back bone

Used spp structure , To some extent, it solves the multi-scale problem .

The first circle is back bone, The second circle is fpn Pyramid structure , The third circle is the inverted pyramid structure .

v4 Overall structure

Optimization strategy

This is a v3 Of , There is a question , Just say when grid cell It's not easy to deal with if the center point of falls on the intersection .

therefore v4 Added one scale factor ,scale Can take 2.

This is v3 The matching of , First take the left 3 individual anchor and gt do iou Calculation , And then look gt Where does the center of the fall grid cell Inside , Just use this. grid cell Of 3 individual anchor in iou Greater than 0.3 As a positive sample , Use it to predict .

This is v4 了 , You can see ,gt The center point of the box falls on a grid cell Inside , Plus the left and top grid cell, Because after the change in the lower right corner , become -0.5 To 1.5 了 , Then this one above grid cell Of the top left corner of x The coordinate distance from the center point is less than 0.5,y The coordinates are also less than 1.5 Of , So this one above grid cell Can be selected .

More information is shown in the figure above .

scale .

v5

https://blog.csdn.net/qq_37541097/article/details/123594351



This picture is the attention mechanism .

Instead of sppf,

Data to enhance .mosaic, Put together four pictures into one picture , Can increase diversity .

Take people down , With giraffes .

Affine transformation , Translate, rotate, zoom, and so on ,

Set the transparency of two pictures and mix them into one picture .