How YOLO Works

YOLO works by deviding images into certain fixed cells and detect whether object is present their or not. If object is present, it will be detected and gives a output vector as below:
pc means probability of class whether there is dog or cat or both or neither. If present, it will be 1.0 else 0.0.
bx,by,bh,bw means bounding box coordinates. It is a rectangle which encloses the object.
c means class. It is the class of the object. That means whether that is dog or cat in the case of detecting dog and cat.
After performing all above task it will perform a IOU which is Intersection Over Union in order to fine tune the output.
IOU is a measure of how much two boxes overlap. If two boxes overlap, it will be 1.0. If they don't overlap, it will be 0.0. Correct if IOU>=0.5 .

If there is not object in the cell, it will be 0.0 in all the output vector.

Architechture of YOLO-V4

There is mainly 4 layers in YOLO-V4 they are:

INPUT :
- This is the input layer. It is the input image.
Backbone :
- This is the backbone of the network. It is the feature extractor.
- For extracting features in the image, it uses CSPDarknet53 model
Neck :
- Here in this block extracted features are collected to feed to the Dense Prediction.
Dense Prediction :
- This is the dense prediction block. It is the final block of the network.
- Here YOLO-V3 used for detection. SO it does object detection, classification and plotting bounding box.

Object Detection Using YOLO-V4 on Custom Dataset

Mount a drive first by running following code:

from google.colab import drive
drive.mount('/content/drive')
!ln -s /content/drive/My\ Drive/ /mydrive
!ls /mydrive

Install Darknet by cloning the repository by following command

!git clone https://github.com/AlexeyAB/darknet

Change makefile to have GPU and OPENCV enabled

Initially OPENCV=0 means inactive but following command makes it activate
Initially GPU=0 means it is deactivated but by following code it activate
Initially CUDNN=0 means it is deactivated but by following code it activate

%cd darknet# 
!sed -i 's/OPENCV=0/OPENCV=1/' Makefile
!sed -i 's/GPU=0/GPU=1/' Makefile
!sed -i 's/CUDNN=0/CUDNN=1/' Makefile
!make

Configuring configuration file for training by making a copy of yolov4-tiny.cfg named namaste.cfg in my case.

After running below code cell go inside darknet folder and then cfg folder, there u will find copy of yolov4-tiny.cfg, in my case it is named as namaste.cfg.
Open that in google docs and update the parameter as follows : Training-> subdivisions=16 max_batches = 6000 for two class and general formula to calculate batch size is given as:

steps = 80% of max_batches, 90% of max_batches
Now go to the bottom of the file where [yolo] resides, there classes are by default 80 you should set it according to how much classes you have. Do this for all classes which is under [yolo].
Now Just above the [yolo] heading there is filters which you should change according to following formula :

- Do this for every [yolo] headings

!cp cfg/yolov4-tiny.cfg cfg/namaste.cfg

Creating obj.names and obj.data files

Following code will create two files inside darknet/data/ and set the path to the train.txt and test.txt by creating it.
It also create obj.data and obj.name file, where obj.data is the file which contains the classes path to train.txt, test.txt, backup folder where our model will reside and a path to the google drive folder and obj.names contains the classes name.

!echo $'Namaste' > data/obj.names



!echo -e 'classes= 2\ntrain  = data/train.txt\nvalid  = data/test.txt\nnames = data/obj.names\nbackup = /content/drive/MyDrive/yolo_data/' > data/obj.data
!mkdir data/obj

Data creation part

This is the most crucial part of the project where you create a dataset for your project.
Once you collect a data that you want to train yolo model, then you need to annotate it. Which can be done using labelimg tool and tutorial for this is here.
After annotating your images upload it to github by making it in a zip file.

Following code will unzip the annotated data that you upload in a google drive.

!unzip /mydrive/yolo_data/multipleimages.zip -d data/obj

Following code will return a list of paths matching a pathname pattern.

import glob
images_list = glob.glob("data/obj/multipleimages/*.jpg")
print(images_list)

Following code will create train.txt file which is essential for training our model.

file = open("data/train.txt", "w") 
file.write("\n".join(images_list)) 
file.close()

Following code will train YOLO-V4 model and save the model in your google drive path which you have given in a obj.data file.

!./darknet detector train data/obj.data cfg/namaste.cfg darknet53.conv.74 -dont_show