Target detection notes SSD

SSD- classical one stage

1.SSD structure

​

​ The figure above shows SSD Network structure diagram , The front part is VGG-16 Part of the network ( We can see by comparing the network diagram above ), Output convolution 4_3 As the first level 38*38 Size of feature layer output , Continue convolution Processing separately obtain 19*19、10*10、5*5、3*3、1*1 Size of feature layer output , There are six outputs of different sizes , Used to predict objects of different sizes , Because the large feature map has more details , Large objects use small feature maps To make predictions , Small objects use large feature maps to predict , Here are some different sizes Generation size of feature map prediction box The choice of .

​ You can see what you finally get The number of prediction boxes is 8732 individual , Every Through one 3*3*(Classes+4) The convolution kernel of magnitude obtain Score of predicted categories , as well as Regression parameters of prediction box position .

2. Selection of positive and negative samples

Positive sample ：1. Use the real goal box to matching IOU The biggest prediction box ;2. Prediction box and real box IOU Greater than 0.5.

Negative sample ： For the remaining negative samples Calculate the maximum confidence loss (highest confidence loss), The greater the loss value, the closer it is to the positive sample , selection among The one with the greatest loss value some , Make the ratio of negative samples to positive samples 3:1.

3. Loss function

​ The loss is divided into category loss and positioning loss , Here is SSD Category loss , Positioning loss and Faster R-CNN The calculation method of positioning loss is the same .