See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/366029268

Real -Time Object Detection in Video's using YOLO v3

Article · December 2022

CITATIONS READS
0 249

1 author:

Tripti Sharma
Maharaja Surajmal Institute Of Technology
31 PUBLICATIONS 348 CITATIONS

SEE PROFILE

All content following this page was uploaded by Tripti Sharma on 06 December 2022.

The user has requested enhancement of the downloaded file.

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

Real - Time Object Detection in Video’s using YOLO v3

Priya Dalal,Department of IT, MSIT, NEW DELHI-110058, [email protected]

Tripti Sharma, Department of IT, MSIT, NEW DELHI-110058, [email protected]

Abstract - Object Detection implying deep learning is actually giving good results. Object
detection is used comprehensively now a days. This methodology helps in detecting real - world
objects and also helps in creation of those objects. Nonetheless many object detection methods
subsists but we are sometimes unable to achieve precision, acceleration and effectiveness.
Consequently, this paper manifests real- time problem solving detection by using YOLOv3
algorithms. Along with YOLOv3 deep learning techniques are also used for real-time detection.
Firstly, the crosswise predictions are made above 3 distinctive scales. The layer which we are
using for recognition or identification is used to make distinguished at peak or high point maps
of three variant sizes. The sizes are having strides as 32,16,8 specifically.this suggests that we
can altogether form location bases on scales of 13X13, 26X26 and 52X52 in the company of
416X416 contribution. For the time being it also utilizes the strategic regression to foresee the
small screen article score.the cross-entropy set adversity is employed to anticipate the classes
that the bounding box can accommodate. The validity is resolute and after that we resolute the
forecast. It leads to accomplishment of multi-label categorization for the objects identified in
images. The average accuracy for small objects is enhanced which is rapid and greater than the
RCNN. MAP expands remarkably.the expansion or increase of MAP leads to reduction in the
errors. By using PyTorch Libraries and YOLOv3 we can find out the Objects in Video streams.

Keywords - YOLO v3, Deep Learning, Clustering of High-Dimensional Data, Object Detection,
PyTorch.

1. INTRODUCTION
There are several aspects where we can apply for instance structuring of mechanized vehicles,
identification of a person’s foot, for applying self-governance, recognizing movement, for
automation of CCTV, object checking, etc. Recently object recognition has elevates student
learning to heights by its faster growth in an outstanding way. The Common methods which are
used excessively for locating targets are divided into two categories. First one is the single step
indication along with location proposition which we call them as recognition methodologies [1].
YOLOv3 (seen just once) has a region containing reclusive improved identifier. This is a fast as
well as intelligible object for innovating location. In contrast with faster RCNNs and SSDs the
accuracy of YOLOv3 is quite less . R-CNN is very faster for small targets but the recognition
speed is quite fast ,therefore it can be used much better for building. At the same timeYOLOv3
is similar to or look like RCNN quicker while precision identification when the objectives are
pro-fused. In addition to ,YOLOv3 is likewise superior to SSD when we talk about the accuracy
and speed of location. Anyhow, the methods for obtaining the models for recognition or
identification by means of preparing a vast amount of tests is mainly uttered through the large
amount of tests. A great number of techniques for object detection includes digital image
processing and three-dimensional object detection.
In addition to the plan which we are using do not fulfil the ideal results up to stable
execution.The indicated write-up have gathered various models[2].the models we have obtained
have lighter objects and these models have better comparison in contrast to the background.
VOLUME 9, ISSUE 11, 2022 PAGE NO: 314
GIS SCIENCE JOURNAL ISSN NO : 1869-9391

Therefore YOLOv3 techniques are used for training models for object detection. This is basically
a two-step process. The earlier object detection model is produced with the first model whereas
the final model is produced by using an alleviated model. Eventually the recognition impact of
these two methods are countersigned and the fundamental end is evaluated.

In this paper we have explained about how to perform object detection in python and to be specific
we have applied YOLO object detection with the use of Opencv. You only look once YOLO is a
ultra modern rel-time object detection system. YOLO is a deep learning algorithm to an image
which came out in may 2016 and it quickly became so popular because it is so fast compared with
the other deep learning object detection models traditionally recurrent convolution neural network
applies regions to localize the object to perform object detection which means that the model is
administered to multiple regions within an image and then model computes scores in an image at
different positions and scales. High scores regions of an image are contemplated as an object is
detected.

Traditionally, R-CNN descent of algorithm uses domains to constrain the objects in images. High
scoring domains of the image are contemplated as object detected.

YOLO follows a completely distinct process. It applies a neural network to the entire image to
predict bounding boxes and their probabilities. High probabilities of the image are considered as
object detected. Since it only scans the image once to make the predictions as compared to other
algorithms which requires multiple scans. It is faster in practice and that is why it is called you only
look once YOLO. The latest version of YOLO is YOLO version 3. it makes use of a few schemes
to ameliorate the training and improve the performance. It Constitutes multiple scale predictions
and better best bone classifiers and few more minor techniques.This recent version is more
powerful than the basic YOLO and also the YOLO version 2.

In this paper we have the YOLO v3 and it is extremely fast and accurate as shown in the picture.

VOLUME 9, ISSUE 11, 2022 PAGE NO: 315

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

YOLO can be applied to

1. Image file
2. Webcam Feed
3. Video file

Pre-trained YOLOv3 can detect 80 objects, such as :

1. Human
2. Bicycle
3. Aircraft
4. Motorbike
5. Motor Car
6. Bus

Etc.

Transfer learning is very important and interesting concept of applying deep learning benefits
because very often we are solving a different but yet somehow similar problems. To take
advantage from others work and to speed up our training process we can usually use partly or
wholly from others betraying the network to accelerate our own training and solve our own
problems. In deep learning this concept is called transfer learning. This means that we are using the
weights in one or more layers from a pre-trained neural network model in a new model by keeping
the rate and fine tuning the weight or adapting the rate entirely when training a new model. In
YOLO we are applying similar concepts. We will simply download the rate and configurations of
YOLO and download the name file which is called coco and use the deep learning framework in
opencv that is compatible with YOLO. The advantage of using this that it works without the need
to install anything except the we need to install opencv and one friendly reminder is that the
versions has to be at least 3.4.2.

First we can download the weights and configuration files. There are 5 models that we can select
on your preference. For example of your concern is the speed we can pick the highest frame per
second fps models that is yolo wavy tiny. But if we want to have a higher accuracy we can pick
YOLO v3416 or YOLO v3608. the weight file is used to train the models and which is the core of
the algorithm to detect the objects and the configuration files cfg files is the setting of the YOLO
algorithm and then we can download the name file from github. The name file contains the name of
the objects that the YOLO algorithm can detect . in other words it contains the 80 object names.
The files contains the labels of the classes that the pre-trained models can classify and finally we
can simply open a terminal to pip install opencv python and then we can put everything inside a
folder that contains the program that we have used.

VOLUME 9, ISSUE 11, 2022 PAGE NO: 316

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

We need to import cv2 and also we need to import NumPy SMP so first of all we also need to load
the YOLO weights and configurations and also the object names so everything is under the same
folder. Opencv provides functions to load rate rate and configuration files without the need to
convert them so this is very convenient and you don’t need to analyze or write your own loading
functions and the functions will return model objects that we can use later on for predictions so first
we just create variables called net and then use the opencv function cv2.dnn.net and then pass the
go low free and also pass the other parameters that is the YOLO re free configuration file. These
networks contains the YOLO feed weights and also YOLO configurations so the next things that
we do is extract the object names from the cocoa file and put everything into leads.

VOLUME 9, ISSUE 11, 2022 PAGE NO: 317

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

2. THEORY
[1] Bounding Box forecasting
For producing Bounding Box predictions to be exact along with sliding windows , we basically
withdraw distinct positions. After extracting various locations we run the classifier across it.
Suppose in our case no box is exactly equivalent in accordance with the location of the object.
So what we do is we make that box the best match. And also if we look at the ground truth
perfect boundary box has moderately wider rectangle or we can say has moderately horizontal
aspect ratio, it is not at all square. We have to find out a method to use this algorithm to output
precise bounding boxes. The best method to achieve more precise output for bounding boxes is
using the YOLO algorithm. The full form of YOLO is you only look once. Suppose if we input
an image of dimension 100X100, then we will put down a grid on the input image we are using
and will use a 3X3 dimension grid for the reason of illustration. Despite the fact that we will
use a finer method for actual implementation like we will use a 19X19 dimension grid, the main
motive is to extract the localization and classification algorithm for the image. After extracting
this we put in to each of the nine grid cells of the image. so the more concrete is to define the
labels which we are using for training . so for each of the nine grid cells we will specify a label
y where the label y is eight dimensional vector.if there is an object associated accurate cell then
say c1 c2 c3 , if you are trying to recognize three classes not counting the background class so
we are trying to recognize pedestrians so we have a vector for each grid cells. We start with
upper left cell.if there is no object then the label vector y for the upper left grid cell would be
zero and then don’t care for the rest of these and the output label y would be the same for this
cell and all the grid cells with no interesting object in them. A package of dimensions was
employed for generating anchor frame in YOLOv3. Also YOLOv3 is a single network,
therefore the deprivation of objectivity is intended separately and the allotment needs to be
intended separately by the network as such. YOLOv3 anticipates the fairness score by
employing the logistic regression. In this method the selection rectangle is first overplayed by
the process entirely on the object of the elementary truth[3]. This imparts a solo bounding box
prior to a terrestrial object also known as faster RCNN divergent and if there is any fault in this
would came both in the assignments and in the recognition deficit which is the objectivity.
Already existing of the selection rectangle that intend to have an objectivity score higher than
the threshold but lower than the finest. These delusions happen only for the recognition
deficiency but not for the allocation.

VOLUME 9, ISSUE 11, 2022

Fig-1 Bounding Box PAGE NO: 318
GIS SCIENCE JOURNAL ISSN NO : 1869-9391

[2] Predicting Class

Each and every classifiers nearly approximates the labels for the output are distinctive at the
same time. The outcome of this is that object classes which are complete are true and hence
YOLO executes a softmax function to transcribe the result into possibilities that hold up one.
YOLOv3 utilizes a collective categorization beside suing the tag[4]. For instance if we take
tags for output as "men" and "women" which are not generalized. (The aggregate of the output
must be larger than one). to fathom the likelihood that the item belongs to or held by a
specific label YOLOv3 changes the softmax function with the unconventional logistic
classifiers to work out or solve. To resolve the categorization loss rather than using the mean
square error YOLOv3 employs the binary privation of cross entropy for every label.by using
this we can lessen the complexity of the computation by ignoring the softmax function.

[3] Predictions over scales

The triple distinct scales which we use for predictions are mentioned here. We basically bring
out various properties from these scales as Feature Pyramid Network for object detection.
Then we integrate these convolutional levels for the Darknet-53 which is basically a function
separator which separates all functions.Class forecasting , objectivity and making of
delimitation tables is covers the last level.for the COCO data set also we uses three different
tables for every scale.the out-turn we receive is basically four remuneration each for the
bounding-box, eighty-class predictions and prediction of objectivity which acts as an output
formula.as a consequence, the map for feature is grabbed from the last two levels. Then a map
of the features of preceding one is withdrawn on the network and we join the feature of the
samples using concatenation.Like we have devised Single Short Detector for Deconvolutional
single short detector similarly decoder-decoder design is a conventional approach.we can
acquire more precise data of the sample features using this method. Also an elaborated
information can be acquired on former map of features. Consequently multiple levels of
convolution are joined to strengthen this map of collaborated functions. They also imparts a
related tensor despite the fact that it is now two times greater. The collection of k mean is
utilized here to perceive a superior or recommended bounding box first and eventually we get
a COCO set of data which is (10 × 13), (16 × 30), (33 × 23), (30 × 61), (62 × 45), (59 × 119),
(116 × 90), (156 × 198) and (373 × 326) which we apply[6].

[4] Feature Extractor: Darknet-53

The tertiary scale of modules in Darknet-53 ranges from layer 0 to layer 74. out of these , we
have 53 convolutional layers and the rest of the layers are known as native or resident layers. It is
same as we have the elementary structure of the system which we are using for withdrawal of
YOLOv3 features[7]. this arrangement makes use of a development of 11 convolutional layers of
3X3. these 11 convolutional layers are obtained by consolidating convolutional layers of
different system structures with substantial demonstration. The design or layout of Darknet-53 is
according to the subsequent. Darknet-53 as compared to Darknet-19 is more preferable. As the
same time 1.5 occasions are more successful than the resnet101 when we aim for the better
implementation. It is similar to the capability of resnet-152 with equivalent effect as resnet-152.
VOLUME 9, ISSUE 11, 2022 PAGE NO: 319
GIS SCIENCE JOURNAL ISSN NO : 1869-9391

At the same time to draw out attributes of YOLOv2 we uses Darknet-19 classification
network[8].Presently, a much great Darknet-53 network is employed in YOLOv3. we can also
use 53 Convolutional stages but one YOLOv2 and the other YOLOv3 employs batch
normalization. The net fault rate are stated from Dominant 1 & Dominant 5 of 1000 image Class
and the Unit. Rather than ResNet-101 ,Darknet-53 provides superior execution and is rapid as
1.5 times[9].Darknet-53 has same performance in comparison to ResNet-152 but is fast two
times.

3. RELATED WORK
For object detection YOLOv2 employs a 19-Layer network along with a conventional deep
architecture as well as 11 layers more.YOLOv2 also provides less object detection even after
having a 30-Layer structure. This may lead to disappearance of exquisite features when the layers
sub sampled the input. To resolve this an identity mapping technique was employed by YOLOv2
where we concatenates the feature maps against preceding layers to seize the low level features.
Whereas the preeminent elements which are currently main in nearly all ultra modern or latest
algorithms are still missing in YOLO v2’s architecture. In YOLO v2 there is no over sampling, no
connections and no leftover blocks where as YOLO v3 has all of these features which are missing
in YOLO v2.

Yolov3 employs a modified version of Darknet which is having a network of 53 layers which are
basically trained on Imagenet.in YOLO v3 for detection purpose 53 layers more are joined on
the Darknet which finally results in 106 layer in total for fully convolutional elementary
architecture. This add up of more layers basically slow down the YOLOv3 as opposed to YOLO
v2. In the image given below we have shown the complete architecture of YOLO.

Fig-2 YOLO v3 Network Architecure

VOLUME 9, ISSUE 11, 2022 PAGE NO: 320

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

Three Scale Detection

The latest architecture of YOLO possesses super sampling and residual skip connections.
YOLOv3 builds detection at three distinct scales which is its most important characteristics.as we
all know YOLOv3 is a fully convolutional network. By pertaining 1X1 detection of kernels on
feature maps we can do the detection in YOLOv3. For these detection three different sizes at three
distinct locations are used in the network.

The structure of the detection kernel is 1X1X(BX(5+C)). Here B denotes the number of
bounding boxes that a call uses on the feature map prediction. “5” basically is used for the
attributes of the bounding box and one for object confidence. C here is the number of classes we
are using. If we train a YOLOv3 on COCO then we can take B=3 and C=80. The kernel size
would be 1X1X255. The feature map created by using these values by this kernel will have same
width ans well as same height as compared to the preceding feature map. Also it includes the
attributes for detection across the depth as we have described above.

Fig-3 Detection at three scales

The projection in YOLOv3 is made at three distinct scales which are exactly attained
by sub sampling. Here the measurements of the image input is given by 8,16 and 32
discretely. At the 82th layer we made the elementary detection.For all the elementary
81 layers the image is sub sampled by the network. This image is sampled in such a
way that the 81st has a pace of 32. Given an image of dimension 416X416 we can
achieve the resultant feature map of size 13X13. If we use the 1 x 1 detection kernel
for making the detection possible , then we will get a detection feature map of size
13X13X255.

VOLUME 9, ISSUE 11, 2022 PAGE NO: 321

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

Prior to sub sampling by 2x to dimensions of 26X26, we carryout to a small number

of one or two convolutional layers to feature map from Layer 79. similarly by
employing depth feature map we can instruct feature map from layer 61. The
indicated feature map is then joined by depth to the feature map from Layer 61. To
combine or to merge the properties from the previous Layer which is Layer 61 we
fuse feature maps once more subjected a few 1X1 convolutional layers.Then the next
recognition is done using the 94th Layer capitulating the recognition of feature map
of 26X26X255.

This same process is observed one more time so that the feature map from Layer 91 is
put through to hardly any convolutional layers prior to connection using depth with a
feature map from Layer 36. ahead hardly any 1X1 convolutional layers observe to join
the information from the preceding layer which is Layer 36. The conclusion of all the
3 is done at 106th Layer, capitulating feature map of size 52X52X255.

Can detect smaller objects with more superiority

When we talk about YOLO v2 the most common problem addressed is that while
working with multiple layers detecting small objects is quite difficult. The layers
which are not examined are basically connected among the preceding layers. This
concatenation conserves the fragile attributes that helps in small object recognition.
The layer which is behind large object recognition is a 13X13 layer . on the other
hand the layers which helps in recognizing the smaller objects is a 26X26 layer and
also this layer can recognize medium-sized objects. Now here different objects are
examined in comparison to the same object by various layers.

Selecting anchor boxes

Total 9 anchor boxes are utilized by YOLO v3.for each scale YOLO v3 uses three
boxes. If we want to work or train on our own data set the YOLO v3 utilizes K-
means Clustering which helps in generating 9 anchors.

After generating anchors we set out the anchors in reverse order by using dimension
as a parameter. For the first scale the largest 3 anchors are allocated. Then for the 2 nd
scale we allocate the other 3, similarly for the last scale we allocate the rest three. In
this way anchors are arranged.

Bounding Boxes per image are better

When we talk about the predictions the predictions for bounding boxes by YOLO v3
are quite large for any input image of the identical size as compared to YOLO v2.
Suppose we have an input image in case of YOLO v2 having resolution of 416X416.
then we can find the number of bounding box as 13X13X5 which is 845 boxes. Here
we are able to detect 5 boxes by suing 5 anchors at every grid cell.
On the other hand if we have an input image in case of YOLO v3 of same resolution
(416X416) as v2 then YOLO v3 is capable of doing predictions of boxes at 3
VOLUME 9, ISSUE 11, 2022 PAGE NO: 322
GIS SCIENCE JOURNAL ISSN NO : 1869-9391

variable scales. Here the prediction of boxes made are 10,647.this clearly shows that
the predictions made by YOLO v3 are 10 times more than the predictions made by
v2. That is why we call YOLO v2 slower. For every scale by using 3 boxes and 3
anchors at every grid, we are using three different scales. Therefore the total number
pf anchor boxes in use are 9 where we are using 3 for every scale.

Loss Function used is alternate

Initially the loss function used by YOLO v2 is this:

Fig-4 Loss Function

This loss function looks scary but if we focus on the last 3 terms. The first term of the
last three is used to show the score prediction of objects where we assume that the
scores for bounding boxes is ideally one. the next term of the last three is basically
used to define the bounding boxes with no objects therefore their score preferably
would be 0. the third term is used to recognize the bounding boxes which can
recognize the objects.

The difference lies in the last three terms which are squared errors in case of YOLO
v2, whereas in YOLO v3 we are using cross-entropy terms for errors instead of
squared errors. This means that we can now recognize the class prediction as well as
object confidence via logistic regression in case of YOLO v3. Here for every
foundation truth box we are preparing the detectors . after training the detectors we
allocate a bounding box to the anchor having largest overlaps with the foundation truth
box.

softmaxing the classes is deceased

VOLUME 9, ISSUE 11, 2022 PAGE NO: 323

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

The older versions of YOLO utilizes softmax for the class score where we select the class of
the object which has the highest score of the object consisting the bounding box. We have
changed this concept in case of YOLO v3. YOLO v3 mainly executes the multi label
categorization for finding out the objects in the input image.
In case of softmaxing we deduce that mutual exclusion exists in the classes or in other words if
some object is associated with one class then it can’t be associated with the other class also.
This scheme works best in case of COCO data set. This scheme or this premise totally fails if
we have classes like Human & Figure in any data set.

That is why we avoid using softmaxing the classes in case of YOLO. The method which is
followed is recognizing every class score by utilizing the logistic regression. An approach is
used to recognize the multiple labels of an object where higher is the class score more is the
threshold allocated to the box
.

Criterion
The performance of YOLO v3 is far better than the latest detectors like RetinaNet.
YOLO v3 is substantially quick while using COCO map 50 criterion. YOLO v3 is
also superior to SSD and it’s alternative came. Fig-5 shows the comparative study of
YOLO v3 and RetinaNet on COCO 50 Criterion

Fig-5 Performance of YOLO v3 in contrast to RetinaNet on COCO 50 Criterion

In spite of that YOLO v3 is totally unsuccessful with COCO criterion as YOLO v3 has larger value
of Intersection over Unit which we utilized to fail detection. Is is almost impossible to describe
the COCO criterion here in this paper as it is far off from the work which we are doing right now.
50 in the COCO criterion is basically used to denote the 0.5 Intersection over Unit which
determines how exact we can predict the bounding boxes when we assign the foundation truth
boxes of the object. If the Intersection over Unit has value less than 0.5 then we can identify that
prediction as a mis-localization. Mis-localisation of prediction is clearly a false positive.

VOLUME 9, ISSUE 11, 2022 PAGE NO: 324

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

When we increase the number in case of criterion in COCO like COCO 75 we need to position the
boxes more accurately so that they won’t get rejected while evaluating the metric. So we can say
here RetinaNet performs better than the YOLO as we are unable to line up or position the bounding
boxes of YOLO and RetinaNet. Given below is a complete table for a broad diversification of
Criterion.

Fig-6 YOLOv2/v3 vs RetinaNet

If we give a quick look at this table we can conclude here that RetinaNet is better in comparison
to YOLO if the criterion followed is COCO 75. we can clearly see that RetinaNet also
outperforms at the AP(for little objects) criterion.

4. RESULTS

We will look at some of the results obtained from object detection using the video
streams.Here we have used video clips or trailers from the very popular and widely
watched movie series of Harry Potter first followed by Fast and Furious 9 and some
more trailers.

Images shown above is from Harry Potter 5, where person’s and tie is detected in the
frames when the code is run on command prompt.Images below are from extraction a
netflix
VOLUME 9, ISSUE 11, movie
2022 and harry potter detecting person and bus. PAGE NO: 325
GIS SCIENCE JOURNAL ISSN NO : 1869-9391

VOLUME 9, ISSUE 11, 2022 PAGE NO: 326

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

5. CONCUSION AND FUTURE SCOPE

We have learn that the foundation model is generally a pre-aquainted (classification)

network, normally positioned on a large image data set to learn a sturdy set of subtle
filters. The foundation network can also be trained in a different way from the beginning
however in the end it draws remarkably prolonged amount if time for detecting when we
are concerned about the logical precision or fidelity.Some of the points which can
strengthen our core research are increase the training data set, Add more features such as
detecting stop signs and rear view-back view of cars so that this project can be applied to
self-driving automobiles, Perform analysis on existing avenues for the application of our
project in order tofine- tune the shortcomings.

VOLUME 9, ISSUE 11, 2022 PAGE NO: 327

GIS SCIENCE JOURNAL ISSN NO : 1869-9391

6. REFERENCES

[1] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The

pascal visual object classes (voc) challenge. International journal of computer vision,
2010.

[2] C.-Y. Fu, W. Liu, A. Ranga, A. Tyagi, and A. C. Berg. Dssd: Deconvolutional
single shot detector, 2017.

[3] D. Gordon, A. Kembhavi, M. Rastegari, J. Redmon, D. Fox, and A. Farhadi. Iqa:

Visual question answering in interactive environments, 2017.

[4] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition.
In Proceedings of the IEEE conference on computer vision and pattern recognition,
2016.

[5] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara, A. Fathi, I. Fischer, Z. Wojna,

Y. Song, S. Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional
object detectors.

[6] I. Krasin, T. Duerig, N. Alldrin, V. Ferrari, S. Abu-El-Haija, A. Kuznetsova, H.

Rom, J. Uijlings, S. Popov, A. Veit, S. Belongie, V. Gomes, A. Gupta, C. Sun, G.
Chechik, D. Cai, Z. Feng, D. Narayanan, and K. Murphy. Openimages: A public
dataset for large-scale multi-label and multi-class image classification. Dataset
available from https://github.com/openimages , 2017.

[7] T.-Y. Lin, P. Dollar, R. Girshick, K. He, B. Hariharan, and S. Belongie. Feature
pyramid networks for object detection. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2017.

[8] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollar. ´ Focal loss for dense object
detection, 2017.

[9] J.Redmon. Darknet: Open source neural networks in c. Available from

http://pjreddie.com/darknet/, 2013–2016.

VOLUME 9, ISSUE 11, 2022 PAGE NO: 328

View publication stats