Multi-level Feature Selection for Oriented Object Detection
Chen Jiang
1
, Yefan Jiang
1
, Zhangxing Bian
3
, Fan Yang
2
and Siyu Xia
1
1
School of Automation, Southeast University, Nanjing, China
2
College of Telecommunications and Information Engineering, NJUPT, Nanjing, China
3
Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, U.S.A.
Keywords:
Object Detection, Rotation Detection, Feature Selection, Remote Sensing, Path Aggregation.
Abstract:
Horizontal object detection has made significant progress, but the representation of horizontal bounding box
still has application limitations for oriented objects. In this paper, we propose an end-to-end rotation detector
to localize and classify oriented targets precisely. Firstly, we introduce the path aggregation module, to shorten
the path of feature propagation. To distribute region proposals to the most suitable feature map, we propose
the feature selection module instead of using selection mechanism based on the size of region proposals.
What’s more, for rotation detection, we adopt eight-parameter representation method to parametrize the ori-
ented bounding box and we add a novel loss to handle the boundary problems resulting from the representation
way. Our experiments are evaluated on DOTA and HRSC2016 datasets.
1 INTRODUCTION
Object detection which benefits from the development
of deep learning methods, especially in deep convo-
lution neural networks, has made significant break-
throughs. The progress expands the applications sce-
narios of object detection, such as in security system,
text detection and aerial images.
Current popular detectors can be divided into two
types by different output representations: horizontal
and oriented bounding box. Horizontal bounding box,
which is always represented by (x,y,w,h) (the coor-
dinate of center point, width, height), is difficult to
locate multi-oriented objects. With the ratio of ro-
tated objects enlarging, horizontal bounding boxes
will include more background noise to the detriment
of model training. For example, Fig. 1 shows, when
rotated objects are close-packed, it will bring difficul-
ties for detection and terrible visual experience. To
address the limitation of horizontal detectors, numer-
ous rotated object detection methods have been pro-
posed and achieve considerable progress. However,
rotation detectors are still faced with several prob-
lems.
The rotated objects are usually labeled by five pa-
rameters including an additional parameter θ to orig-
inal horizontal bounding box representation (x,y,w,
h). In the five-parameter representation method from
OpenCV, when the parameter θ reaches its range
(a) Horizontal bounding box (b) Oriented bounding box
Figure 1: Different representation ways on small and clut-
tered object detection. Horizontal bounding box (HBB) in-
cludes extra target while single target is framed by oriented
bounding box (OBB).
boundary, such as 1
◦
and −89
◦
, two bounding boxes
will be very approximate through exchanging width
and height. The loss of these three parameters might
be huge, even though the position of boxes almost
doesn’t change at all. Moreover, IOU is sensitive to
minor angle fluctuation, leading to the decline of ob-
ject detection performance. Some other methods la-
bel rotated boxes through recording the coordinates
of four vertices. However, the coordinates are dis-
ordered and ambious—the same rotated object can be
represented by several sets of values—resulting in ab-
normal loss. Another method (Xu et al., 2020) glides
each vertex of the horizontal bounding box to form
a quadrangle to represent the rotated object, which
also has problems when the object is nearly horizon-
tal. Moreover, it is difficult for rotated bounding box
regression when the prediction of horizontal bound-
36
Jiang, C., Jiang, Y., Bian, Z., Yang, F. and Xia, S.
Multi-level Feature Selection for Oriented Object Detection.
DOI: 10.5220/0010213000360043
In Proceedings of the 10th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2021), pages 36-43
ISBN: 978-989-758-486-2
Copyright
c
 2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

Figure 2: Architecture of our method. Our network consists of four modules: (a) backbone of network for feature extraction
(using ResNet50 and FPN), (b) path aggregation module (PAM) for shortening the information path, (c) feature selection
module (FSM) for selecting the optimal feature map, (d) ROI Align module and branches of regression and classification.
ing box exists deviations.
In this paper, we propose an effective and fast
framework for multi-oriented object detection. We
adopt the gliding vertex method (Xu et al., 2020) to
label the rotated objects and attempt to solve the fun-
damental problems arising from this way. In gliding
vertex method, there are eight parameters, through
adding four gliding offset variables on the basis of
classic horizontal bounding box representation, as is
shown in Fig. 3. The representation way can pre-
vent the sequential problems arising from directly re-
gressing four vertices, because each offset is corre-
sponding to the relative side of horizontal bounding
boxes. However, it shares the same boundary prob-
lem with five-parameter representation method when
the angle and offset variables reach the range bound-
ary. We propose a new modulated loss in view of this
situation to ensure that the prediction result can be
obtained most simply and directly. In addition, the
final prediction quadrangle depends on the accuracy
of horizontal bounding box regression. We design an
adaptive feature selection module to improve the re-
gression results of horizontal bounding boxes. Re-
gion of interest generated from region proposal net-
work (Ren et al., 2015) will be distributed to best fea-
ture layer through feature selection module. We also
add bottom-up path augmentation module to preserve
low-level localization signals, which is beneficial to
following regression and classification. In summary,
the main contributions of this paper include:
• We design a novel modulated loss function to
solve the boundary problems based on gliding ver-
tex method, to improve the accuracy of prediction
results when the detection object is nearly hori-
zontal.
• We propose a novel FSM to flexibly distribute
multi scale ROI to the most suitable feature map
and add PAM to our model, increasing the accu-
racy of the horizontal bounding box prediction re-
sults.
• We propose an effective multi-oriented object de-
tection network, and reach substantial gains on
DOTA and HRSC2016.
2 RELATED WORK
2.1 Horizontal Region Object Detection
Since (Girshick et al., 2014) proposed R-CNN, clas-
sic object detection has made considerable break-
through. Based on this seminal work, Fast R-CNN
(Girshick, 2015), Faster R-CNN (Ren et al., 2015)
and R-FCN (Dai et al., 2016) are proposed subse-
quently, which improve the accuracy and efficiency
of detection. Faster R-CNN includes object detection
network and region proposal network (RPN), is the
representative two-stage method. On the other hand,
one-stage detectors, which predict bounding boxes di-
rectly from feature maps, are proposed to improve the
speed for the simple architecture, such as SSD (Liu
et al., 2016), YOLO (Redmon et al., 2016) and Reti-
naNet (Lin et al., 2017b). In particular, YOLO se-
ries have shown great performance through optimiza-
tion of several versions. RetinaNet along with focal
loss function is presented to handle class imbalance of
samples. (Lin et al., 2017a) considers the scale vari-
ance in images and proposes Feature Pyramid Net-
work (FPN) to address the problems of multi-scale
objects. To get rid of the disadvantages of anchor-
based networks, anchor-free detectors become the re-
search focus in recent years. CenterNet (Duan et al.,
2019), FCOS (Tian et al., 2019), and ExtremeNet
(Zhou et al., 2019) are prototypical one-stage detec-
tors. (Cai and Vasconcelos, 2018) introduce the idea
of cascade and propose a multi-stage detector called
Cascade R-CNN, that achieve high performance in
both localization and classification. However, the
above detectors merely generate horizontal bounding
boxes, and still have application limitations in several
Multi-level Feature Selection for Oriented Object Detection
37

Figure 3: Representation of oriented bounding box. We
adopt eight parameters (x,y,w,h,α
1
,α
2
,α
3
,α
4
) to label the
oriented object, where (x,y,w,h) represents the horizontal
bounding box.
real-world scenarios, especially in aerial images and
multi-oriented scene texts.
2.2 Oriented Region Object Detection
Since horizontal detectors are not suitable for detec-
tion tasks in aerial images and scene texts, more de-
tectors for rotated objects spring up. In scene text
detection, RRPN (Ma et al., 2018) is improved in
the framework of Faster R-CNN and propose rotated
region proposal network (RRPN). TextBox++ (Liao
et al., 2018a) adopts quadrilateral prediction based
on SSD. RRD (Liao et al., 2018b) decouples clas-
sification and bounding box regression on rotation-
invariant and rotation sensitive features to further im-
prove TextBox++. R
2
CNN (Jiang et al., 2017) also
generates rotated bounding boxes to perform fast and
effective text detection.
For object detection in aerial images, the com-
plexity of the remote sensing background, the multi
scales of samples and the huge number of dense,
cluttered and rotated objects are extremely difficult,
which calls for robust detectors. RoI Transformer
(Ding et al., 2019) extracts rotated region of interest
to locate and classify. SCRDet (Yang et al., 2019b)
combines multi-dimensional attention network and
refined sampling network and achieve state-of-the-art
performance. R3Det (Yang et al., 2019a) proposes a
refined single-stage detector with feature refinement
to solve the feature misalignment problem. Glid-
ing Vertex (Xu et al., 2020) and RSDet (Qian et al.,
2019) reach SOTA performance on DOTA datasets by
quadrilateral regression.
(a) Convolutional network (b) Feature pyramid network
Figure 4: Illustrations of two feature extraction network
structures. (a) use the top feature map for fast prediction.
(b) use feature pyramid network for more accurate predic-
tion.
2.3 Multi-level Features
Features from different layers include distinct seman-
tic information, which are useful for multi-scale ob-
ject detection. SharpMask (Pinheiro et al., 2016),
LRR (Ghiasi et al., 2016) and (Peng et al., 2017) use
feature fusion to obtain more details. FCN (Long
et al., 2015) and U-Net (Ronneberger et al., 2015)
fused features from lower layers by skip-connections.
FPN (Lin et al., 2017a) introduce a top-down path and
combine semantic features from top layers and high-
resolution information from lower layers for segmen-
tation. PANet (Liu et al., 2018) firstly augments an
additional down-top path in the framework of FPN
and achieves better prediction. ASFF (Liu et al.,
2019), NAS-FPN (Ghiasi et al., 2019) and BiFPN
(Tan et al., 2020) adopt complex two-path integration
for further improvement.
3 PROPOSED METHOD
3.1 Overview
The architecture of our network is shown in Fig. 2.
It includes four modules: (a) the backbone of our
network. We adopt FPN in the framework. (b) the
path aggregation module (PAM), which is proposed
by PANet(Liu et al., 2018) to improve the results of
multi-scale feature extraction. (c) the feature selec-
tion module (FSM), which is designed to distribute
ROI to the optimal feature map. (d) the full connec-
tion layer. We first introduce PAM in Sec. 3.2. Next,
the detail of FSM will be explained in Sec. 3.3. Fi-
nally, we will introduce the boundary problem of glid-
ing vertex representation method and propose a novel
loss to handle it.
3.2 Path Aggregation Module
CNN usually applied the structure of Fig. 4a in the
past, which used the final feature map for prediction.
ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods
38

Figure 5: Illustration of feature selection module. Region
proposals from RPN will be mapped to all feature maps and
the optimal level is selected through calculating the layer
with minimum IOU-loss.
However, pooling and other operations reduce the res-
olution of the convolution feature map, which cannot
meet the needs of small target detection, resulting in
low positioning accuracy and missing detection. As
is shown in Fig. 4b, FPN (Lin et al., 2017a), which
adds a top-down path, is proposed to ameliorate the
problem. The up-bottom path transfers deep seman-
tic information to shallow layers combined with low-
level features, which enhances the robustness of the
network to objects with different scales.
Meanwhile, the low-level features of the target,
such as the edge, play a crucial role in the position-
ing. In order to enhance the results of localization, we
need to make full use of low-level local feature infor-
mation with high spatial resolution. PAM (Fig. 2(b))
introduces an additional bottom-up path aggregation
path, which greatly shortens the propagation path of
local feature information. Since the network aggre-
gates the propagation path, it can better extract the
local feature information, such as texture and edge of
the target, and semantic feature information, and im-
prove the ability of the network to detect multi-scale
targets in remote sensing images. Ablation study is
given in Sec. 4.4.
3.3 Feature Selection Module
After obtaining the multi-scale region proposals gen-
erated from RPN(Ren et al., 2015), the problem to be
solved is how to allocate the ROI of different scales
to the corresponding feature map. The common dis-
tribution method is that ROI will be mapped to differ-
ent layers according to the size. Assuming the size of
ROI is w ×h, according to Formula (1), it would be
mapped to feature map P
k
:
k =
"
k
0
+ log
2
√
wh
224
#
(1)
k
0
corresponds to the feature layer of the ROI with
224×224 area. It illustrates that the ROI with smaller
size will be mapped to the feature map with higher
spatial resolution, and the larger ROI will be mapped
to the feature map with lower resolution. However,
there are limitations of this simple method, which
might not be the optimal plan. For example, two re-
gion proposals with a difference of about 10 pixels in
size may be assigned to different feature layers under
this method, but in fact the two regions may be very
similar. Feature maps from PAM combine low-level
feature information with high-level feature informa-
tion, and it is vital to select the most suitable fea-
ture map, which is beneficial to final regression and
classification. The architecture of FSM is illustrated
in Fig. 5. Firstly, each region proposal generated
from RPN (the grey regions) will be mapped to differ-
ent feature levels. Next, the ground truth (horizontal
bounding box) will be mapped to these feature maps
and calculate IOU-Loss individually on each feature
map. Finally, we will compare all loss values and
choose the minimum to decide the optimal level to be
pooled through ROIAlign. Better than proposal size,
the smallest IOU-Loss value includes the most feature
information. Through FSM, we transform the simple
feature selection mechanism to IOU-based adaptive
method. Ablation study is given in Sec. 4.4.
3.4 Loss Functions
We adopt a simple representation for rotated object,
which is intuitively displayed in Fig. 3. The black
horizontal bounding box B
h
denoted by (v
0
1
,v
0
2
,v
0
3
,v
0
4
),
has four sliding vertices on each edge (v
1
,v
2
,v
3
,v
4
),
which construct a quadrilateral B
r
to represent the ro-
tated object, the orange one in the figure. (v
1
,v
2
,v
3
,v
4
)
are corresponding to the top, right, bottom and left
side of B
h
. The horizontal bounding box is repre-
sented by (x,y,w,h), where (x,y) is the center, w is
width and h is height. The oriented bounding box
is denoted by (x,y,w,h,α
1
,α
2
,α
3
,α
4
), where the extra
variables are defined as follows:
α
1
=
t
1
w
α
2
=
t
2
h
(2)
α
3
=
t
3
w
α
4
=
t
4
h
(3)
where t
i
= ||v
0
1
−v
1
||, i ∈
{
1, 2, 3, 4
}
represents the dis-
tance between v
0
1
and v
1
.
However, both α
i
=0 or 1 can represent horizontal
bounding box, which is confused for bounding box
regression. For example, the ground truth is horizon-
tal, whose α
i
are set to 1, and the predicted offset α
i
are all nearly 0. Two regions are highly coincident but
the loss would be far more than 0. In addition, when
objects are similar to horizontal, it might be simpler
to regress to near 0 rather than 1. Clearly, the final
prediction result can be obtained in the simpler and
more direct way when the target is nearly horizontal.
Multi-level Feature Selection for Oriented Object Detection
39

Figure 6: Example results of our method. The top row is from DOTA and the bottom row is from HRSC2016.
We propose a simple and novel loss function to
handle this problem. When all the α
0
i
of ground truth
are nearly 0 or 1, we take it as the horizontal object
and decide which regression direction to choose ac-
cording to the smaller loss value. The regression loss
function is defined as follows:
L
reg
= L
h
+ L
r
(4)
L
r
=



4
∑
i=0
smooth
L1
(α
i
,
e
α
i
), l
1
≤
e
α
i
≤ l
2
L
min
else
(5)
L
min
= min







4
∑
i=0
smooth
L1
(α
i
, 0)
4
∑
i=0
smooth
L1
(α
i
, 1)
(6)
where L
h
is the regression loss function for horizon-
tal box, the same as that in Faster R-CNN, and l
1
, l
2
are the threshold to determine whether to be taken as
horizontal boxes. The total loss function is given by
L =
1
N
cls
∑
i
L
cls
+
λ
N
reg
∑
i
p
∗
i
L
reg
(7)
where N
cls
, N
reg
indicate the number of mini-batch
size and positive targets in ground truth respectively,
i denotes the index of a proposal in a mini-batch. p
∗
i
is a binary value (p
∗
i
= 0 for background and p
∗
i
= 1
for foreground). The hyper-parameter λ controls the
trade-off, which is set to be 1 by default. The loss of
RPN follows the Faster R-CNN (Ren et al., 2015).
4 EXPERIMENTS
We evaluate our proposed method on the challenge
datasets DOTA and HRSC2016 for object detection
in remote sensing, which both include a mass of
arbitrary-oriented objects. The ablation study is con-
ducted on HRSC2016, which is full of multi-oriented
ships with different scales. Some qualitative results
on HRSC2016 and DOTA are shown in Fig. 6
4.1 Datasets
DOTA is one of the largest and most challenging
datasets in aerial image detection with quadrangle an-
notations. DOTA contains 2,806 aerial images from
different sensors and platforms including 15 object
categories with 188,182 instances, the size of which
ranges from around 800 ×800 to 4, 000 ×4, 000 pix-
els. It is split into training, validation and testing sets,
accounting for 1/2, 1/6, 1/3 of the whole data set,
respectively. The categories are: Plane (PL), Swim-
ming pool (SP), Baseball diamond (BD), Ground field
track (GTF), Large vehicle (LV), Ship (SH), Tennis
court (TC), Soccer-ball field (SBF), Basketball court
(BC), Storage tank (ST), Bridge (BR), Roundabout
(RA), Harbor (HA), Small vehicle (SV) and Heli-
copter (HC).
HRSC2016 is a dataset in ship detection with large
range of aspect ratio and wide variety of orientations,
which contains 1061 images with 29 categories. The
size of each image in HRSC2016 is various, ranging
ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods
40

Table 1: Comparisons with other methods on DOTA.
Methods PL BD BR GTF SV LV SH TC BC ST SBF RA HA SP HC mAP
FR-O(Xia et al., 2018) 79.09 69.12 17.17 63.49 34.2 37.16 36.2 89.19 69.6 58.96 49.4 52.52 46.69 44.8 46.3 52.93
R-DFPN(Yang et al., 2018) 80.92 65.82 33.77 58.94 55.77 50.94 54.78 90.33 66.34 68.66 48.73 51.76 55.1 51.32 35.88 57.94
R
2
CNN(Jiang et al., 2017) 80.94 65.67 35.34 67.44 59.92 50.91 55.81 90.67 66.92 72.39 55.06 52.23 55.14 53.35 48.22 60.67
RRPN(Ma et al., 2018) 88.52 71.2 31.66 59.3 51.85 56.19 57.25 90.81 72.84 67.38 56.69 52.84 53.08 51.94 53.58 61.01
ICN(Azimi et al., 2018) 81.4 74.3 47.7 70.3 64.9 67.8 70 90.8 79.1 78.2 53.6 62.9 67 64.2 50.2 68.2
RADet(Li et al., 2020) 79.45 76.99 48.05 65.83 65.46 74.4 68.86 89.7 78.14 74.97 49.92 64.63 66.14 71.58 62.16 69.09
RoI-Transformer(Ding et al., 2019) 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
Ours 89.2 86.5 46.7 72.2 69.5 68 76.2 90.5 78.6 84.4 58.5 69.9 67.4 70.6 53.2 72.1
from 300×300 to 500 ×500. The training, validation
and testing sets contain 436, 181 and 444 images.
4.2 Implementation Details
Network Setting. The proposed method is imple-
mented on the framework of “maskrcnn benchmark”
with Ubuntu 16.04, NVIDIA GTX 1080, and 8G
Memory. We adopt ResNet50 as backbone and the
batch size is set to 2 because of the limited memory.
Stochastic gradient descent (SGD) is used in all ex-
periments with weight decay and momentum set to
0.0001 and 0.9, respectively. The base learning rate is
set to 0.0025 and is divided by a factor of 10 at each
decay step. The threshold l
1
, l
2
in Formula (5) is set
to 0.1 and 0.9, respectively.
Dataset Setting. For DOTA, the model is trained for
80k iterations and the learning rate decays by 10 after
54k and 64k steps from an initial value of 2.5e
−4
to
2.5e
−6
. We use random flipping and random rotate
from (0,90,180,270) degree for data augmentation in
training. For HRSC2016, the model is trained for 40k
iterations and the learning rate decays at 28k. Multi
scale testing is applied with (0.5,0.8,1.0).
4.3 Comparisons with the
State-of-the-Art Methods
We compare our proposed method with the state-of-
the-art algorithms on DOTA and HRSC2016. The
compared results of DOTA are depicted in Table 1.
The results reported are achieved from the official
DOTA evaluation server. The compared method in-
clude R
2
CNN and RRPN, the methods for scene text
detection, along with RoI-transformer and ICN, meth-
ods for aerial image detection. The results show that,
our method obtains 72.1% mAP of OBB task and
2.54% better than the state-of-the-art method, RoI-
Transformer. The compared results of HRSC2016 are
shown in Table 2. Some qualitative results are se-
lected for comparison, and we obtain 93.16% mAP
in this dataset with ResNet50 as backbone.
4.4 Ablation Study
Effects of Path Aggregation Module. We use the
ResNet50 as backbone and regress (x,y,w,h,α
1
,α
2
,α
3
,
α
4
) with smooth − l1 loss for baseline. It shows
that by adding PAM, it can improve performance by
1.04% mAP on the HRSC2016 in Table 3. PAM aims
to shorten the propagation path of local feature in-
formation and achieve better results than independent
FPN. It also reveals that PAM works on oriented ob-
ject detection.
Effects of Feature Selection Module. The purpose
of feature selection module is to help region propos-
als to select the most suitable feature map for further
regression and classification, which substitutes the se-
lection mechanism based on the size of ROI. The per-
formance is improved from 89.88% mAP to 90.21%
mAP in Table 3, which increases by 0.33% and is
not significant. The feature maps from FPN include
less information than those from PAM, which leads
to the limited effect of FSM. When we combine PAM
with FSM, the result increases from 89.88% mAP to
91.43% mAP, which increases by 1.55% and is more
than the sum of two modules improvement.
Effects of L
min
Loss. The compared results in Ta-
ble 3 shows the performance of smooth −l1 loss and
the performance by adding our L
min
loss in the fifth
row. The L
min
loss is proposed to detect nearly hori-
zontal objects more effectively and directly. The per-
formance increases by 0.99% through adding this loss
function and improves more than 2% by adding three
modules.
5 CONCLUSIONS
In this paper, we present an effective framework for
oriented and multi-scale object detection. We intro-
duce the PAM to shorten the feature propagation path
from low level to high level and enhance the local-
ization capability. To allocate ROI of different scales
to the corresponding feature map, we propose FSM
instead of the allocation scheme anchored in the size
Multi-level Feature Selection for Oriented Object Detection
41

Table 2: Comparisons with other methods on HRSC2016.
Method Backbone mAP
R
2
CNN(Jiang et al., 2017) ResNet101 79.73
RRPN(Ma et al., 2018) ResNet101 85.64
RetinaNet-H(Yang et al., 2019a) ResNet101 89.27
DRN(Pan et al., 2020) Hourglass104 92.70
Ours ResNet50 93.16
Table 3: Ablation study on HRSC2016.
FPN PAM FSM L
min
AP
X 89.88
X X 90.92
X X 90.21
X X 90.87
X X X 91.43
X X X X 91.91
of region proposals. In addition, we adopt the eight-
parameter method to represent the oriented object and
propose a novel modulated loss function to address
the problem of the representation way when the tar-
get is nearly horizontal. We conduct extensive exper-
iments to illustrate the achievements of our method
across two datasets DOTA and HRSC2016 in compar-
ison with several qualitative approaches and the effect
of each module is shown by ablation study.
REFERENCES
Azimi, S. M., Vig, E., Bahmanyar, R., K
¨
orner, M., and
Reinartz, P. (2018). Towards multi-class object de-
tection in unconstrained remote sensing imagery. In
Asian Conference on Computer Vision, pages 150–
165. Springer.
Cai, Z. and Vasconcelos, N. (2018). Cascade r-cnn: Delving
into high quality object detection. In Proceedings of
the IEEE conference on computer vision and pattern
recognition, pages 6154–6162.
Dai, J., Li, Y., He, K., and Sun, J. (2016). R-fcn: Object de-
tection via region-based fully convolutional networks.
In Advances in neural information processing sys-
tems, pages 379–387.
Ding, J., Xue, N., Long, Y., Xia, G.-S., and Lu, Q. (2019).
Learning roi transformer for oriented object detection
in aerial images. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition,
pages 2849–2858.
Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., and Tian, Q.
(2019). Centernet: Keypoint triplets for object detec-
tion. In Proceedings of the IEEE International Con-
ference on Computer Vision, pages 6569–6578.
Ghiasi, G., Fowlkes, C. C., et al. (2016). Laplacian recon-
struction and refinement for semantic segmentation.
arXiv preprint arXiv:1605.02264, 4(4).
Ghiasi, G., Lin, T.-Y., and Le, Q. V. (2019). Nas-fpn: Learn-
ing scalable feature pyramid architecture for object
detection. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 7036–
7045.
Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE
international conference on computer vision, pages
1440–1448.
Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2014).
Rich feature hierarchies for accurate object detec-
tion and semantic segmentation. In Proceedings of
the IEEE conference on computer vision and pattern
recognition, pages 580–587.
Jiang, Y., Zhu, X., Wang, X., Yang, S., Li, W., Wang, H.,
Fu, P., and Luo, Z. (2017). R2cnn: rotational region
cnn for orientation robust scene text detection. arXiv
preprint arXiv:1706.09579.
Li, Y., Huang, Q., Pei, X., Jiao, L., and Shang, R. (2020).
Radet: Refine feature pyramid network and multi-
layer attention network for arbitrary-oriented object
detection of remote sensing images. Remote Sensing,
12(3):389.
Liao, M., Shi, B., and Bai, X. (2018a). Textboxes++: A
single-shot oriented scene text detector. IEEE trans-
actions on image processing, 27(8):3676–3690.
Liao, M., Zhu, Z., Shi, B., Xia, G.-s., and Bai, X. (2018b).
Rotation-sensitive regression for oriented scene text
detection. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 5909–
5918.
Lin, T.-Y., Doll
´
ar, P., Girshick, R., He, K., Hariharan, B.,
and Belongie, S. (2017a). Feature pyramid networks
for object detection. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition,
pages 2117–2125.
Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Doll
´
ar, P.
(2017b). Focal loss for dense object detection. In
Proceedings of the IEEE international conference on
computer vision, pages 2980–2988.
Liu, S., Huang, D., and Wang, Y. (2019). Learning spatial
fusion for single-shot object detection. arXiv preprint
arXiv:1911.09516.
Liu, S., Qi, L., Qin, H., Shi, J., and Jia, J. (2018). Path ag-
gregation network for instance segmentation. In Pro-
ceedings of the IEEE conference on computer vision
and pattern recognition, pages 8759–8768.
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.,
Fu, C.-Y., and Berg, A. C. (2016). Ssd: Single shot
multibox detector. In European conference on com-
puter vision, pages 21–37. Springer.
ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods
42

Long, J., Shelhamer, E., and Darrell, T. (2015). Fully con-
volutional networks for semantic segmentation. In
Proceedings of the IEEE conference on computer vi-
sion and pattern recognition, pages 3431–3440.
Ma, J., Shao, W., Ye, H., Wang, L., Wang, H., Zheng, Y.,
and Xue, X. (2018). Arbitrary-oriented scene text de-
tection via rotation proposals. IEEE Transactions on
Multimedia, 20(11):3111–3122.
Pan, X., Ren, Y., Sheng, K., Dong, W., Yuan, H., Guo,
X., Ma, C., and Xu, C. (2020). Dynamic refinement
network for oriented and densely packed object de-
tection. In Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition, pages
11207–11216.
Peng, C., Zhang, X., Yu, G., Luo, G., and Sun, J. (2017).
Large kernel matters–improve semantic segmentation
by global convolutional network. In Proceedings of
the IEEE conference on computer vision and pattern
recognition, pages 4353–4361.
Pinheiro, P. O., Lin, T.-Y., Collobert, R., and Doll
´
ar, P.
(2016). Learning to refine object segments. In Eu-
ropean conference on computer vision, pages 75–91.
Springer.
Qian, W., Yang, X., Peng, S., Guo, Y., and Yan, C. (2019).
Learning modulated loss for rotated object detection.
arXiv preprint arXiv:1911.08299.
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.
(2016). You only look once: Unified, real-time object
detection. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 779–
788.
Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster
r-cnn: Towards real-time object detection with region
proposal networks. In Advances in neural information
processing systems, pages 91–99.
Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:
Convolutional networks for biomedical image seg-
mentation. In International Conference on Medical
image computing and computer-assisted intervention,
pages 234–241. Springer.
Tan, M., Pang, R., and Le, Q. V. (2020). Efficientdet: Scal-
able and efficient object detection. In Proceedings of
the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 10781–10790.
Tian, Z., Shen, C., Chen, H., and He, T. (2019). Fcos: Fully
convolutional one-stage object detection. In Proceed-
ings of the IEEE international conference on com-
puter vision, pages 9627–9636.
Xia, G.-S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J.,
Datcu, M., Pelillo, M., and Zhang, L. (2018). Dota:
A large-scale dataset for object detection in aerial im-
ages. In Proceedings of the IEEE Conference on Com-
puter Vision and Pattern Recognition, pages 3974–
3983.
Xu, Y., Fu, M., Wang, Q., Wang, Y., Chen, K., Xia, G.-S.,
and Bai, X. (2020). Gliding vertex on the horizon-
tal bounding box for multi-oriented object detection.
IEEE Transactions on Pattern Analysis and Machine
Intelligence.
Yang, X., Liu, Q., Yan, J., Li, A., Zhang, Z., and Yu, G.
(2019a). R3det: Refined single-stage detector with
feature refinement for rotating object. arXiv preprint
arXiv:1908.05612.
Yang, X., Sun, H., Fu, K., Yang, J., Sun, X., Yan, M., and
Guo, Z. (2018). Automatic ship detection in remote
sensing images from google earth of complex scenes
based on multiscale rotation dense feature pyramid
networks. Remote Sensing, 10(1):132.
Yang, X., Yang, J., Yan, J., Zhang, Y., Zhang, T., Guo, Z.,
Sun, X., and Fu, K. (2019b). Scrdet: Towards more ro-
bust detection for small, cluttered and rotated objects.
In Proceedings of the IEEE International Conference
on Computer Vision, pages 8232–8241.
Zhou, X., Zhuo, J., and Krahenbuhl, P. (2019). Bottom-
up object detection by grouping extreme and center
points. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 850–
859.
Multi-level Feature Selection for Oriented Object Detection
43