Generating Proposals from Corners in RPN to Detect Bees in Dense
Yassine Kriouile
1,2 a
, Corinne Ancourt
1 b
, Katarzyna Wegrzyn-Wolska
1,2 c
and Lamine Bougueroua
2 d
Mines ParisTech, PSL University, Centre de Recherche en Informatique, 35 rue Saint Honor
e, 77300 Fontainbleau, France
EFREI Paris, AliansTIC, 30/32 Avenue de la R
epublique, 94800 Villejuif, France
{katarzyna.wegrzyn, lamine.bougueroua},
Bees, Object Detection, Faster RCNN, RPN, High Object Density, Corners.
Detecting bees in beekeeping is an important task to help beekeepers in their work, such as counting bees, and
monitoring their health status. Deep learning techniques could be used to perform this automatic detection.
For instance Faster RCNN is a neural network for object detection that is suitable for this kind of tasks. But its
accuracy is degraded when it comes to images of bee frames due to the high density of objects. In this paper,
we propose to extend the RPN sub-neural network of Faster RCNN to improve detection recall. In addition to
detect bees from centers, four branches are added to detect bees from their corners. We constructed a dataset of
images and annotated it. We compared this approach to the standard Faster RCNN. It improves the detection
accuracy. Code is available at
In recent years, bees suffer from many problems like
varroa infestation (Sipos et al., 2021). This prob-
lem could cause a negative impact on the health of
bees. To overcome this kind of issues, beekeepers
should monitor bees and their conditions. Computer
vision methods based on deep learning could help
beekeepers in bee detection. In fact, there are al-
ready advanced neural networks for object detection
like Faster RCNN. This network is detecting objects
in an image through two stages. First RPN generates
proposals then they are classified and enhanced using
another sub-network. RPN takes a set of feature maps
calculated by a backbone neural network. For each
position in the map, it predicts a predefined number of
possible objects. To achieve this prediction, a set of
bounding boxes called anchors are generated at each
feature map position. RPN predicts whether each an-
chor matches an object or not and gives the offset co-
ordinates to correct the anchor position. The anchors
are centered on the position. An anchor corresponds
to an object when its IoU (intersection over union)
with a ground truth is high. This means that anchors
generated at the centers of the objects are more likely
classified as objects, while anchors generated at the
corners of the objects are more likely classified as
non-objects. In dense scenes, there are objects that
are partially visible; only the corners are visible. The
standard approach could miss these kinds of objects.
To overcome this issue, we propose to enhance anchor
generation by generating anchors from object corners
in addition to object centers. Concretely, four addi-
tional types of anchors are used; each one corresponds
to an object corner (top left, top right, bottom left, bot-
tom right). At each feature map position, the corner
anchor is generated so that the feature map point is
located at the anchor corner. We duplicate the RPN
prediction components and losses to handle the cor-
ner data. To test our approach, we constructed and
annotated a specific dataset composed by images of
bee frames. To analyze our approach, we made dif-
ferent modifications on our proposed approach and
tested them on two databases of different density lev-
els. In this paper, our contribution consists on:
An extension of standard Faster RCNN architec-
ture to predict objects from corners and improve
recall in dense scenes.
Kriouile, Y., Ancourt, C., Wegrzyn-Wolska, K. and Bougueroua, L.
Generating Proposals from Corners in RPN to Detect Bees in Dense Scenes.
DOI: 10.5220/0010815000003124
In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 5: VISAPP, pages
ISBN: 978-989-758-555-5; ISSN: 2184-4321
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
A dataset of bee frames annotated with bounding
boxes corresponding to bees.
Our work aims to improve bee detection using deep
learning. In this section, we cite some works that
are related to our subject. First, we focus on exist-
ing methods dealing with bee detection in Section 2.1,
then we present the state of the art of neural networks
for object detection in Sections 2.2 and 2.3. Sec-
tion 2.4 presents the dense scene issue, and the ex-
isting approach dealing with this situation. Anchors
are also an interesting topic as our method is based on
the generation of a new set of anchors, related work
about anchors is described in Section 2.5. Finally we
cite some already proposed approaches based on cor-
ner data to detect objects in Section 2.6.
2.1 Bee Detection
There are works treating the problem of detecting
bees in images. (Magnier et al., 2018) propose an
approach to detect and track bees from videos with
white background. This detection is based on back-
ground subtraction, ellipse approximation, border de-
tection and color threshold, but its FPS (frame per sec-
ond) is low. In (Magnier et al., 2019), they suggest a
method to track bees by estimating their trajectories
through polynomial and interpolation.
(Kulyukin and Mukherjee, 2018) estimate the bee
traffic by detecting movements and classifying im-
ages based on machine learning methods as CNN,
SVM and Random Forest. They published a public
database containing annotated images and videos for
machine learning.
(Tiwari, 2018) tries to recognize bees and track
them from videos using CNN. The recording is man-
aged by BeePi which is a specific hardware dedicated
to beehives and makes it possible to collect other
kinds of data like sound and temperature.
(Tu et al., 2016) aim to track bee behavior and
evaluate the condition of the hive. This is achieved
by counting bees at the beehive entrance and estimat-
ing their in-out activity using linear regression. Their
method does not handle complicated cases such as
high bee density and bee occlusion.
(Kulyukin and Reka, 2016) track the traffic of for-
aging bees using sound and images. After recording
videos thanks to BeePi, tracking is performed by pixel
separation algorithm and contour detection of binary
In (Dembski J., 2020), the authors try to detect
bees on video images using three steps: determine the
regions of interest ROI for each frame using motion
detection, then classify each region whether it con-
tains a bee or not thanks to a convolutional deep neu-
ral network, and finally group regions using clustering
2.2 Two Stage Object Detection
Two stage object detection is an approach of detect-
ing objects using two phases ; generate proposals then
classify them. R-CNN is based on a vision algorithm
generating proposals which are provided to a convolu-
tional network which performs the classification (Gir-
shick et al., 2014). Fast-RCNN improves object de-
tection by using a ROI pooling layer responsible of
extracting features from shared feature maps instead
of calculating feature map for each proposal (Gir-
shick, 2015). Faster-RCNN uses the RPN neural net-
work as proposal generator (Ren et al., 2015). This ar-
chitecture allows to learn predicting object bounding
boxes. R-FCN is a two stage object detection neural
network built only by convolutional layers, the prin-
ciple is to use many feature maps, each of which con-
tains information about different object regions (Dai
et al., 2016).
2.3 One Stage Object Detection
One stage object detectors are faster and less resource
greedy than two stage detectors. SSD and YOLO are
the state of art neural networks which detect objects
without generating proposals (Liu et al., 2016), (Red-
mon et al., 2016). The drawback of these approaches
is the imbalance between positive and negative ex-
amples during training, this leads to limited perfor-
mance. To overcome this issue, RetinaNet was pro-
posed, their authors suggest to use Focal loss which
increases the loss of misclassified object compared to
well classified objects (Lin et al., 2017b).
2.4 Dense Scene
In the field of object detection, a dense scene means
a scene where the density of objects is very high,
in other words it means that the number of objects
per unit area is large
. In this case, there is a high
probability that the image contains two objects that
are adjacent or occluding each other. Performing
accurate detection using machine learning for com-
puter vision in these conditions could be challeng-
ing. There are many works dealing with these issues
1 density
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
in the context of detecting people and pedestrians in
crowded scenes, such as (Liu et al., 2019) where au-
thors present a method which predicts whether two
overlapping bounding boxes match the same object
or two different objects. In (Zhang et al., 2018),
the authors try to improve the detection in pedestrian
crowded scenes by proposing another way of calculat-
ing the loss which aims to improve the compactness
of the predicted bounding boxes around the ground-
truth, and to improve the prediction of the human
body by detecting its parts. The paper (Xi et al., 2020)
deals with the detection of human faces in crowded
scenes with low resolution by exploring the similar-
ity between detected objects. Finally (Zhang et al.,
2019) propose to use two anchors to detect people in
a crowd; one for the head and one for the body.
Concerning object detection in a general context;
the paper (G
ahlert et al., 2020) attempts to improve
object detection in images through using pixel-base
bounding box annotations. (Goldman et al., 2019)
detect objects in a dense scene using a new layer of
prediction called ”Soft-IoU” which predicts IoU of a
predicted object bounding box compared to the real
object position.
2.5 Anchors
In the state of the art of object detection neural net-
works, anchors are used to detect objects. They are
bounding boxes of different sizes and ratios generated
by the neural network. It generates a predefined num-
ber of anchors at each position of the feature map.
For each of them, a score is predicted, it corresponds
to the probability that the anchor contains an object.
The learning phase is based on the comparison of an-
chors with ground-truth bounding boxes. Each anchor
is centered on its corresponding feature map position,
this matching enables the neural network to predict
objects from their centers.
On the other side, there are anchor-free detectors
such as (Tian et al., 2019) which do not use anchors.
In this approach, for each feature map point, object
bounding box coordinates are directly predicted. This
approach allows to predict objects from all their pix-
2.6 Corner Detection
The paper (Qiu et al., 2020) presents an approach
called BorderDet which aims to improve the detec-
tion of objects in dense scenes by using feature maps
of borders in addition to standard feature maps of cen-
ters. This way allows not to miss partially hidden ob-
jects. But the limit of this method is that it needs a
first coarse prediction to extract border points.
(Zhou et al., 2019) suggest to detect the extreme
points and the central point of an object. The final
detection is obtained by grouping these points. The
proposed approach is based on the detection of key
points using heat maps and (Law and Deng, 2018).
This method needs extreme points annotations.
(Wei et al., 2020) use point-set anchor instead of a
rectangular box anchor, this permits to represent ob-
jects more accurately. This approach is mainly inter-
esting in the case of image segmentation and pose es-
timation, where ground-truth annotations are not stan-
dard bounding boxes.
(Duan et al., 2020) propose an anchor-free and
two stage object detector, based on corner propos-
als. These proposals are generated by predicting two
kinds of corner heat maps; top-left and bottom-right.
Corners are extracted from these maps and a fixed
number of proposals are then generated.
In standard Faster RCNN, RPN generates proposals
from object centers. In fact, it takes feature maps as
input, and generates anchors for each point in each
map. An anchor is placed in a way that the point is
located at the anchor center. The network predicts
for each anchor, objectness (a score corresponding
to the probability of the anchor bounding box to be
an object) and offsets to correct the anchor position
and match the real object. The training phase com-
pares each anchor to ground truth bounding boxes,
using IoU metric (Intersection over Union). The an-
chors with big IoU are considered as objects (ground-
truth objectness is equal to 1). The anchors gener-
ated at corner regions are likely to be classified as
non-objects as their IoUs with ground truth bounding
boxes are small. If the image contains objects whose
only visible parts are corners, they are more likely to
be missed by the network. In the case of bee images,
this situation is frequent due to the high density of
bees in the bee frame image. The RPN has a direct
influence on the recall of Faster-RCNN; if an object
is missed during proposal generation, it is almost im-
possible to detect it by Faster RCNN. We think that
this issue can be resolved by generating another kind
of anchors, using another way of placing them on fea-
ture map points. Each one of this new type of anchors
is placed in a way that the point is located at a cor-
ner anchor. This way would enable RPN training to
consider anchors generated at object corner regions
as objects (ground truth objectness is equal to 1) and
avoid missing partially hidden bees. Our work is mo-
Generating Proposals from Corners in RPN to Detect Bees in Dense Scenes
tivated by these elements:
In standard anchor-based neural networks for ob-
ject detection, prediction is based on object center
information, corner information is not taken into
Improving anchor generation could be used in all
anchor-based neural networks like YOLO.
Generating anchors influences on detection recall.
It is important to generate anchors able to take into
account the type of visible region (center or cor-
Bee detection is an interesting task for beekeep-
ing domain. Improve accuracy of this detection
remains an important objective to reach.
Faster-RCNN is a two stage neural network. It uses
RPN as a proposal generator. This component gener-
ates anchors centered on the feature map positions.
This way of placing anchors favors objects whose
central regions are visible. We propose to generate an-
other kind of anchors in order to predict objects from
corners and improve detection recall.
4.1 General Architecture
The proposed neural network is based on Faster
RCNN. Figure 1 shows the structure of the net-
work. Our work focuses on RPN, the purpose of this
sub-network is to generate region proposals which
are then classified and enhanced using another sub-
network. The first step computes feature maps using
a pre-trained classification neural network like VGG
or Resnet. Then objectness and offsets are predicted.
After that the predicted offsets are re-used by anchor
generation mechanism to create proposals which are
ordered using objectness scores. In the standard ap-
proach, proposals are generated from object centers.
Our approach consists on generating proposals from
object corners. To achieve this objective, we modi-
fied the anchor generation mechanism and extended
the neural network to predict corner data. Section 4.2
explains the types of generated anchors. Section 4.3
presents the added corner predictors. Section 4.4 de-
scribes how generated anchors and prediction data are
combined to create proposals. Section 4.5 clarifies the
used method for labeling anchors during training. The
last section (4.6) cites the used losses for learning.
4.2 Anchor Generation
The aim of the standard anchor generation is to detect
objects from their centers. In the case of dense scenes,
where usually objects are partially occluded, the cen-
tral region of the objects could be hidden, therefore
some objects could be missed. Anchors are a set of
bounding box candidates for predicted objects. They
are generated from feature maps. In the case of FPN
(Lin et al., 2017a), more than one feature maps are
used to detect objects of different sizes. We extend
the anchor generation mechanism by adding four an-
chor generators: top left anchors, top right anchors,
bottom left anchors and bottom right anchors. For
each generator, K anchors (in our case K = 3), of dif-
ferent sizes and ratios, are generated at each position
for each feature map. Let s be the dimension of a
feature map stride (the image region corresponding to
the feature map position), w the anchor width, h the
anchor height, and x1, x2, y1 and y2 anchor coordi-
nates in the format X1X2Y 1Y 2 (X1: x-coordinate of
left border, X2: x-coordinate of right border, Y 1: y-
coordinate of top border, Y 2: y-coordinate of bottom
border), in the coordinate system whose origin is the
center of the stride:
In the case of central anchor:
x1 = w/2, x2 = w/2, y1 = h/2, y2 = h/2
In the case of top left anchor:
x1 = s/2, x2 = ws/2, y1 = s/2, y2 = hs/2
In the case of top right anchor:
x1 = w + s/2, x2 = s/2, y1 = s/2, y2 = hs/2
In the case of bottom left anchor:
x1 = s/2, x2 = ws/2, y1 = h +s/2, y2 = s/2
In the case of bottom right anchor:
x1 = w+s/2, x2 = s/2, y1 = h+s/2, y2 = s/2
Figure 2 illustrates the ve types of generated an-
4.3 Objectness and Offset Prediction
Instead of using one layer for predicting objectness
and offsets, four other parallel layers are added to pre-
dict corner data. In total there are five branches with
different weights; (1) the first to generate proposals
from object centers, (2) the second to generate pro-
posals from object top left corners, (3) the third to
generate proposals from object top right corners, (4)
the fourth to generate proposals from object bottom
left corners, (5) the fifth to generate proposals from
object bottom right corners. Each branch is made up
of two layers; the first one takes as input the feature
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
Figure 1: General Architecture of Faster-RCNN based on corner detection. l: number of feature maps, N: number of images,
Hi: height of feature map i, Wi: width of feature map i, C: number of channels, K: number of anchors per feature map
maps generated by the backbone, and applies a convo-
lution operation. The second layer contains two sub-
branches, the first is a convolutional layer to predict
objectness at each feature map position, the second is
a convolutional layer to predict box offset coordinates
to correct the positions of the proposal.
4.4 Generating Proposals
Proposal generation is performed by combining the
prediction data and the generated anchors. We dupli-
cate this step four times to generate proposals from
corner data. It consists in applying predicted offsets
on generated anchors, then a predefined number of
best proposals is preserved from each feature map,
the NMS (non-maximum suppression) algorithm is
applied on each set of feature map proposals to re-
duce multidetection, and finally a predefined number
of best proposals is preserved from all proposals. A
proposal is better than the other when its objectness
score is higher.
The final set of proposals is simply a union of pro-
posals generated by the five generators.
Generating Proposals from Corners in RPN to Detect Bees in Dense Scenes
Figure 2: Types of generated anchors from feature maps.
4.5 Anchor Ground Truth Labeling and
Offset Assigning
The anchor labeling consists on labeling each anchor
with 0 or 1 which corresponds to whether the an-
chor contains an object or not. Offset assigning is
the process of assigning to each anchor the offsets
that must be applied to the anchor to reach the cor-
rect position. The aim of training is, for each anchor,
to predict the objectness and offsets. Our approach
uses standard labeling and assigning; for each anchor,
the ground truth bounding box with which the anchor
has the greatest overlapping score (IoU: Intersection
over Union) is mapped to the anchor, if this IoU is
greater than a fixed threshold (0.7), the anchor is la-
beled with 1, if it is less than a fixed threshold (0.3),
the anchor is labeled with 0, otherwise it is ignored.
The differences between the anchor borders and the
borders of the mapped bounding box are the assigned
ground truth offsets.
4.6 RPN Losses
As demonstrated in figure 3, there are two losses in
RPN; objectness loss and offset loss. To balance be-
tween positive and negative samples, a sampling is
performed. The training consists in optimizing the
sum of these two losses:
Objectness loss:
L =
) + (1 t
)log(1 p
) (1)
Offset loss:
L =
) (2)
Where R, N
, N
, p
, t
, o
are respectively the
smooth L1 loss (Girshick, 2015), the number of an-
chors to classify, the number of anchors to regress,
the predicted objectness, the ground truth label, the
predicted offsets, and the ground truth offsets.
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
For training on different layers, we add similar
losses for the predicted corner information. The fi-
nal loss is an equally weighted sum of the five losses.
In Section 5.1, we describe the dataset sources from
which we retrieve images for training and testing our
approach. In Section 5.2 we present tools used for
image annotation. Our code was developed using a
specific Python library for object detection, and an ef-
ficient hardware. Section 5.3 explains these software
and hardware features. We chose standard metrics to
evaluate our approach, they are cited in Section 5.4.
5.1 Data Source
Our main objective is to improve bee detection us-
ing an object detection neural network. Therefore,
our image database is only made up of images of bee
The images used for training and testing our ap-
proach come from three sources: (1) images down-
loaded from the Internet, (2) photos taken using cam-
era and phone in an apiary, and (3) images provided
by our project partners.
The main preprocessing that we perform on im-
ages is cropping to make annotation task easier. In
fact there were images with many bees, the complete
annotation of such images takes a lot of time and ef-
5.2 Image Annotation
We used two tools to annotate our images. The first
is Imagetagger which is a web application for anno-
tating images with bounding boxes
. This tool en-
abled us to work remotely in collaboration with our
partners. We installed it and configured it to make
it accessible through the Internet. Our project part-
ners used it to help us in annotation task. The second
tool is the Matlab application Matlab Image Labler
Through this tool, we annotated some images of our
database with bounding boxes, then exported the an-
notations to matlab workspace. A matlab code was
developed to export annotations from Matlab format
to Json format.
5.3 Framework and Environment
Our work is based on the detectron framework, its
source code is hosted on this Github repository
. The
pre-built version used in our work is located on this
. Detectron is a Facebook framework based
on Pytorch library, and contains implementations of
state of art neural networks for object detection and
image segmentation, such as Faster RCNN, Mask
RCNN and RetinaNet. We have chosen this frame-
work among others because it contains official imple-
mentations of Faster RCNN which is appropriate to
our case because of its good performance despite of
its slowness. The extendibility and modularity of the
framework enabled us to create and integrate our cus-
tom subcomponents like anchor generators and RPN
The training and tests were executed on a machine
with these characteristics:
32 Go of RAM
4 cores, 8 CPUs
16 Go Nvidia GPU: it was used for training and
5.4 Metrics
To assess the accuracy of our approach, we used con-
ventional object detection metrics: average recall and
average accuracy. Since our approach consists of
modifying RPN, we tested the recall of RPN in ad-
dition to the accuracy of the whole neural network.
In Section 6.1, we describe the created datasets for
training and testing. Then, in Section 6.2 we explain
the library parameters that we set for our use case.
In Section 6.3, we discuss and compare the accuracy
results that we obtained after training and testing the
standard approach and our approach.
6.1 Train and Test Data
We created three image sets; (1) the first set was an-
notated by our partners, it contains arbitrary images
coming from our dataset sources. They were not com-
pletely annotated; highly occluded bees were gen-
erally not annotated, because of the effort required
linux x86 64.whl
Generating Proposals from Corners in RPN to Detect Bees in Dense Scenes
Figure 3: RPN losses. These two losses are duplicated five times to train predictors on center and corner data.
by these annotations. (2) The images of the second
set were purposely selected to obtain a database with
high object density. Moreover, they were almost en-
tirely annotated. (3) The third set contains images
with a lower object density. We took 22 images from
our first set with 1086 annotations. We created two
test sets: (1) the first one contains 11 images with
535 annotations coming from our first image set, (2)
the second contains 25 images with 1691 annotations
coming from the second set of images. By using two
image sets we aimed at comparing our approach un-
der two situations: images with normal density and
images with higher density.
6.2 Parameters
We mainly used the default parameter values set by
the Detectron library except for some parameters. For
training, a pretrained Faster-RCNN model with FPN
architecture was used. This model uses a pretrained
ResNet-50 backbone trained on ImageNet dataset.
The model was trained on COCO dataset. As long as
transfer learning is used, 8000 iterations of training on
bee images was enough to achieve state of art results.
The number of preserved proposals before and after
NMS algorithm were also fixed through two specific
Detectron parameters. We changed the default values
because the relevance of our approach depends on that
6.3 Result Analysis
Table 1 shows the results of the tests for the standard
approach against the proposed approach. Recall and
precision are significantly improved by the corner ap-
proach. The recall evaluates the number of detected
objects compared to ground truth ones. The preci-
sion corresponds to the proportion of correct detec-
tions among all detections.
To compare the standard approach to the corner
approach, the number of RPN output proposals must
be the same. The number of proposals preserved be-
fore NMS is 200. When FPN is used, 200 best pro-
posals are retrieved from each feature map. There are
5 feature maps, it means that 1000 proposals are pre-
served. After NMS, in the standard case, 1000 best
proposals are preserved, it means that all the propos-
als are preserved. In the case of our approach, 200
best proposals are preserved from each center/corner
prediction after NMS, before merging the five sets to
obtain 1000 proposals. These parameter values were
chosen because of the constraint of the inference. In
fact NMS is a greedy algorithm, it requires time and
resources. Providing a big number of proposals ttopo
NMS is not very convenient if quick detection is re-
The results demonstrate that our approach detects
more objects than the standard. This is due to the pre-
dicted corner data which enables the neural network
to detect bees from corners.
To understand our results, and explore our approach
advantages and limits, we have analyzed some aspects
of them. First, we verified the accuracy of the center
predictor alone, this is explained in Section 7.1. In
Section 7.2 we talk about the influence of the num-
ber of the preserved proposals on the neural network
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
Table 1: Results showing accuracy metrics of standard approach and our proposed approach. RPN AR@1000: RPN average
recall on 1000 first detections using 0.5 IoU, Faster RCNN AR: Faster RCNN average recall on 100 first detections using
0.5 IoU, Faster RCNN AR: Faster RCNN average precision on 100 first detections using 0.5:0.95 IoU. Number of proposals
before RPN NMS is 200, number of porposals after RPN NMS is 1000.
Metric RPN AR@1000 Faster RCNN AR Faster RCNN AP
Metric type Recall Recall Precision
Standard Approach
Normal density 76.45 76.45 46.03
High density 56.06 53.87 27.6
Corner Approach
Normal density 86.73 82.8 52.17
High density 66.88 55.82 28.54
Table 2: Results showing accuracy metrics of our approach when some modifications are applied: (1) One loss instead of five
losses, (2) One branch for detecting corner data instead of four branches, (3) Using shifted convolution in corner branches
instead of normal convolution. RPN AR@1000: RPN average recall on 1000 first detections using 0.5 IoU, Faster RCNN
AR: Faster RCNN average recall on 100 first detections using 0.5 IoU, Faster RCNN AR: Faster RCNN average precision on
100 first detections using 0.5:0.95 IoU. Number of proposals before RPN NMS is 200, number of porposals after RPN NMS
is 1000.
Metric RPN AR@1000 Faster RCNN AR Faster RCNN AP
Metric type Recall Recall Precision
One loss
Normal density 85.79 81.31 51.95
High density 65.58 52.27 28.35
One branch
Normal density 82.82 82.61 48.71
High density 64.7 59.43 30.02
Shifted convolution
Normal density 84.67 81.87 51.94
High density 63.87 54.35 29.02
accuracy. In Section 7.3 we focus on the possibility
of merging the proposals before NMS. On the other
hand, we explored using one loss instead of ve losses
and one branch instead of five branches as explained
in Section 7.4 and Section 7.5. In Section 7.6, the
new convolution method that we used to enhance cor-
ner data representation is described. Finally, in Sec-
tion 7.7, we analyze the results that we obtained when
we trained our neural network on bee images with low
7.1 Accuracy of Center Data
When we compared the accuracy of the standard neu-
ral network and the network based on our approach
using only center prediction, the standard gives bet-
ter results. Indeed, it is more difficult for the training
to find backbone parameters that satisfy all the ob-
jective losses. Nevertheless, the results demonstrated
that the proposed approach detects corner data which
compensates the reduced accuracy of center data pre-
7.2 Number of Preserved Proposals
before NMS
We noticed that the accuracy of our approach is not
good when the number of preserved proposals from
each feature map exceeds 200. It remains a limitation
of this approach that have to be addressed in the fu-
ture. But the results show that corner data prediction
is a promising way to detect objects in dense scenes.
7.3 Proposal Concatenation
The proposed approach is based on the concatenation
of the proposals after NMS. Merging these propos-
als before NMS algorithm is another possibility: in
this case, the predicted offsets for each branch (cen-
ter/corner) are applied to the anchors to obtain pro-
posals, these bounding boxes are concatenated in a
set, then they are ordered by their objectness score,
and NMS is applied to reduce multi-detection. The
advantage of this method is that it executes the NMS
algorithm once instead of five times. But its results
were not satisfactory. Furthermore, in this alternative
approach, NMS should be applied to a greater number
of proposals which require a greater use of resources.
7.4 One Loss vs Five Losses
Instead of using one loss for each corner or center,
only one loss could be used. In fact, this can be im-
plemented by combining corner and center predicted
data in a single layer. The five anchor generators are
replaced by one generator which generates the five
types of anchors. The labeling and mapping step is
performed on all the generated anchors. As demon-
Generating Proposals from Corners in RPN to Detect Bees in Dense Scenes
Table 3: Results showing accuracy metrics of standard approach and our proposed approach when training on dataset with
low object density. RPN AR@1000: RPN average recall on 1000 first detections using 0.5 IoU, Faster RCNN AR: Faster
RCNN average recall on 100 first detections using 0.5 IoU, Faster RCNN AR: Faster RCNN average precision on 100 first
detections using 0.5:0.95 IoU.
Metric RPN AR@1000 Faster RCNN AR Faster RCNN AP
Metric type Recall Recall Precision
Standard Approach
Normal density 62.99 45.61 23.01
High density 49.67 21.94 10.81
Corner Approach
Normal density 65.42 55.33 27.69
High density 53.93 31.52 14.27
strated by Table 2 this approach gives lower results
than the one based on five different losses.
7.5 One Branch for Corner Data
A similar approach was also tested. Instead of using
four different branches for the corner data, one branch
is used. This method has the advantage of reducing
the number of parameters to learn. Despite of the con-
straint of representing different types of corners with
the same set of parameters, the approach results are
promising as shown in Table 2.
7.6 Shifted Convolution
Our approach is based on predicting objectness and
offsets from object corners. Whether or not a fea-
ture map region corresponds to a corner depends on
the region content and its surroundings. Regarding
the corners, the prediction is more influenced by the
region containing the object. This region is not en-
circling the corner as it is the case for object center,
but it is shifted. For instance, for a top left corner,
instead of using a surrounding centered on the corner
point, this surrounding should be shifted to right and
bottom. Therefore instead of convolving the points of
standard surrounding, the convolution is made on this
new surrounding which contains more relevant cor-
ner data. The results of this approach are presented
by Table 2.
7.7 Sparse Images as Training Set
To check the relevance of our approach, we trained
the neural network on a simple database made up of
22 images with only 424 annotations. In fact these
images contain sparse bees. The objective was to ver-
ify whether the approach could detect partially visi-
ble objects using only visible parts which are mainly
bee corners. So we used a sparse image set to pre-
vent the network to learn hidden objects by consid-
ering them as small. The results are shown in Table
3. The limits of those results are that the accuracy is
very low compared to the values of state of the art. In
fact it is certainly due to the limited number of annota-
tions used for training. But the advantage of that case,
is that there is no limit to the number of considered
proposals before NMS. More investigation should be
conducted in the future to understand the relation be-
tween the training annotations and the performance of
the approach. Figure 4 shows that our approach de-
tects some occluded bees while the normal approach
could not predict them.
In this paper, we propose a new approach to detect
objects in dense scenes. It is an extension of the stan-
dard Faster RCNN. In addition to predicting whether
the positions of the feature map correspond to the ob-
ject centers, we predict whether they correspond to
the object corners. This is achieved through gener-
ating a new set of anchors and predicting their cor-
responding objectness and offset data. Our approach
aims to improve object detection in the case of bee
frame images. Therefore, we built and annotated a
specific dataset. We executed the training and test-
ing to evaluate our approach. When the number of
the preserved proposals before NMS is less than 200,
our approach performs better than the standard. So it
can be used in situations where computation resources
are limited. This case can occur when there are con-
straints to obtain images. Moreover, our proposed
neural network can be used in other object detection
contexts. Since our approach is only to extend an-
chor generation mechanism, it is not specific to Faster
RCNN, but it can be exploited in other anchor based
neural networks like YOLO. We hope that our ap-
proach can help beekeepers to monitor their beehives
more accurately, and that our work can contribute to
the advancement in computer vision research.
Nevertheless, there are some limitations that must
be addressed in the future. How to improve recall
independently to the number of preserved proposals
should be treated in more depth. On the other hand,
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications
(a) Detected bees in the first image using
standard approach.
(b) Detected bees in the first image using
corner approach.
(c) Detected bees in the second image using
standard approach.
(d) Detected bees in the second image using
corner approach.
(e) Detected bees in the third image using
standard approach.
(f) Detected bees in the third image using
corner approach.
Figure 4: The detected bees in three images using standard and proposed approaches. (a) and (b) relate to the same image. (c)
and (d) relate to the same image. (e) and (f) relate to the same image.
Generating Proposals from Corners in RPN to Detect Bees in Dense Scenes
our approach should be tested in other contexts and
neural networks.
The authors gratefully acknowledge the Ministry of
Agriculture and Food which funds PNAPI through
CASDAR (the special appropriation account Agri-
culture and Rural Development”) under project num-
ber 18 ART 1831 as well as the support and help of
Alexandre Dangl
eant, ITSAP (Technical and Scien-
tific Institute of Beekeeping and Pollination).
Dai, J., Li, Y., He, K., and Sun, J. (2016). R-fcn: Object de-
tection via region-based fully convolutional networks.
In International Conference on Neural Information
Processing Systems. Curran Associates Inc.
Dembski J., S. J. (2020). Weighted clustering for bees de-
tection on video images. In Computational Science.
Duan, K., Xie, L., Qi, H., Bai, S., Huang, Q., and Tia, Q.
(2020). Corner proposal network for anchor-free, two-
stage object detection. In European Conference on
Computer Vision – ECCV.
ahlert, N., Hanselmann, N., Franke, U., and Denzler, J.
(2020). Visibility guided nms: Efficient boosting of
amodal object detection in crowded traffic scenes.
Girshick, R. (2015). Fast r-cnn. In 2015 IEEE International
Conference on Computer Vision (ICCV).
Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2014).
Rich feature hierarchies for accurate object detection
and semantic segmentation.
Goldman, E., Herzig, R., Eisenschtat, A., Goldberger, J.,
and Hassner, T. (2019). Precise detection in densely
packed scenes. In 2019 IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR).
Kulyukin, V. and Mukherjee, S. (2018). On video analysis
of omnidirectional bee traffic: Counting bee motions
with motion detection and image classification. Ap-
plied Sciences.
Kulyukin, V. A. and Reka, S. K. (2016). Toward sustainable
electronic beehive monitoring: Algorithms for omni-
directional bee counting from images and harmonic
analysis of buzzing signals.
Law, H. and Deng, J. (2018). Cornernet: Detecting objects
as paired keypoints. In European Conference on Com-
puter Vision (ECCV).
Lin, T.-Y., Doll
ar, P., Girshick, R., He, K., Hariharan, B.,
and Belongie, S. (2017a). Feature pyramid networks
for object detection. In IEEE Conference on Computer
Vision and Pattern Recognition (CVPR).
Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Doll
ar, P.
(2017b). Focal loss for dense object detection. In
IEEE International Conference on Computer Vision
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu,
C.-Y., and Berg, A. C. (2016). Ssd: Single shot multi-
box detector. Lecture Notes in Computer Science.
Liu, Y., Liu, L., Rezatofighi, H., Do, T.-T., Shi, Q., and
Reid, I. (2019). Learning pairwise relationship for
multi-object detection in crowded scenes.
Magnier, B., Ekszterowicz, G., Laurent, J., Rival, M., and
Pfister, F. (2018). Bee hive traffic monitoring by track-
ing bee flight paths. In Proceedings of the 13th Inter-
national Joint Conference on Computer Vision, Imag-
ing and Computer Graphics Theory and Applications.
SCITEPRESS - Science and Technology Publications.
Magnier, B., Gabbay, E., Bougamale, F., Moradi, B., Pfis-
ter, F., and Slangen, P. R. (2019). Multiple honey bees
tracking and trajectory modeling. In Multimodal Sens-
ing: Technologies and Applications. SPIE.
Qiu, H., Ma, Y., Li, Z., Liu, S., and Sun, J. (2020). Bor-
derdet: Border feature for dense object detection. In
European Conference on Computer Vision – ECCV.
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.
(2016). You only look once: Unified, real-time object
detection. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR).
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 Informa-
tion Processing Systems. Curran Associates, Inc.
Sipos, T., Donk
o, T., J
ak, I., and Keszthelyi, S.
(2021). Study of morphological features in pre-
imaginal honey bee impaired by varroa destructor by
means of computer tomography. Insects.
Tian, Z., Shen, C., Chen, H., and He, T. (2019). Fcos:
Fully convolutional one-stage object detection. In
IEEE/CVF International Conference on Computer Vi-
sion (ICCV).
Tiwari, A. (2018). A deep learning approach to recognizing
bees in video analysis of bee traffic. PhD thesis, Utah
state university.
Tu, G. J., Hansen, M. K., Kryger, P., and Ahrendt, P. (2016).
Automatic behaviour analysis system for honeybees
using computer vision. Computers and Electronics in
Wei, F., Sun, X., Li, H., Wang, J., and Lin, S. (2020). Point-
set anchors for object detection, instance segmenta-
tion and pose estimation. In European Conference on
Computer Vision – ECCV.
Xi, Y., Zheng, J., He, X., Jia, W., Li, H., Xie, Y., Feng, M.,
and Li, X. (2020). Beyond context: Exploring seman-
tic similarity for small object detection in crowded
scenes. Pattern Recognition Letters.
Zhang, K., Xiong, F., Sun, P., Hu, L., Li, B., and Yu, G.
(2019). Double anchor r-cnn for human detection in a
Zhang, S., Wen, L., Bian, X., Lei, Z., and Li, S. Z. (2018).
Occlusion-aware r-cnn: Detecting pedestrians in a
crowd. In Computer Vision ECCV. Springer Inter-
national Publishing.
Zhou, X., Zhuo, J., and Kr
uhl, P. (2019). Bottom-
up object detection by grouping extreme and center
points. In IEEE/CVF Conference on Computer Vision
and Pattern Recognition (CVPR).
VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications