yolov5,yolov4,yolov3 mess

Structural comparison

Don't ask where such a beautiful image came from, ask就是Goodness and falsehoods (toulai)
Let’s briefly talk about the changes I think are more obvious in the picture

yolov4

backbone:
①All use the Mish activation function instead of yolov3's Leakyrelu
②Start using concat in the backbone
neck:
①The complexity is soaring
②The spp module is introduced

yolov5

backbone:
①Introduced Focus module
②If you follow The content in Jiang Xiaobai’s article The activation function is changed back to Leakyrelu (Is it repeated horizontally every time? )
But some bloggers say it is silu
I will study it here when I have time

③CBL and CSP1_X structures are repeated
④CSP1_X adds BN+ activation function after concat
neck:
①spp also adds CBL structure before and after

The overall trend is the continuous subdivision and reintegration process of image information

Analyze why the activation function has to jump repeatedly

This part mainly refers to learningThis blogger's content.

(1) LeakyReLU function

The Dead ReLU problem occurred after Relu was introduced (when the input is negative, ReLU completely fails)
Attach a Relu image first!

So give a small component to the negative part to solve the Dead ReLU problem

Swish function

Suddenly turned to Zhihu, a blogger wrote a very complete Look at this!!!
I don't want to write anymore hhhh!!!!

Forget it and write a summary...

focus module

According to the horizontal and vertical coordinate directions, double interval sampling is performed to retain the basic characteristics of the image. Focus can play the following roles: reduce the number of layers, reduce the amount of parameters, reduce the amount of calculation, reduce the memory usage of cuda, and affect the mAPIn a small case, it improves the inference speed and gradient back-propagation speed.Among them, reducing the number of layers, reducing the amount of parameters, and reducing the amount of calculation refers to comparing with YOLOV3. The author believes that such a Focus layer can equal the three convolutional layers of YOLOV3