Asian Games countdown! AI target detection helps host the Asian Games!

Pay attention to the parallel stars

Never get lost

Institute of computer vision

official account ID｜ComputerVisionGzq

Study Group ｜ Scan the code to get the join mode on the homepage

Computer Vision Institute column

author ：Edison_G

object detection It is now the most popular research topic , At present, the most popular is Yolo Series frame , Recently, our computer vision research institute also shared a lot of work and practice of target detection , All are Yolo-Base frame , Today we share a modified version of Yolov5, Effect of real-time detection ！

Open source code ：https://github.com/ultralytics/yolov5/releases

01

Preface

object detection It is now the most popular research topic , Now there are more and more frameworks , But the novelty of technology has reached the bottleneck , At present, it is becoming popular Transformer Mechanism , And it can also be greatly improved in the field of target detection , It is also a hot spot in current research .

The upcoming Asian Games , More advanced technology will be invested , such as 3D imaging 、 Attitude estimation 、 object detection 、 Tracking and identification ！ Let's go to the world and see a different China , Different technology , Different Asian Games . Today, let's talk about the embodiment of target detection in the Asian Games ！

At present, the most popular is Yolo Series frame , Recently we “ Institute of computer vision ” Also shared a good understanding of target detection work and practice , All are Yolo-Base frame .

• gaussian YoloV3 object detection （ The source code link is provided in the article ）
  

• Yolo Lightweight networks , Ultralight algorithm can achieve industrial detection effect in all hardware （ With source code ）
  

• No longer just Yolo, Now lightweight detection networks emerge in endlessly （ Framework analysis and deployment practice ）
  

• CVPR21 A new framework for target detection ： No more YOLO, But only one layer of features （ Dry cargo is full. , Recommended collection ）

02

The new framework improves

Today we share a simple optimized Yolov5, Temporarily named ：Pad-YoloV5, stay IPad It can detect in real time ！ be based on YoloV5 frame , Familiar students should not have to explain more .

YoloV4 stay YoloV3 Based on the research results of recent two years , as follows ：

1. Input terminal adopts mosaic Data to enhance

2. Backbone I used CSPDarknet53、Mish Activation function 、Dropblock Methods such as .（cspnet It can reduce the amount of calculation and ensure the accuracy ）

3. Mish Function is ：

   

4. Neck Have adopted the SPP、FPN+PAN Structure ,

5. The output end adopts CIOU_Loss、DIOU_nms operation

YoloV5 Major changes , as follows ：

1. Input end ：Mosaic Data to enhance 、 Adaptive anchor frame calculation

2. Backbone：Focus structure ,CSP structure

3. Neck：FPN+PAN structure

4. Prediction：GIOU_Loss

This time, the main optimization is , yes YoloV5 When data is enhanced , Use random scaling 、 Random cutting 、 Random layout of the way to splice , The detection effect for small targets is still very friendly . It is found through experiments that , This random splicing and regular splicing , The final result is a little different .

First, I modify the data enhancement strategy , Start to make statistics on the overall data set （ That is, data preprocessing and analysis ）, I'm roughly divided into three areas . Randomly splice the largest and smallest , The final result is really better than the overall random effect ！

secondly , Slightly modified the adaptive image scaling strategy ,Yolov5 In the code datasets.py Of letterbox Function has been modified , Add the least black edge to the original image adaptively . I am the picture after adaptive scaling , I'll fill in the lower right corner , In fact, most of the data have not changed , Just change it casually , Because online is all about Yolo Add new strategies in recent years , Indeed, there was a certain increase in the effect of the final examination .

The last modification , It's hard work Transformer The mechanism is added YoloV5 In the basic framework of , Training has really accelerated , But for the results of training with notebooks , Still not obvious enough . This is also the first time recently to share some careful thoughts about the practical process , The specific details we “ Institute of computer vision ” Later, I will share with you through a detailed work article ！

 THE END 

Please contact the official account for authorization.

The learning group of computer vision research institute is waiting for you to join ！

Institute of computer vision Mainly involves Deep learning field , It's mainly about Face detection 、 Face recognition , Multi target detection 、 Target tracking 、 Image segmentation, etc Research direction . research institute Next, we will continue to share the latest papers, algorithms and new frameworks , The difference of our reform this time is , We need to focus on ” Research “. After that, we will share the practice process for the corresponding fields , Let's really understand Get rid of the theory The real scene of , Develop the habit of hands-on programming and brain thinking ！

Sweep code Focus on

Institute of computer vision

official account ID｜ComputerVisionGzq

Study Group ｜ Scan the code to get the join mode on the homepage

  Previous recommendation  

• Label,Verify,Correct： A simple Few Shot Target detection method
  

• SPARSE DETR： Efficient end-to-end target detection with learnable sparsity （ Source code download ）
  

• Adaptive feature fusion is used for Single-Shot object detection （ With source code download ）
  

• object detection ：SmartDet、Miti-DETR and Few-Shot Object Detection
  

• RestoreDet： Target detection in low resolution image
  

• Yolo-Z： The improved YOLOv5 For small target detection （ Attached is the original paper download ）
  

• Zero sample target detection ： Robust region feature synthesizer is used for target detection （ Attached thesis download ）
  

• Target detection innovation ： A region based semi supervised method , Some labels are enough （ Attached is the original paper download ）
  

• Use advanced technology to protect our country ： Deep learning for small target detection （ Suitable for beginners ）
  

• Multigrid redundant bounding box annotation for accurate target detection
  

• Huawei Noah Ark laboratory products ： Experimental study of additive neural network in target detection
  

• Multi scale feature fusion ： Learn better semantic information for detection （ Attached thesis download ）