[deep learning] [original] how to detect targets and draw map and other parameter maps without yolov5 weights or models

When you see this blog , Your first reaction must be bullshit , Target detection without models ？ Yes, you did , I can simulate the target detection results without weight , Including pictures , Video and parametric curve generation . First, let's take a look at my generated parameter graph

It can be seen that there is almost no difference between the actual training parameters , And the simulation parameters can be controlled by code , Like trying to put map Set to 0.8,loss from 0.1 Start descending ,epoches Change to 500 wait , Can be completed . In addition, I have realized the batch annotation drawing of pictures to pictures and yolov5 The real test results are identical , You can even set the confidence to a fixed value ！ In fact, as long as you know yolov5 The drawing principle can easily achieve a high degree of simulation of test results . And recently I have implemented analog video detection , The fundamental reason for the application background of this technology is ：

Many students have just come into contact with target detection , Then I don't know how to do it , Actually, the environment is built , model training , test , A bunch of operations like preparing datasets take a long time , The key is the hardware GPU To complete . Some students in order to complete homework or hand in homework , It is necessary to obtain the test results in time , There are also video results and parameter diagrams , All these require a high-precision model to complete such a task , If the generation can be simulated, the hardware preparation can be omitted , Dataset annotation , Model training and a series of operations , And finish your homework quickly , Although this can solve the urgent problem , But I still recommend you to train your model honestly , Only in this way can we gain real knowledge , But there is a kind of speculation in using my method , Of course, you can also exercise your coding ability . Especially simulation capability , But in the final analysis, it is the king to follow the normal process . Okay , If there is such a request for private letter to me . This is the end of this article .

Last sentence , Actually yolov5 It's still a good framework , Many practical problems have been solved , The recent discovery yolov6 It's already out , You can learn about the new target detection algorithm , It is said that transformer It is the trend of target detection in the future , The reason is that the structure is simple , A lot of algorithm principles are omitted , Let Algorithm Engineers free themselves from complex algorithms .