zhen Zhu2 , Serge Belongie3 ,
Jiebo Luo4 , Mihai Datcu5 , Marcello Pelillo6 , Liangpei Zhang1
arXiv:1711.10398v3 [cs.CV] 19 May 2019

1
State Key Lab. LIESMARS, Wuhan University, China
{guisong.xia, jding, zlp62}@whu.edu.cn
2
EIS, Huazhong Univ. Sci. and Tech., China
{xbai, zzhu}@hust.edu.cn
3
Computer Science Depart., Cornell University, USA
[email protected]
4
Computer Science Depart., Rochester University, USA
[email protected]
5
German Aerospace Center (DLR), Germany
[email protected]
6
DAIS, University of Venice, Italy
[email protected]

May 21, 2019

Abstract
Object detection is an important and challenging problem in computer vision. Although the
past decade has witnessed major advances in object detection in natural scenes, such successes
have been slow to aerial imagery, not only because of the huge variation in the scale, orientation
and shape of the object instances on the earth’s surface, but also due to the scarcity of well-
annotated datasets of objects in aerial scenes. To advance object detection research in Earth
Vision, also known as Earth Observation and Remote Sensing, we introduce a large-scale Dataset
for Object deTection in Aerial images (DOTA). To this end, we collect 2806 aerial images from
different sensors and platforms. Each image is of the size about 4000 × 4000 pixels and contains
objects exhibiting a wide variety of scales, orientations, and shapes. These DOTA images are
then annotated by experts in aerial image interpretation using 15 common object categories. The
fully annotated DOTA images contains 188, 282 instances, each of which is labeled by an arbitrary
(8 d.o.f.) quadrilateral. To build a baseline for object detection in Earth Vision, we evaluate
state-of-the-art object detection algorithms on DOTA. Experiments demonstrate that DOTA well
represents real Earth Vision applications and are quite challenging.
∗
DOTA dataset is available at http://captain.whu.edu.cn/DOTAweb or https://captain-whu.github.io/DOTA.
†
Equal contributions

1
1 Introduction
Object detection in Earth Vision refers to localizing objects of interest (e.g., vehicles, airplanes) on the
earth’s surface and predicting their categories. In contrast to conventional object detection datasets,
where objects are generally oriented upward due to gravity, the object instances in aerial images often
appear with arbitrary orientations, as illustrated in Fig. 1, depending on the perspective of the Earth
Vision platforms.
Extensive studies have been devoted to object detection in aerial images [3, 14, 17–19, 21, 23, 30,
31, 37], drawing upon recent advances in Computer Vision and accounting for the high demands of
Earth Vision applications. Most of these methods [3, 18, 31, 37] attempt to transfer object detection
algorithms developed for natural scenes to the aerial image domain. Recently, driven by the successes
of deep learning-based algorithms for object detection, Earth Vision researchers have pursued ap-
proaches based on fine-tuning networks pre-trained on large-scale image datasets (e.g., ImageNet [6]
and MSCOCO [13]) for detection in the aerial domain, see e.g. [2, 3, 18, 29].
While such fine-tuning based approaches are a reasonable avenue to explore, images such as Fig. 1
reveals that the task of object detection in aerial images is distinguished from the conventional object
detection task in the following respects:
- The scale variations of object instances in aerial images are huge. This is not only because of the
spatial resolutions of sensors, but also due to the size variations inside the same object category.

- Many small object instances are crowded in aerial images, for example, the ships in a harbor
and the vehicles in a parking lot, as illustrated in Fig. 1. Moreover, the frequencies of instances
in aerial images are unbalanced, for example, some small-size (e.g. 1k × 1k) images contain
1900 instances, while some large-size images (e.g. 4k × 4k) may contain only a handfull of small
instances.
- Objects in aerial images often appear in arbitrary orientations. There are also some instances
with an extremely large aspect ratio, such as a bridge.
Besides these distinct difficulties, the studies of object detection in Earth Vision are also challenged
by the well-known dataset bias problem [28], i.e.the degree of generalizability across datasets is often
low. In order to alleviate such biases, the dataset should be annotated to reflect the demands of real
world applications.
Therefore, it is not surprising that the object detectors learned from natural images are not suitable
for aerial images. However, existing annotated datasets for object detection in aerial images, such as
UCAS-AOD [39] and NWPU VHR-10 [2], tend to use images in ideal conditions (clear backgrounds
and without densely distributed instances), which cannot adequately reflect the problem complexity.
To advance the object detection research in Earth Vision, this paper introduces a large-scale
Dataset for Object deTection in Aerial images (DOTA). We collect 2806 aerial images from different
sensors and platforms with crowdsourcing. Each image is of the size about 4000 × 4000 pixels and
contains objects of different scales, orientations and shapes. These DOTA images are annotated
by experts in aerial image interpretation, with respect to 15 common object categories. The fully
annotated DOTA dataset contains 188,282 instances, each of which is labeled by an oriented bounding
box, instead of an axis-aligned one, as is typically used for object annotation in natural scenes. The
main contributions of this work are:

2
 (a)

(b) (c) (d)

car ship harbor roundabout
(e)
#Instances

#pixels
(f)
#Instances

radians

Figure 1: An example taken from DOTA. (a) Typical image in DOTA consisting of many instances
across multiple categories. (b) Illustration of the variety in instance orientation and size. (c),(d)
Illustration of sparse instances and crowded instances, respectively. Here we show four out of fifteen
of the possible categories in DOTA. Examples shown in (b),(c),(d) are cropped from source image
(a). The histograms (e),(f) exhibit the distribution of instances with respect to size and orientation
in DOTA.

3
 - To our knowledge, DOTA is the largest annotated object dataset with a wide variety of categories
in Earth Vision. It can be used to develop and evaluate object detectors in aerial images. We
will continue to update DOTA, to grow in size and scope and to reflect evolving real world
conditions.
- We also benchmark state-of-the-art object detection algorithms on DOTA, which can be used
as the baseline for future algorithm development.
In addition to advancing object detection studies in Earth Vision, DOTA will also pose interesting
algorithmic questions to conventional object detection in computer vision.

2 Motivations
Datasets have played an important role in data-driven research in recent years [6, 13, 34, 36, 38]. Large
datasets like MSCOCO [13] are instrumental in promoting object detection and image captioning
research. When it comes to the classification task and scene recognition task, the same is true for
ImageNet [6] and Places [38], respectively.
However, in aerial object detection, a dataset resembling MSCOCO and ImageNet both in terms of
image number and detailed annotations has been missing, which becomes one of the main obstacles to
the research in Earth Vision, especially for developing deep learning-based algorithms. Aerial object
detection is extremely helpful for vehicle counting, remote object tracking and unmanned driving.
Therefore, a large-scale and challenging aerial object detection benchmark, being as close as possible
to real-world applications, is imperative for promoting research in this field.
We argue that a good aerial image dataset should possess four properties, namely, 1) a large
number of images, 2) many instances per categories, 3) properly oriented object annotation, and
4) many different classes of objects, which make it approach to real-world applications. However,
existing aerial image datasets [15, 17, 24, 39] share in common several shortcomings: insufficient data
and classes, lack of detailed annotations, as well as low image resolution. Moreover, their complexity
is inadequate to be considered as a reflection of the real world.

Dataset Annotation way #main categories #Instances #Images Image width

NWPU VHR-10 [2] horizontal BB 10 3651 800 ∼1000
SZTAKI-INRIA [1] oriented BB 1 665 9 ∼800
TAS [9] horizontal BB 1 1319 30 792
COWC [20] one dot 1 32716 53 2000∼19,000
VEDAI [24] oriented BB 3 2950 1268 512, 1024
UCAS-AOD [39] oriented BB 2 14,596 1510 ∼1000
HRSC2016 [17] oriented BB 1 2976 1061 ∼1100
3K Vehicle Detection [15] oriented BB 2 14,235 20 5616
DOTA oriented BB 14 188,282 2806 800∼4000
Table 1: Comparison among DOTA and object detection datasets in aerial images. BB is short
for bounding box. One-dot refers to annotations with only the center coordinates of an instance
provided. Fine-grained categories are not taken into account. For example, DOTA consist of 15
different categories but only 14 main categories, because small vehicle and large vehicle are both
sub-categories of vehicle.
Datasets like TAS [9], VEDAI [24], COWC [20] and DLR 3K Munich Vehicle [15] only focus on

4
vehicles. UCAS-AOD [39] contains vehicles and planes while HRSC2016 [17] only contains ships even
though fine-grained category information are given. All these datasets are short in the number of
classes, which restricts their applicabilities to complicated scenes. In contrast, NWPU VHR-10 [2] is
composed of ten different classes of objects while its total number of instances is only around 3000.
Detailed comparisons of these existing datasets are shown in Tab. 1. Compared to these aerial datasets,
as we shall see in Section 4, DOTA is challenging for its tremendous object instances, arbitrary but
well-distributed orientations, various categories and complicated aerial scenes. Moreover, scenes in
DOTA is in coincidence with natural scenes, so DOTA is more helpful for real-world applications.
When it comes to general objects datasets, ImageNet and MSCOCO are favored by researchers due
to the large number of images, many categories and detailed annotations. ImageNet has the largest
number of images among all object detection datasets. However, the average number of instances per
image is far smaller than MSCOCO and our DOTA, plus the limitations of its clean backgrounds and
carefully selected scenes. Images in DOTA contain an extremely large number of object instances,
some of which have more than 1,000 instances. PASCAL VOC Dataset [7] is similar to ImageNet in
instances per image and scenes but the inadequate number of images makes it unsuitable to handle
most detection needs. Our DOTA resembles MSCOCO in terms of the instance numbers and scene
types, but DOTA’s categories are not as many as MSCOCO because objects which can be seen clearly
in aerial images are quite limited.
Besides, what makes DOTA unique among the above mentioned large-scale general object detection
benchmarks is that the objects in DOTA are annotated with properly oriented bounding boxes (OBB
for short). OBB can better enclose the objects and differentiate crowded objects from each other.
The benefits of annotating objects in aerial images with OBB are further described in Section 3. We
draw a comparison among DOTA, PASCAL VOC, ImageNet and MSCOCO to show the differences
in Tab. 2.
Image BBox Avg. BBox
Dataset Category
quantity quantity quantity
PASCAL VOC
20 21,503 62,199 2.89
(07++12)
MSCOCO
80 123,287 886,266 7.19
(2014 trainval)
ImageNet
200 349,319 478,806 1.37
(2017train)
DOTA 15 2,806 188,282 67.10

Table 2: Comparison among DOTA and other general object detection datasets. BBox is short for
bounding boxes, Avg. BBox quantity indicates average bounding box quantity per image. Note that
for the average number of instances per image, DOTA surpasses other datasets hugely.

3 Annotation of DOTA
3.1 Images collection
As mentioned in [5], in aerial images, the resolution and variety of sensors being used are factors
to produce dataset biases. To eliminate the biases, images in our dataset are collected from multiple
sensors and platforms (e.g. Google Earth) with multiple resolutions. To increase the diversity of data,

5
we collect images shot in multiple cities carefully chosen by experts in aerial image interpretation. We
record the exact geographical coordinates of the location and capture time of each image to ensure
there are no duplicate images.

3.2 Category selection

Fifteen categories are chosen and annotated in our DOTA dataset, including plane, ship, storage tank,
baseball diamond, tennis court, basketball court, ground track field, harbor, bridge, large vehicle, small
vehicle, helicopter, roundabout, soccer ball field and basketball court.
The categories are selected by experts in aerial image interpretation according to whether a kind
of objects is common and its value for real-world applications. The first 10 categories are common
in the existing datasets, e.g., [2, 15, 20, 39], We keep them all except that we further split vehicle
into large ones and small ones because there is obvious difference between these two sub-categories in
aerial images. Others are added mainly from the values in real applications. For example, we select
helicopter considering that moving objects are of significant importance in aerial images. Roundabout
is chosen because it plays an important role in roadway analysis.
It is worth discussing whether to take “stuff” categories into account. There are usually no clear
definitions for the ”stuff” categories (e.g. harbor, airport, parking lot), as is shown in the SUN dataset
[32]. However, the context information provided by them may be helpful for detection. We only adopt
the harbor category because its border is relatively easy to define and there are abundant harbor
instances in our image sources. The final extended category is soccer field.
In Fig.2, we compare the categories of DOTA with NWPU VHR-10 [2], which has the largest
number of categories in previous aerial object detection datasets. Note that DOTA surpass NWPU
VHR-10 not only in category numbers, but also the number of instances per category.

3.3 Annotation method

We consider different ways of annotating. In computer vision, many visual concepts such as region
descriptions, objects, attributes, and relationships, are annotated with bounding boxes, as shown in
[12]. A common description of bounding boxes is (xc , yc , w, h), where (xc , yc ) is the center location,
w, h are the width and height of the bounding box, respectively.
Objects without many orientations can be adequately annotated with this method. However,
bounding boxes labeled in this way cannot accurately or compactly outline oriented instances such as
text and objects in aerial images. In an extreme but actually common condition as shown in Fig. 3 (c)
and (d), the overlap between two bounding boxes is so large that state-of-the-art object detection
methods cannot differentiate them. In order to remedy this, we need to find an annotation method
suitable for oriented objects.
An option for annotating oriented objects is θ-based oriented bounding box which is adopted in
some text detection benchmarks [35], namely (xc , yc , w, h, θ), where θ denotes the angle from the
horizontal direction of the standard bounding box. A flaw of this method is the inability to compactly
enclose oriented objects with large deformation among different parts. Considering the complicated
scenes and various orientations of objects in aerial images, we need to abandon this method and choose
a more flexible and easy-to-understand way. An alternative is arbitrary quadrilateral bounding boxes,
which can be denoted as {(xi , yi ), i = 1, 2, 3, 4}, where (xi , yi ) denotes the positions of the oriented
bounding boxes’ vertices in the image. The vertices are arranged in a clockwise order. This way is

6
 soccer ball field
helicopter
swimming pool
roundabout
large vehicle
small vehicle
bridge
harbor
ground track field
basketball court
tennis court
baseball diamond
storage tank
ship
plane
100 1000 10000 100000

NWPU VHR 10 DOTA

Figure 2: Comparison between DOTA and NWPU VHR-10 in categories and responding quantity of
instances.

widely adopted in oriented scene text detection benchmarks [11]. We draw inspiration from these
researches and use arbitrary quadrilateral bounding boxes to annotate objects.
To make a more detailed annotation, as illustrated Fig. 3, we emphasize the importance of the
first point (x1 , y1 ), which normally implies the “head” of the object. For helicopter, large vehicle,
small vehicle, harbor, baseball diamond, ship and plane, we carefully denote their first point to enrich
potential usages. While for soccer-ball field, swimming pool, bridge, ground track field, basketball
court and tennis court, there are no visual clues to decide the first point, so we normally choose the
top-left point as the starting point.
Some samples of annotated patches (not the whole original image) in our dataset are shown in
Fig. 4.
It is worth noticing that, Papadopoulos et al. [22] have explored an alternative annotation method

7
 (a) (b) (c) (d)
Figure 3: Visualization of adopted annotation method. The yellow point represents the starting point,
which refers to: (a) top left corner of a plane, (b) the center of sector-shaped baseball diamond, (c)
top left corner of a large vehicle. (d) is a failure case of the horizontal rectangle annotation, which
brings high overlap compared to (c).

and verify its efficiency and robustness. We assume that the annotations would be more precise
and robust with more elaborately designed annotation methods, and alternative annotation protocols
would facilitate more efficient crowd-sourced image annotations.

3.4 Dataset splits

In order to ensure that the training data and test data distributions approximately match, we randomly
select half of the original images as the training set, 1/6 as validation set, and 1/3 as the testing set.
We will publicly provide all the original images with ground truth for training set and validation set,
but not for the testing set. For testing, we are currently building an evaluation server.

4 Properties of DOTA
4.1 Image size
Aerial images are usually very large in size compared to those in natural images dataset. The original
size of images in our dataset ranges from about 800 × 800 to about 4000 × 4000 while most images
in regular datasets (e.g. PASCAL-VOC and MSCOCO) are no more than 1000 × 1000. We make
annotations on the original full image without partitioning it into pieces to avoid the cases where a
single instance is partitioned into different pieces.

4.2 Various orientations of instances

As shown in Fig.1 (f), our dataset achieves a good balance in the instances of different directions,
which is significantly helpful for learning a robust detector. Moreover, our dataset is closer to real
scenes because it is common to see objects in all kinds of orientations in the real world.

4.3 Spatial resolution information

We also provide the spatial resolution for each image in our dataset, which implies the actual size of
an instance and plays a significant role in aerial object detection. The importance of spatial resolution
for detection task are two folds. First, it allows the model to be more adaptive and robust for varieties

8
Figure 4: Samples of annotated images in DOTA. We show three samples per each category, except
six for large-vehicle.

(a) (b) (c)

Figure 5: Statistics of instances in DOTA. AR denotes the aspect ratio. (a) The AR of horizontal
bounding box. (b) The AR of oriented bounding box. (c) Histogram of number of annotated instances
per image.

of objects of the same category. It’s known that objects appear smaller when seen from a distance.

9
The same object with different sizes will trouble the model and hurt classification. However, a model
can pay more attention to the shape with resolution information provided instead of objects’ size.
Second, it’s better for fine-grained classification. For example, it will be simple to distinguish a small
boat from a large warship.
Spatial resolution can also be used to filter mislabeled outliers in our dataset because intra-class
varieties of actual sizes for most categories are limited. Outliers can be found by selecting the objects
whose size is far different from those of the same category in a small range of spatial resolution.

4.4 Various pixel size of categories

Following the convention in [33], we refer to the height of a horizontal bounding box, which we call
pixel size for short, as a measurement for instance size. We divide all the instances in our dataset
into three splits according to their height of horizontal bounding box: small for range from 10 to 50,
middle for range from 50 to 300, and large for range above 300. Tab. 3 illustrates the percentages of
three instance splits in different datasets. It is clear that the PASCAL VOC dataset, NWPU VHR-10
dataset and DLR 3K Munich Vehicle dataset are dominated by middle instances, middle instances and
small instances, respectively. However, we achieve a good balance between small instances and middle
instances, which is more similar to real-world scenes and thus, helpful to better capture different size
of objects in practical applications.
It’s worth noting that pixel size varies in different categories. For example, a vehicle may be as
small as 30, however, a bridge can be as large as 1200, which is 40 times larger than a vehicle. The
huge differences among instances from different categories make the detection task more challenging
because models have to be flexible enough to handle extremely tiny and huge objects.

Dataset 10-50 pixel 50-300 pixel above 300 pixel

PASCAL VOC 0.14 0.61 0.25
MSCOCO 0.43 0.49 0.08
NWPU VHR-10 0.15 0.83 0.02
DLR 3K Munich Vehicle 0.93 0.07 0
DOTA 0.57 0.41 0.02
Table 3: Comparison of instance size distribution of some datasets in aerial images and natural images.

4.5 Various aspect ratio of instances

Aspect ratio is an essential factor for anchor-based models, such as Faster RCNN [26] and YOLOv2 [25].
We count two kinds of aspect ratio for all the instances in our dataset to provide a reference for bet-
ter model design: 1) Aspect ratio of minimally circumscribed horizontal rectangle bounding box, 2)
Aspect ratio of original quadrangle bounding box. Fig. 5 illustrates these two types of distribution
of aspect ratio for instances in our dataset. We can see that instances varies greatly in aspect ratio.
Moreover, there are a large number of instances with a large aspect ratio in our dataset.

4.6 Various instance density of images

It is common for aerial images to contain thousands of instances, which is different from natural
images. For example, images in ImageNet [6] contain on the average 2 categories and 2 instances,
while MSCOCO contains 3.5 categories and 7.7 instances, respectively. Our dataset is much richer in

10
instances per image, which can be up to 2000. Fig. 5 illustrates the number of instances in our DOTA
dataset.
With so many instances in a single image, it is unavoidable to see areas densely crowded with
instances. For COCO, instances are not annotated one by one because occlusion makes it difficult to
distinguish an instance from its neighboring instances. In these cases, the group of instances is marked
as one segment with attribute named “crowd”. However, this is not the case for aerial images because
there are rarely occlusion due to the perspective from the above. Therefore, we can annotate all the
instances in a dense area one by one. Fig. 4 shows examples of densely packed instances. Detecting
objects in these cases poses an enormous challenge for the current detection methods.

5 Evaluations
We evaluate the state of the art object detection methods on DOTA. For horizontal object detection,
we carefully select Faster R-CNN1 [26], R-FCN2 [4], YOLOv23 [25] and SSD2 [16] as our benchmark
testing algorithms for their excellent performance on general object detection. For oriented object
detection, we modify the original Faster R-CNN algorithm such that it can predict properly oriented
bounding boxes denoted as {(xi , yi ), i = 1, 2, 3, 4}.
Note that, the backbone networks are ResNet-101 [8] for R-FCN and Faster R-CNN, Incep-
tionV2 [10] for SSD and customized GoogLeNet [27] for YOLOv2, respectively.

5.1 Tasks
To comprehensively evaluate the state of the art deep learning based detection methods on DOTA,
we propose two tasks, namely detection on horizontal bounding boxes (HBB for short) and detection
on oriented bounding boxes (OBB for short). To be more specific, we evaluate those methods on two
different kinds of ground truths, HBB or OBB, no matter how those methods were trained.

5.2 Evaluation prototypes

Images in DOTA are so large that they cannot be directly sent to CNN-based detectors. Therefore, we
crop a series of 1024 × 1024 patches from the original images with a stride set to 512. Note that some
complete objects may be cut into two parts during the cropping process. For convenience, we denote
the area of the original object as Ao , and the area of divided parts Pi , (i = 1, 2) as ai , (i = 1, 2). Then
we compute the parts areas over the original object area:
ai
Ui = .
Ao
Finally, we label the part Pi with Ui < 0.7 as difficult and for the other one, we keep it the same as
the original annotation. For the vertices of the newly generated parts, we need to ensure they can be
described as an oriented bounding box with 4 vertices in the clockwise order with a fitting method.
In the testing phase, first we send the cropped image patches to obtain temporary results and then
we combine the results together to restore the detecting results on the original image. Finally, we
1
https://github.com/msracver/Deformable-ConvNets
2
https://github.com/tensorflow/models/tree/master/research/object_detection
3
https://github.com/pjreddie/darknet

11
 YOLOv2 [25] R-FCN [4] FR-H [26] SSD [16]
Plane 76.9 81.01 80.32 57.85
BD 33.87 58.96 77.55 32.79
Bridge 22.73 31.64 32.86 16.14
GTF 34.88 58.97 68.13 18.67
SV 38.73 49.77 53.66 0.05
LV 32.02 45.04 52.49 36.93
Ship 52.37 49.29 50.04 24.74
TC 61.65 68.99 90.41 81.16
BC 48.54 52.07 75.05 25.1
ST 33.91 67.42 59.59 47.47
SBF 29.27 41.83 57 11.22
RA 36.83 51.44 49.81 31.53
Harbor 36.44 45.15 61.69 14.12
SP 38.26 53.3 56.46 9.09
HC 11.61 33.89 41.85 0
Avg. 39.2 52.58 60.46 29.86

Table 4: Numerical results (AP) of baseline models evaluated with HBB ground truths. 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-H means Faster R-CNN [26] trained
on Horizontal bounding boxes. FR-O means Faster R-CNN [26] trained on Oriented bounding boxes.
YOLOv2 [25] R-FCN [4] SSD [16] FR-H [26] FR-O
Plane 52.75 39.57 41.06 49.74 79.42
BD 24.24 46.13 24.31 64.22 77.13
Bridge 10.6 3.03 4.55 9.38 17.7
GTF 35.5 38.46 17.1 56.66 64.05
SV 14.36 9.1 15.93 19.18 35.3
LV 2.41 3.66 7.72 14.17 38.02
Ship 7.37 7.45 13.21 9.51 37.16
TC 51.79 41.97 39.96 61.61 89.41
BC 43.98 50.43 12.05 65.47 69.64
ST 31.35 66.98 46.88 57.52 59.28
SBF 22.3 40.34 9.09 51.36 50.3
RA 36.68 51.28 30.82 49.41 52.91
Harbor 14.61 11.14 1.36 20.8 47.89
SP 22.55 35.59 3.5 45.84 47.4
HC 11.89 17.45 0 24.38 46.3
Avg. 25.492 30.84 17.84 39.95 54.13

Table 5: Numerical results (AP) of baseline models evaluated with OBB ground truths. The short
names are defined the same as depicted in Tab. 4. Note that only FR-O [26] is trained with OBB.

use non-maximum suppression (NMS) on these results based on the predicted classes. We keep the
threshold of NMS as 0.3 for the HBB experiments and 0.1 for the oriented experiments. In this way,
we indirectly train and test CNN-based models on DOTA.
For evaluation metrics, we adopt the same mAP calculation as for PASCAL VOC.

5.3 Baselines with horizontal bounding boxes

Ground truths for HBB experiments are generated by calculating the axis-aligned bounding boxes
over original annotated bounding boxes. To make it fair, we keep all the experiments’ settings and
hyper parameters the same as depicted in corresponding papers [4, 16, 25, 26].
The experimental results of HBB prediction are shown in Tab. 4. Note that results of SSD is

12
much lower than other models. We suspect it should be attributed to the random crop operation in
SSD’s data augmentation strategies, which is quite useful in general object detection while degrades
in aerial object detection for tremendous small training instances. The results further indicate the
huge differences between aerial and general objects with respect to instance sizes.

5.4 Baselines with oriented bounding boxes

Prediction of OBB is difficult because the state of the art detection methods are not designed for
oriented objects. Therefore, we choose Faster R-CNN as the base framework for its accuracy and
efficiency and then modify it to predict oriented bounding boxes.
RoIs (Region of Interests) generated by RPN (Region Proposal Network) are rectangles which can
be written as R = (xmin , ymin , xmax , ymax ), for a more detailed interpretation, R = {(xi , yi ), i = 1, 2, 3, 4},
where x1 = x4 = xmin , x2 = x3 = xmax , y1 = y2 = ymin , y3 = y4 = ymax . In R-CNN procedure, each
RoI is attached to a ground truth oriented bounding box written as G = {(gxi , gyi ), i = 1, 2, 3, 4}.
Then R-CNN’s output target T = {(txi , tyi ), i = 1, 2, 3, 4} is calculated by following equations,

txi = (gxi − xi )/w, (1)

tyi = (gyi − yi )/h, (2)

where i = 1, 2, 3, 4, w = xmax − xmin , and h = ymax − ymin .

Other settings and hyper parameters are kept the same as depicted in Faster R-CNN [26]. The
numerical results are shown in Tab. 5. To make a comparison to our implemented Faster R-CNN for
OBB, we evaluate YOLOv2, R-FCN, SSD and Faster R-CNN trained on HBB with the OBB ground
truth. As shown in Tab.5, the results of those methods trained on HBB are much lower than Faster
R-CNN trained on OBB, indicating that for oriented object detection in aerial scenes, those methods
should be adjusted accordingly.

5.5 Experimental analysis

When analyzing the results exhibited in Table. 4, performances in categories like small vehicle, large
vehicle and ship are far from satisfactory, which attributes to their small size and densely crowded
locations in aerial images. As a contrast, large and discrete objects, like planes, swimming pools and
tennis courts, the performances are rather fair.
In Fig. 6, we compare the results between object detection experiments of HBB and OBB. For
densely packed and oriented objects shown in Fig. 6 (a) and (b), location precision of objects in HBB
experiments are much lower than OBB experiments and many results are suppressed through post-
progress operations. So OBB regression is the correct way for oriented object detection that can be
really integrated to real applications. In Fig. 6 (c), large aspect ratio objects annotated in OBB style
like (harbor, bridge) are hard for current detectors to regress. But in HBB style, those objects usually
have normal aspect ratios and as a consequence, results seem to be fairly good as shown in Fig. 6 (d).
However in extremely dense scenes, e.g in Fig. 6 (e) and (f), results of HBB and OBB are all not
satisfying which implies the defects of current detectors.

13
 (a) (c) (e)

(b) (d) (f)

Figure 6: Visualization results of testing on DOTA using well-trained Faster R-CNN. TOP and
Bottom respectively illustrate the results for HBB and OBB in cases of orientation, large aspect
ratio, and density.

6 Cross-dataset validations
The cross dataset generalization [28] is an evaluation for the generalization ability of a dataset. We
choose the UCAS-AOD dataset [39] to do cross-dataset generalization for its comparatively large
number of data comparing to other aerial object detection datasets. For there are no official data
splits for UCAS-AOD, we randomly select 1110 for training and 400 for testing. We choose YOLOv2
as the testing detector for all experiments described below and HBB-style annotations for all ground
truths. Input image size is changed to 960 × 544 around the original image sizes in UCAS-AOD while
other setting kept unchanged.
Results are shown in Tab. 6. The performance difference across two datasets is 35.8 for YOLOv2-A
and 15.6 for YOLOv2-D models, respectively. It suggests that DOTA hugely covers UCAS-AOD and
furthermore has more patterns and properties that are not shared in UCAS-AOD. And both models
get a low results on DOTA which reflects that DOTA is much more challenging.

7 Conclusion
We build a large-scale dataset for oriented objects detection in aerial images which is much larger
than any existing datasets in this field. In contrast to general object detection benchmarks, we

14
 Testing set Detector Plane Small-vehicle Avg.
YOLOv2-A 90.66 88.17 89.41
UCAS-AOD
YOLOv2-D 87.18 65.13 76.15
YOLOv2-A 62.92 44.17 53.55
DOTA
YOLOv2-D 74.83 46.18 60.51
Table 6: Results of cross-dataset generalization. Top: Detection performance evaluated on UCAS-
AOD. Bottom: Detection performance evaluated on DOTA. YOLOv2-A and YOLOv2-D are
trained with UCAS-AOD and DOTA, respectively.

annotate a huge number of well-distributed oriented objects with oriented bounding boxes. We assume
this dataset is challenging but very similar to natural aerial scenes, which are more appropriate for
practical applications. We also establish a benchmark for object detection in aerial images and show
the feasibility to produce oriented bounding boxes by modifying a mainstream detection algorithm.
Detecting densely packed small instances and extremely large instances with arbitrary orientations
in a large picture would be particularly meaningful and challenging. We believe DOTA will not only
promote the development of object detection algorithms in Earth Vision, but also pose interesting
algorithmic questions to general object detection in computer vision.

8 Acknowledgement
We thank Fan Hu, Pu Jin, Xinyi Tong, Xuan Hu, Zhipeng Dong, Liang Wu, Jun Tang, Linyan Cui,
Duoyou Zhou, Tengteng Huang, and all the others who involved in the annotations of DOTA.

References
[1] C. Benedek, X. Descombes, and J. Zerubia. Building development monitoring in multitemporal remotely
sensed image pairs with stochastic birth-death dynamics. IEEE TPAMI, 34(1):33–50, 2012.
[2] G. Cheng, P. Zhou, and J. Han. Learning rotation-invariant convolutional neural networks for object
detection in VHR optical remote sensing images. IEEE Trans. Geosci. Remote Sens., 54(12):7405–7415,
2016.
[3] G. Cheng, P. Zhou, and J. Han. Rifd-cnn: Rotation-invariant and fisher discriminative convolutional neural
networks for object detection. In CVPR, pages 2884–2893, 2016.
[4] J. Dai, Y. Li, K. He, and J. Sun. R-FCN: object detection via region-based fully convolutional networks.
In NIPS, pages 379–387, 2016.
[5] A.-M. de Oca, R. Bahmanyar, N. Nistor, and M. Datcu. Earth observation image semantic bias: A
collaborative user annotation approach. IEEE J. of Selected Topics in Applied Earth Observations and
Remote Sensing, 2017.
[6] J. Deng, W. Dong, R. Socher, L. Li, K. Li, and F. Li. Imagenet: A large-scale hierarchical image database.
In CVPR, pages 248–255, 2009.
[7] M. Everingham, L. V. Gool, C. Williams, J. Winn, and A. Zisserman. The pascal visual object classes
(VOC) challenge. IJCV, 88(2):303–338, 2010.
[8] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, June 2016.
[9] G. Heitz and D. Koller. Learning spatial context: Using stuff to find things. In ECCV, pages 30–43, 2008.
[10] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal
covariate shift. CoRR, abs/1502.03167, 2015.

15
[11] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. K. Ghosh, A. D. Bagdanov, M. Iwamura, J. Matas,
L. Neumann, V. R. Chandrasekhar, S. Lu, F. Shafait, S. Uchida, and E. Valveny. ICDAR 2015 competition
on robust reading. In Proc. ICDAR, 2015.
[12] R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz, S. Chen, Y. Kalantidis, L.-J. Li, D. A.
Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annota-
tions. IJCV, 123(1):32–73, 2017.
[13] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft
COCO: common objects in context. In ECCV, pages 740–755, 2014.
[14] Y. Lin, H. He, Z. Yin, and F. Chen. Rotation-invariant object detection in remote sensing images based
on radial-gradient angle. IEEE Geosci.Remote Sensing Lett., 12(4):746–750, 2015.
[15] K. Liu and G. Máttyus. Fast multiclass vehicle detection on aerial images. IEEE Geosci. Remote Sensing
Lett., 12(9):1938–1942, 2015.
[16] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed, C. Fu, and A. C. Berg. SSD: single shot multibox
detector. In ECCV, pages 21–37, 2016.
[17] Z. Liu, H. Wang, L. Weng, and Y. Yang. Ship rotated bounding box space for ship extraction from
high-resolution optical satellite images with complex backgrounds. IEEE Geosci. Remote Sensing Lett.,
13(8):1074–1078, 2016.
[18] Y. Long, Y. Gong, Z. Xiao, and Q. Liu. Accurate object localization in remote sensing images based on
convolutional neural networks. IEEE Trans. Geosci. Remote Sens., 55(5):2486–2498, 2017.
[19] T. Moranduzzo and F. Melgani. Detecting cars in uav images with a catalog-based approach. IEEE Trans.
Geosci. Remote Sens., 52(10):6356–6367, 2014.
[20] T. N. Mundhenk, G. Konjevod, W. A. Sakla, and K. Boakye. A large contextual dataset for classification,
detection and counting of cars with deep learning. In ECCV, pages 785–800, 2016.
[21] A. Ö. Ok, Ç. Senaras, and B. Yüksel. Automated detection of arbitrarily shaped buildings in complex
environments from monocular VHR optical satellite imagery. IEEE Trans. Geosci. and Remote Sens.,
51(3-2):1701–1717, 2013.
[22] D. P. Papadopoulos, J. R. R. Uijlings, F. Keller, and V. Ferrari. Extreme clicking for efficient object
annotation. CoRR, abs/1708.02750, 2017.
[23] J. Porway, Q. Wang, and S. C. Zhu. A hierarchical and contextual model for aerial image parsing. IJCV,
88(2):254–283, 2010.
[24] S. Razakarivony and F. Jurie. Vehicle detection in aerial imagery: A small target detection benchmark. J
Vis. Commun. Image R., 34:187–203, 2016.
[25] J. Redmon and A. Farhadi. YOLO9000: better, faster, stronger. CoRR, abs/1612.08242, 2016.
[26] S. Ren, K. He, R. B. Girshick, and J. Sun. Faster R-CNN: towards real-time object detection with region
proposal networks. IEEE TPAMI, 39(6):1137–1149, 2017.
[27] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabi-
novich. Going deeper with convolutions. In CVPR, pages 1–9, 2015.
[28] A. Torralba and A. A. Efros. Unbiased look at dataset bias. In CVPR, pages 1521–1528, 2011.
[29] M. Vakalopoulou, K. Karantzalos, N. Komodakis, and N. Paragios. Building detection in very high reso-
lution multispectral data with deep learning features. In IGARSS, pages 1873–1876, 2015.
[30] L. Wan, L. Zheng, H. Huo, and T. Fang. Affine invariant description and large-margin dimensionality
reduction for target detection in optical remote sensing images. IEEE Geosci. Remote Sensing Lett., 2017.
[31] G. Wang, X. Wang, B. Fan, and C. Pan. Feature extraction by rotation-invariant matrix representation
for object detection in aerial image. IEEE Geosci.Remote Sensing Lett., 2017.
[32] J. Xiao, J. Hays, K. Ehinger, A. Oliva, and A. Torralba. SUN database: Large-scale scene recognition from
abbey to zoo. In CVPR, pages 3485–3492, 2010.

16
[33] S. Yang, P. Luo, C. C. Loy, and X. Tang. WIDER FACE: A face detection benchmark. In CVPR, pages
5525–5533, 2016.
[34] B. Yao, X. Yang, and S.-C. Zhu. Introduction to a large-scale general purpose ground truth database:
Methodology, annotation tool and benchmarks. In EMMCVPR 2007, pages 169–183, 2007.
[35] C. Yao, X. Bai, W. Liu, Y. Ma, and Z. Tu. Detecting texts of arbitrary orientations in natural images. In
CVPR, 2012.
[36] Q. You, J. Luo, H. Jin, and J. Yang. Building a large scale dataset for image emotion recognition: The
fine print and the benchmark. In AAAI, pages 308–314, 2016.
[37] F. Zhang, B. Du, L. Zhang, and M. Xu. Weakly supervised learning based on coupled convolutional neural
networks for aircraft detection. IEEE Trans. Geosci. Remote Sens., 54(9):5553–5563, 2016.
[38] B. Zhou, À. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. Learning deep features for scene recognition
using places database. In NIPS, pages 487–495, 2014.
[39] H. Zhu, X. Chen, W. Dai, K. Fu, Q. Ye, and J. Jiao. Orientation robust object detection in aerial images
using deep convolutional neural network. In ICIP, pages 3735–3739, 2015.

17