Analysis of YOLOv3 algorithm for target detection

Fundamental

feature network

Input and output
Input $416\ast 416\ast 3$ -sized pictures (not unique, but the picture size must be a multiple of 32), output feature maps of 3 scales, respectively$13\ast 13\ast 255$，$26\ast 26\ast 255$，$52\ast 52\ast 255$ , that is divided into$13\ast 13$,$26\ast 26$,$52\ast 52$ grid cells.
Each grid cell generates 3 anchors. Each anchor corresponds to a prediction box. Each prediction box has$5+80$个参数， { （ x , y , w , h , c ） , 80   k i n d s   o f   c l a s s } \{（x,y,w,h,c）,80 \space kinds \space of \space class\} { （x,y,w,h,c）,80 kinds of class}

Output analysis
(this picture is taken from Zhihu blogger )

$13\ast 13\ast 255$，$26\ast 26\ast 255$，$52\ast 52\ast 255$ predict large, medium and small objects respectively.
$13\ast 13\ast 255$ is the feature obtained by downsampling 32 times;
$26\ast 26\ast 255$ is 16 times downsampling and$13\ast 13$ Features obtained by combining one-time upsampling;
$52\ast 52\ast 255$ is downsampled 8 times and$26\ast 26$ features obtained by combining one-time upsampling;

Positive and negative samples: Positive
samples are anchors whose IOU between the anchor and the real box is greater than the specified threshold, and the maximum IOU;
negative samples are anchors whose IOU between the anchor and the real box is less than the specified threshold.

The loss function
consists of coordinate loss, confidence loss and category loss for positive samples, and confidence loss for negative samples.
$$\begin{gathered} l_{coord}\sum _{i=0}^{S^2}\sum _{J=0}^B{1}_{i,j}^{obj}[(b_x-\widehat{b_x}{)}^2+(b_y-\widehat{b_y}{)}^2+(b_w-\widehat{b_w}{)}^2+(b_h-\widehat{b_h}{)}^2] \\ +\sum _{i=0}^{S^2}\sum _{J=0}^B{1}_{i,j}^{obj}[-log(p_c)+\sum _{i=1}^{n}BCE(c_i,\widehat{c_i})] \\ +l_{noo}\sum _{i=0}^{S^2}\sum _{J=0}^B{1}_{i,j}^{noo}[-log(1-p_c)] \end{gathered}$$
$S^2$ is the total number of grid cells,$B$ is the number of anchors in each grid cell.
The first line calculates the coordinate loss of the positive sample and the coordinate loss of the real frame; the
second line calculates the confidence and category loss of the positive sample,$1_{i,j}^{obj}$Indicates whether it is a positive sample; $-log(p_c)$ , if$p_c$The closer it is to 1, then $-log\left(p_c\right)$ is closer to 0; in category loss, for each of the 80 detected categories, a binary entropy loss operation is performed.
The third line is the confidence loss of negative samples,$-log(1-p_c)$ in$p_c$The closer it is to 0, the smaller the formula is, and it is closer to 0

performance