Before we talk about fast-rcnn, let's take a look at spp net, right? ~
There are two main improvements to r-cnn by spp net :
1. Extract features only once from the original image. The input is a picture, and the position of the candidate frame is found at the corresponding position of the feature layer.
2. Combine the spatial pyramid method to realize the input of cnns to different scales. Realize multi-scale input of data. Divide different small blocks to pool the features.
Compared with r-cnn, the speed is about 100 times faster
Question: How to train the network for extracting features?
Process: marquee, CNN, SVM, regression
fast rcnn：
The problem solved: It takes time to input a lot of graphs and repeat the calculation.
Solution: Learn from the idea of spp net, 1. After the picture passes the network once, select the frame at the feature layer; 2. Propose roi pooling, divide the candidate frame into 2*2 regions, extract the features, and ensure that all the features extracted by the candidate frames are Is consistent.
Process changes: without SVM, classification and regression branches are directly connected after roi pooling
faster rcnn：
Problem solved: Selective search to find all candidate boxes is also very time-consuming.
Solution: Use region proposal network (rpn) --> to extract the edge neural network.
Specific method: Put rpn after the last convolutional layer. rpn directly trains to obtain candidate regions.
Detailed explanation of rpn specific algorithm: https://www.cnblogs.com/wangyong/p/8513563.html
Detailed explanation of non-maximum suppression algorithm: https://blog.csdn.net/zchang81/article/details/70211851
Process changes: Add RPN to select boxes in the feature layer, and finally determine about 300 boxes with higher probability to continue to do ROI POOLING, etc., NMS moves to the network and finally decides which 300 boxes are left