2020 IEEE Region 10 Symposium (TENSYMP), 5-7 June 2020, Dhaka, Bangladesh

CPU Based YOLO: A Real Time Object Detection

Algorithm
Md. Bahar Ullah
Department of Electrical and Electronic Engineering
University of Chittagong
Chittagong, Bangladesh
[email protected]

Abstract—This paper describes CPU Based YOLO, a real

time object detection model to run on Non-GPU computers
that may facilitate the users of low configuration computer.
There are a lot of well improved algorithms for object
detection such as YOLO, Faster R-CNN, Fast R-CNN, R-CNN,
Mask R-CNN, R-FCN, SSD, RetinaNet etc. YOLO is a Deep
Neural Network algorithm for object detection which is most
fast and accurate than most other algorithms. YOLO is
Fig. 1. Convolutional Neural Networks design
designed for GPU based computers which should have above
12GB Graphics Card. In our model, we optimize YOLO with
OpenCV such a way that real time object detection can be II. RELATED WORK
possible on CPU based Computers. Our model detects object The evaluation in computer vision with Deep Learning
from video in 10.12 – 16.29 FPS and with 80-99% confidence has been established and accomplished with time, primarily
on several Non –GPU computers. CPU Based YOLO achieves over one particular well-known algorithm named
31.05% mAP. Convolutional Neural Networks (CNN). Fig. 1 shows the
Keywords—object detection, real time, YOLO, CPU, deep architecture of CNN. It has a convolution layer where a
learning filter is convolved with various parts of the input to generate
the output. The use of a convolution layer permits for
I. INTRODUCTION relative patterns to be drawn from associate input.
Computer vision is associate knowledge base field that Moreover, a convolution layer tends to own less weight that
has been gaining enormous amounts of traction within the require to be learned than a completely connected layer as
recent years and self-driving cars have taken center stage. filters don't need assigned weight from each input to each
Another integral part of computer vision is real time object output [1], [2].
detection. Real time object detection aids in cause
estimation, vehicle detection, traffic control, CCTV A. YOLO
monitoring etc. For this reason, day to day progress in To generate a single step procedure including object
technology is necessary so that the ability to give vision in detection and classification You Only Look Once (YOLO)
computer is possible. So, low cost technology and learning [7] was introduced. YOLO is successful to gain 45 FPS and
tools are in requirement [1], [2], [3]. further a lite version named Tiny-YOLO, achieves around
The techniques which are currently used for computer 244 FPS (Tiny-YOLOv2 [11]) on GPU computers.
vision are more costly than previous because of developing Why YOLO is differ from other existing networks? The
larger and deeper networks to gain higher accuracy [4], [5], answer is: at the same time bounding box predictions and
[6]. Techniques that are entirely computer based mostly need class predictions can be done by YOLO. First, the input
vast quantity of GPU power and even then aren't always real image is sliced up into S × S grids [7]. Second, B bounding
time, making them inappropriate for day to day applications. boxes are included in each grid cell, each with a score of
The methods such as YOLO(You Only Look Once) [7], confidence. The probability of an object exists in each
Faster R-CNN [8], Fast R-CNN [9] etc. have affluently bounding box is known as confidence and is defined as:
obtained an efficient, fast and accurate model with high = ( )∗ (1)
mean average precision (mAP); But, their frames per second
(FPS) on non-GPU computers drive them unsuitable for real- Where, IOU is the abbreviation of intersection over union
time application. In this paper, we are proposing a model which describes a measure (a fraction between 0 and 1) of
named “CPU Based YOLO” to run YOLO [7] algorithm on overlap between two bounding box. Intersection is the
non-GPU computer with OpenCV [10]. Our model is able to overlapped area between the predicted bounding box and
execute videos from external source or from webcam with ground truth (Object), and union is the total area of predicted
minimum 10.12FPS, 80-99% confidence and with 31.05% and ground truth boxes as drawn in Fig. 2. For accurate
mAP that is suitable for real time application within low cost detection, the IOU should be near to 1, so that the predicted
and less effort. bounding box lies closer to the ground truth. IOU can be
expressed as:

978-1-7281-7366-5/20/$31.00 ©2020 IEEE

552

Authorized licensed use limited to: Auckland University of Technology. Downloaded on December 18,2020 at 13:18:33 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. COMPARISON BETWEEN DIFFERENT YOLO VERSIONS
= (2)
Performance Evaluation
Predicted Model Input
Predicted Image Train Set Test Set mAP FPS
Size
Intersection YOLOv1 448x448 VOC 2007+2012 VOC 63.4 45
[13] 2007
Fast YOLOv1 448x448 VOC 2007+2012 VOC 52.7 155
Ground Truth Ground Truth
2007
YOLOv2 416x416 VOC 2007+2012 VOC 76.8 67
(a) Area of Overlap (b) Area of Union 2007

Fig. 2. Intersection Over Union Tiny – YOLOv2 [11] 416x416 VOC 2007+2012 VOC 57.1 207
2007
YOLOv2 608x608 COCO [14] COCO 48.1 40
At the same time, while generating bounding boxes, each
YOLOv3 [12] 608x608 COCO COCO 57.9 20
grid cell predicts C conditional class probability of the
object. The class-specific probability for each grid cell is:
YOLO is later upgraded with various versions such as
( | )∗ ( )∗ YOLOv2 or YOLOv3[12] in order to optimize localization
= ( )∗ (3) errors and increase mean average precision (mAP). The FPS
and mAP of different version of YOLO is shown in Table I.
YOLO utilizes the equation below for calculating loss
B. Frameworks
function and finally improve confidence [7]:
To develop our model we need to install a deep learning
LossFunction: framework where we will run the YOLO algorithm. There
are few frameworks for running an algorithm which are
1 ( − ) +( − ) discussed below:
• TensorFlow: It is a deep learning framework
created by Google which can be used for designing,
+ 1 − + ℎ − ℎ building, and training models. But it needs huge
amount of GPU power and is favorable to Linux
[15].
+ 1 − • Darknet: It was developed by the developer of
YOLO based on Linux. And it runs better on GPU
based computers [16].
+ 1 −
• Darkflow: It was made by adapting darknet to
Tensorflow and works very fast in GPU based
+ 1 ( ) − ̂ ( ) . (4) computers. It also runs in CPU based computers but
installing in windows is terrible and very slow [17].
∈
• Opencv: It was built by Intel and also it has a deep
To legitimate the center and the bounding box of each learning framework. It works only in CPU and
prediction, the loss function is used. Every input image is installation in windows is easy [10].
divided into an S × S grid, with B bounding boxes for each
grid. The and variables are coordinates of center of III. CPU BASED YOLO ARCHITECTURE
each prediction, while and ℎ refer to dimensions of Our aim is to build the model (CPU Based YOLO) for
bounding box. The variable is used to increase CPU Based real time object detection. Developers of YOLO
emphasis on boxes with objects, and variable decrease used Darknet [16] framework for running the algorithm. We
the emphasis on boxes with no objects. C refers to the run YOLOv3 with DarkFlow [17] and OpenCv [10] as only
confidence score, and ( ) refers to the prediction of they are favorable for CPU Based YOLO. Our aim was to
classification. The 1 is 1 if the bounding box in the develop the model in Windows operating system but we
found that DarkFlow installation in windows is really a
cell predicts the object, an else 1 is 0. If the object is in complex task. However, we installed it and run YOLOv3
cell i then 1 is 1 and 0 otherwise. The loss function [12] on it, but the result was terrible. The FPS was too low
indicates the performance of the model, if loss is low then and the starting time was too much. Then we start to do the
performance will be high [7]. The accuracy of predictions task on OpenCv. We input a video though OpenCv and
generated by models in object detection is calculated through found a good FPS in the output. The architecture is shown in
the average precision equation defined as [7]: Fig. 3.
A. Setup
= ( )∆ ( ) (5) We loaded YOLOv3 and Dataset (COCO) [14] through
the framework OpenCv. Procedure of loading programme is
( ) refers to the precision at threshold k and ∆ ( ) refers shown in Fig.4.
to the fluctuate in recall [7].

553

Authorized licensed use limited to: Auckland University of Technology. Downloaded on December 18,2020 at 13:18:33 UTC from IEEE Xplore. Restrictions apply.
 above 16 but not so accurate for tiny objects like cup, book,
YOLO cellphone. The result is shown in Fig. 7. We run the same
task in many videos and found almost same result. As we
Real Time
used OpenCv, GPU was not used by the framework.
OpenCv Object
Detection

COCO
dataset

Fig. 3. Architecture of CPU Based YOLO

(a) Blob: 90x90 (b) Blob: 220x220
When we input a video then we found that the system was
detecting object from the input but the FPS was too low
(around 2FPS) but higher than that of DarkFlow. Then we
focused on increasing FPS and began to reduce the Blob size
in the input video and observed the detecting performance
for testing. The loading programme is shown in Fig. 4.

(c) Blob: 64x64(not detecting the car)

Fig. 7. FPS and Confidence in Different Blob Size from Experiment-1

C. Checking in Another Computer

To justify the system performance in another
computer we done the task in a very low configuration
Fig. 4. Programme of Loading YOLOv3 Using OpenCV
laptop which is built by Intel® Core™ i3-5010U CPU @
2.10GHz with 4GB RAM. We found 4.47FPS when blob
B. Blob Size Reduction
size was 90x90 and 7.7FPS when blob size was 64x64,
Blob size is a group of connected pixels in the input shown in Fig. 8. This FPS is also higher than most other real
image or video that assigns several common properties of the time object detection models.
object class. We used several blob size for finding perfect
FPS in CPU based computers. First, we used blob size
220x220 in the OpenCv [10] function names
“cv2.dnn.blobFromImage()”, we found 2.40FPS and 99% to
100% confidence in output from webcam video which is
above most of the current algorithms. Then we did the same
task in several times using the blob size 416x416, 320x320,
128x128, 90x90, 64x64 etc. using several inputs which is
shown in Fig. 5. The Computer configuration was AMD Blob: 90x90 Blob: 64x64
Ryzen™ 3 2200G with Radeon™ Vega 8 Graphics 3.50GHz Fig. 8. FPS and Confidence in Different Blob Size from Experiment-2
with 8GB RAM, shown in Fig. 6. We found that low blob
size increase the FPS but decrease the detection accuracy. D. mAP calculation
We perform mAP calculation in COCO dataset [14] in
our computer using OpenCV, JSON [20], Matplotlib and
found 31.05% mAP, shown in Fig. 9. The mAP is low due to
Fig. 5. Blob Size Reduction Programme reducing the blob size of the input video. But it will not be a
big fact for low configuration computer users.

Fig. 6. AMD CPU Used for Experiment -1

But we need an optimum value of blob size for CPU Based

YOLO. We found that in 90x90 the detection was good
accurate and also the FPS is above 10. In 64x64, the FPS is Fig. 9. mAP from Experiment

554

Authorized licensed use limited to: Auckland University of Technology. Downloaded on December 18,2020 at 13:18:33 UTC from IEEE Xplore. Restrictions apply.
 80

70

60

50

40 mAP

30 FPS

20

10

0
YOLO(GPU) [7] Fast R-CNN(GPU) [9] DPM(GPU) [18] Faster R-CNN(GPU) [8] SSD(GPU) [19] CPU Based
YOLO(Proposed Model)

Fig. 10. Comparison with other Networks

[5] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi, “Inception-v4,

inception-resnet and the impact of residual connections on learning.”
IV. RESULTS AND COMPARISON WITH OTHER ALGORITHMS In AAAI, vol. 4, 2017, p. 12. 1
[6] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for
We perform several experiments in several computers for image recognition,” in Proceedings of the IEEE conference on
finding an optimum value of blob size so that we can do the computer vision and pattern recognition, 2016, pp. 770–778. 1
task with better accuracy and FPS. We found 31.05% mAP [7] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look
and 16FPS (maximum) which is good enough for low once: Unified, real-time object detection,” in Proceedings of the IEEE
configuration CPU based computers. The results for the conference on computer vision and pattern recognition, 2016, pp.
779–788. 1, 2
experiments are shown in the Table II. From the comparison
with other GPU based algorithms, it is clear that our model is [8] S. Ren, K. He, R. B. Girshick, and J. Sun, “Faster R-CNN: towards
real-time object detection with region proposal networks,” CoRR, vol.
eligible for CPU based applications. The comparison is abs/1506.01497, 2015. [Online]. Available: http://arxiv.org/abs/1506.
shown in Fig. 10. 01497. 1
[9] R. B. Girshick, “Fast R-CNN,” CoRR, vol. abs/1504.08083, 2015.
TABLE II. RESULTS FROM EXPERIMENTS ON TWO CPU [Online]. Available: http://arxiv.org/abs/1504.08083. 1
[10] G. Bradski, A. Kaehler, 2008, Learning OpenCV – Computer Vision
Computer /Laptop RAM Dataset mAP FPS with the OpenCV Library, O’Reilly Media Inc. Sebastopol. 1, 2, 3
AMD Ryzen™ 3 2200G 8GB COCO 31.05% 16 [11] J. Redmon and A. Farhadi, “Yolo9000: Better, faster, stronger,” arXiv
CPU 3.50GHz [14] preprint, 2017. 2
Intel® Core™ i3-5010U 4GB COCO 31.05% 7.7 [12] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,”
CPU @ 2.10GHz arXiv preprint arXiv:1804.02767, 2018. 2
[13] M. Everingham, L. Van Gool, C. Kl. Williams, J. Winn, and A.
Zisserman. The pascal visual object classes (voc) challenge. JJCV,
2010. 2
V. CONCLUSION AND FUTURE WORK
[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev,
CPU Based YOLO gains the purpose of real time object Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C.
detection on Non - Computers. First, we selected a right Lawrence Zitnick, Piotr Dollár. Microsoft COCO: Common Objects
framework for doing object detection task in CPU then we in Context. In ECCV. 2014. 2, 3, 4
utilized YOLO algorithm in a right manner and done the task [15] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo,
Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey
successful. We can apply this model to normal Desktop or Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew
Laptop for executing YOLO. Thus we can do real time Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal
object detection for several purposes like video surveillance, Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan
traffic monitoring, face tracking etc. In future, we will Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike
perform the tasks by training the machine with custom Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal
Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan,
dataset. We will increase the mAP and FPS by optimizing Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Wattenberg,
the model. Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. Tensorflow: Large-
scale machine learning on heterogeneous distributed systems, 2016,
REFERENCES arXiv:1603.04467. 2
[1] J. Schmidhuber, “Deep learning in neural networks: An overview,” [16] J. Redmon. Darknet: Open source neural networks in c.
Neural networks, vol. 61, pp. 85–117, 2015. 1 http://pjreddie.com/darknet/, 2013–2016. [Accessed January 06,
[2] C. Szegedy, A. Toshev, and D. Erhan, “Deep neural networks for 2020]. 2
object detection,” in Advances in neural information processing [17] Darkflow, 2019. https://github.com/thtrieu/darkflow. [Accessed
systems, 2013, pp. 2553–2561. 1 December 25, 2019]. 2
[3] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature [18] P.F. Felzenszwalb, R.B. Girshick, D. McAllester, and D. Ramanan.
hierarchies for accurate object detection and semantic segmentation,” Object detection with discriminatively trained part-based models.
in Proceedings of the IEEE conference on computer vision and IEEE Transactions on Pattern Analysis and Machine Intelligence,
pattern recognition, 2014, pp. 580–587. 1 (9):1627–1645, Sept 2010.4
[4] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, [19] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.- Y. Fu, and
“Rethinking the inception architecture for computer vision,” in A. C. Berg. Ssd: Single shot multibox detector. In European
Proceedings of the IEEE conference on computer vision and pattern conference on computer vision, pages 21–37. Springer, 2016. 4
recognition, 2016, pp. 2818–2826. 1
[20] C. Severance, "Discovering JavaScript Object Notation,”
Computer, vol. 45, pp. 6–8, 2012.4

555

Authorized licensed use limited to: Auckland University of Technology. Downloaded on December 18,2020 at 13:18:33 UTC from IEEE Xplore. Restrictions apply.