Neural Network mini-project#1 Spring 2022

UNIVERSITY OF TEHRAN

COLLEGE OF ENGINEERING

DEPARTMENT OF ELECTRICAL AND COMPUTER ENGINEERING

NEURAL NETWORK & DEEP LEARNING

S IAVASH SHAMS
MOHAMMAD HEYDARI

SCHOOL OF ELECTRICAL AND COMPUTER ENGINEERING

Apr. 2022

1
Neural Network mini-project#1 Spring 2022

1 CONTENTS
Object Detection .............................................................................................................................................3
YOLOv2 Vs YOLOv1 ................................................................................................................... 3
YOLOv5 and YOLOv4 Advances ................................................................................................. 4
One stage Vs Two Stage Object Detection ..................................................................................... 5
YOLOv5 for Bocce Balls............................................................................................................... 6

2
Neural Network mini-project#1 Spring 2022

OBJECT DETECTION

YOLOV2 VS YOLOV1

Figure31. Yolov1

The changes from YOLOv2 to YOLO v1:

 Batch Normalization: it normalizes the input layer by altering slightly and scaling the
activations. Batch normalization decreases the shift in unit value in the hidden layer and
by doing so it improves the stability of the neural network.
 Higher Resolution Classifier: the input size in YOLO v2 has been increased from
224*224 to 448*448.
 Anchor Boxes: one of the most notable changes which can visible in YOLO v2 is the
introduction the anchor boxes. YOLO v2 does classification and prediction in a single
framework. These anchor boxes are responsible for predicting bounding box and this
anchor boxes are designed for a given dataset by using clustering (k-means clustering)
 Fine-Grained Features: one of the main issued that has to be addressed in the YOLO
v1 is that detection of smaller objects on the image. This has been resolved in the YOLO
v2 divides the image into 13*13 grid cells which is smaller when compared to its previous
version. This enables the yolo v2 to identify or localize the smaller objects in the image
and also effective with the larger objects.
 Multi-Scale Training: on YOLO v1 has a weakness detecting objects with different
input sizes which says that if YOLO is trained with small images of a particular object. it
has issues detecting the same object on image of bigger size.
 Darknet 19: YOLO v2 uses Darknet 19 architecture with 19 convolutional layers and 5
max pooling layers and a softmax layer for classification objects. The architecture of the
Darknet 19 has been shown below. Darknet is a neural network framework written in
Clanguage and CUDA. It’s really fast in object detection which is very important for
predicting in real-time.

3
Neural Network mini-project#1 Spring 2022

Table1. Darknet 19 architecture

YOLOv2 has seen a great improvement in detecting smaller objects with much more accuracy
which it lacked in its predecessor version.

YOLOV5 AND YOLOV4 ADVANCES

YOLO4. uses Cross Stage Partial Network (CSPNet) in Darknet, creating a new feature extractor
backbone called CSPDarknet53. The convolution architecture is based on modified DenseNet. It
transfers a copy of feature map from the base layer to the next layer through dense block. The
advantages of using DenseNet include the diminishing gradient vanishing problems, boosting
backpropagation, removal of the computational bottleneck, and improved learning.
However, YOLOv5 is different from the previous releases. It utilizes PyTorch instead of
Darknet. It utilizes CSPDarknet53 as backbone. This backbone solves the repetitive gradient
information in large backbones and integrates gradient change into feature map that reduces the
inference speed, increases accuracy, and reduces the model size by decreasing the parameters.

Table2.

4
Neural Network mini-project#1 Spring 2022

Figure32. Average precision against GPU Time for different versions of YOLO

ONE STAGE VS TWO STAGE OBJECT DETECTION

One-stage detectors have high inference speeds and two-stage detectors have high localization and
recognition accuracy. The two stages of a two-stage detector can be divided by a RoI (Region of
Interest) Pooling layer. One of the prominent two-stage object detectors is Faster R-CNN. It has
the first stage called RPN, a Region Proposal Network to predict candidate bounding boxes. In
5
Neural Network mini-project#1 Spring 2022

the second stage, features are by RoI pooling operation from each candidate box for the following
classification and bounding-box regression tasks. In contrast, a one-stage detector predicts
bounding boxes in a single-step without using region proposals. It leverages the help of a grid box
and anchors to localize the region of detection in the image and constraint the shape of the object.

Figure33. Average precision against GPU Time for different versions of YOLO

Some of one-stage object detection methods are: YOLO, EfficientDet, RefineDet

YOLOV5 FOR BOCCE BALLS

We trained our YOLOv5 model for the Booceball dataset which has 6 objects to detect; Balls
with 5 different colors and vlines.
Training time: 1hour 28minutes on Tesla K80 GPU and 12GB of RAM, 100 epochs
First, we upload the dataset with a .yaml file on the RoboFlow website and then use the
roboflow library in python with the API obtained from the website to load the data in the Colab
environment. Then we git clone the YOLOv5 repository in Github and build our model
(anchors, backbone, head). Then we train our model on the dataset for 100 epochs using 16
batch sizes. After training, we will show precision and accuracy graphs for validation and training
data. Then we will show an augmented (the augmentation procedure is explained in the
"README.roboflow.txt") image from the training dataset and a ground truth image with labels
from the validation dataset along with the image with the predicted labels by the model. Then we
will run the "detect.py" from the YOLOv5 repository to perform object detection on the test
dataset and show the results.

6
Neural Network mini-project#1 Spring 2022

Figure34. Average precision against GPU Time for different versions of YOLO

Figure35. Average precision against GPU Time for different versions of YOLO

7
Neural Network mini-project#1 Spring 2022

Figure36. Average precision against GPU Time for different versions of YOLO

Figure37. Average precision against GPU Time for different versions of YOLO

8
Neural Network mini-project#1 Spring 2022

Figure38. Average precision against GPU Time for different versions of YOLOv5