Faster RCNN.pdfFaster RCNN 网络结构，详细解读Faster RCNN的各个

文件名称: Faster RCNN.pdf

所属分类: 深度学习

开发工具:

文件大小: 227kb

下载次数: 0

上传时间: 2019-09-14

提供者: wqwqqw******

下载 (227kb)

不能下载？报告错误

详细说明：Faster RCNN 网络结构，详细解读Faster RCNN的各个网络模块，有助于对two stage模型的理解anchor target layer output labels res labels 去除1 labels 4*1*333*50 4*16650 reshape 1024 cr。 isEntropy rpn cls lo rPn CIs score 去除1对应的, rPn CIs score 4*2*333*50 reshape 1024 RPN bbox pred 4*36*37*50 L1 loss rpn box loss rpn loss anchor target layeroutput bbox target 4*36-37*50 anchor_ target layer output outside_ weight 权重 4*36*37*50 anchor target layer output inside weight 指示 4*36*37*50 Base feature align pooled feature fc7+reshape pooled feature bbox pred bbox pred bbox pred 4*512*37*50 1024*5127*7 1024*4096 1024*84 1024*21*4 1024*4 proposal target layer output rois 4*256*5 proposal target layer output label reshape label 4*256 1024 选择 cIs score softmax 1024*21 1024*21 proposal target layer output bbox pred reshape 4*256*5 L1 oss mean cnn box loss rpn box lOSs bbox pred 1024*4 loss hape s pro 1024*21 CrossEntropy rann cls loss rpn cls loss proposal target layer output label batch 4*256 anchors 丢弃有ⅰmae anchors overlap max overlap大L大于h置1 16650*4 外部分的 4*5944*4 4*5944*20 4*5944 小于 bg_thres置0 labels初始值为-1 gt boxes lax gt idx 4*20*5 4*20 label Sample fg和b, labels Um reshpae+permute labels 4*5944 将多余的晋为1 4*5944 4*16650 4*1*333*50 overlap 根据☆的对应 bbox target Umap bbox target reshpaetpermute bbox target 5944*4 计算 Itrans 4*5944*4 4*16650*4 4*36*37*50 gt boxes 4*20*5 outside weight Umap outside_ weightreshpaetexpand outside weight 平衡 rpn cls_los和rpn_ reg los的权重 4*256*4 416650 4*36*37*50 inside weight ap inside eshpae+expand tht 指示哪些 proposal参与los计算 4*256*4 4*16650 4*36*37*50

(系统自动生成,下载前可以参看下载内容)