Yolov5 learning notes (I) -- principle overview

Catalog

One 、 Target detection Overview

1.1 Data set introduction

1.2 Performance indicators  

1.2.1 Confusion matrix

1.2.2 IOU（ Bounding box regression ）

1.2.3 AP&mAP

1.2.4 Detection speed

1.3 YOLO The history of

​1.3.1 Algorithmic thought

1.3.2  YOLOv5 Network architecture

---

One 、 Target detection Overview

1.1 Data set introduction

• PASCAL VOC

• MS COCO

1.2 Performance indicators  

1.2.1 Confusion matrix

1.2.2 IOU（ Bounding box regression ）

IOU=1 Is completely overlapping

according to IOU Set the threshold to judge TP still FP , For example, the overlap is 0.5

1.2.3 AP&mAP

AP It is to measure the quality of the learned model in each category

mAP It is to measure the quality of the learned model in all categories AP Average value

•   Different data sets for AP and mAP The concept of is still different

• AP Calculation  

1.2.4 Detection speed

1.3 YOLO The history of

1.3.1 Algorithmic thought

• yolov3 Frame diagram

First, the characteristic image is obtained by convolution neural network , Then mesh the image , Each grid carries out frame detection and category probability map separately , The final result .

Each small box contains the coordinates of the bounding box 、 Goal score and category score

• Multiscale fusion

  After convolution neural network, we can get different sizes of characteristic graphs , The fusion of feature maps of different sizes is conducive to the detection of small targets .

Image convolution passes 32 Times down sampling 19*19 Pictures of the , Each grid will predict and draw anchor boxes separately  

Preset the size of some bounding boxes , Each scale has several anchor frames

 

• Loss function  

1.3.2  YOLOv5 Network architecture

A network often has a backbone network （Backbone）+ neck （Neck）+ Head （Head） form  

 

• visualization  

pip install onnx>=1.7.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install coremltools==4.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
python models/export.py --weights weights/yolov5s.pt --img 640 --batch 1