[target detection] Introduction to yolor theory + practical test visdrone data set

Preface

YOLOR yes 2021 An algorithm proposed in , One of its works Chien-Yao Wang( Taiwan ) At the same time, it has just come out recently YOLOv7 The first author of .

Address of thesis ：https://arxiv.org/abs/2105.04206
Address of the thesis project ：https:// github.com/WongKinYiu/yolor

A brief introduction to the theory

YOLOR Your paper is not easy to chew , Because it writes a lot of mathematical derivation concepts , I don't do in-depth derivation , Simply retell his thoughts .

YOLOR The core idea of is also based on the physiological structure of human brain . Human beings can actively learn through normal senses and experience , The knowledge learned in this part is called Explicit knowledge , meanwhile , Human beings can also learn through subconsciousness , This part of the knowledge learned is called Tacit knowledge .

In the traditional neural network model , It often extracts neuron features and discards the learning and application of implicit knowledge , The author defines directly observable knowledge as explicit knowledge , Knowledge hidden in neural networks and unobservable is defined as implicit knowledge .
therefore , The author has made the following research contributions to this point ：

1. A unified network that can complete various tasks is proposed , It learns general representation by integrating implicit knowledge and explicit knowledge , People can accomplish various tasks through this general expression . The proposed network at a very small additional cost ( Less than 1 The parameter quantity and calculation quantity of ten thousand ) The performance of the model is effectively improved .
2. Align the kernel space 、 Prediction refinement and multi task learning are introduced into the implicit knowledge learning process , And verify their effectiveness .
3. Using vectors 、 The method of modeling tacit knowledge by means of neural network and matrix decomposition , And its effectiveness is verified .
4. It is proved that the proposed implicit representation can accurately correspond to specific physical features , And we also present it in a visual way .

The following specific content is too complex and boring , For in-depth understanding, you can see the original paper or blog [1] Summary mind map ：

The following is the experimental results of the author using different models shown in the paper ：

Practice testing

Is it good , It depends on the actual use .
Seeing the source code is basically based on YOLOv5 Changed , So read my previous blog 【 object detection 】YOLOv5 Run through VisDrone Data sets Our readers should be familiar with Runtong YOLOR I'm familiar with it .

I will use VisDrone Data set and yolor_csp_x_star This model runs 100 individual epoch, Retest , Here are the test results , Same as before 【 object detection 】TPH-YOLOv5： be based on transformer Improvement yolov5 UAV target detection Make an experimental model for comparison , give the result as follows ：

Algorithm	[email protected]	[email protected]:.95s
yolov5-5.0	34.9%	20.6%
yolov5-6.1	33.1%	18.7%
tph-yolov5	37.4%	21.7%
yolov6	32.5%	17.4%
yolor	37.7%	22.8%

notes ： It's just 100 individual epoch What you get best.pt Test results , Not achieving optimal performance , Usually the test results shown in the paper are 300 individual epoch above .

Just look mAP Numerical words ,YOLOR The value of is iterating 100 Student: the wheel case , The highest value , explain YOLOR Indeed, it has improved the effect .

Next, I import the small target human detection video I tested before , You can see YOLOR The detection effect of is not as ideal as expected , Explain the situation of small targets ,YOLOR Not very applicable ( This is a conclusion reached without enough experiments , For reference only ).

YOLOR_VisDrone Data set detection effect

Code backup

run YOLOR It is obvious that the author's eyes are not limited to AP, occasionally ,AR In some specific tasks , Than AP It is more important . The author optimized this point , After training , Unlike the original only save best.pt and last.pt, Instead, save multiple models .

It may be due to some mistakes of the author , The original code has some small at run time bug. I made a simple correction , The code is backed up as follows ( contain yolor_csp_x_star.pt Weight file and some results of my running )：
https://pan.baidu.com/s/1lXiu82q0sQPTCU0C4UmcAQ?pwd=8888

References

[1]https://blog.csdn.net/weixin_42917352/article/details/120050525