RPN structure in Faster-RCNN
Recently, I am looking at related papers on target detection. The most recent one is faster-RCNN. Compared with fast-CNN, faster-RCNN uses the neural network to do the resolution proposal module originally obtained by regional search. In faster-RCNN , this part is called RPN (region proposal network). I was puzzled about this part at first. Today, I saw a blog written by a friend. Mao Said, and wrote down his understanding here for future reference:
The network structure of RPN is: the first part is a 3*3 convolutional layer added after the fifth convolutional network of the VGG network. First, after five layers of convolution, the size of the feature map becomes 1 of the original image. /16; that is. The original image is 1000*600, and after the five-layer convolution of VGGNet, it becomes a feature map of 60*40 size. For each point on this feature map, according to different aspect ratios, select 9-scale boxes ( anchor), then there are 60*40*9 anchors for a feature map, that is, 60*40*9 candidate regions
Collection URL:
CNN target detection (1): Faster RCNN in detail
Faster RCNN github: https://github.com/rbgirshick/py-faster-rcnn
Faster RCNN:paper: https://arxivorg/abs/1506.01497
faster-rcnn , Understanding of RPN
Spatial pyramid pooling (SPP)-net (spatial pyramid pooling) notes
The anchor, sliding windows, proposals of rpn in faster rcnn?
Target detection is broad and profound, I want to practice hard