Pyramidal feature hierarchy In SSD, the feature maps, conv4-3,conv-7(FC7),conv6-2,conv7-2,conv8_2,conv9_2, etc. is used to predict bounding boxes and their class labels. Prior Box Small objects Large objects layer?{?? ??name:?"conv4_3_norm_mbox_priorbox"?? ??type:?"PriorBox"?? ??bottom:?"conv4_3_norm"?? ??bottom:?"data"?? ??top:?"conv4_3_norm_mbox_priorbox"?? ??prior_box_param?{?? ????min_size:?30.0?? ????max_size:?60.0?? ????aspect_ratio:?2?? ????flip:?true?? ????clip:?false?? ????variance:?0.1?? ????variance:?0.1?? ????variance:?0.2?? ????variance:?0.2?? ????step:?8?? ????offset:?0.5?? ??}?? }?? Permute,Flatten And Concat Layers R-FCN Object Detection via?Region-based Fully Convolutional Networks RFCN framework RPN consistent with Faster RCNN Base CNN network position sensitive prediction Recap of Faster RCNN RFCN Position-sensitive score maps Thank you! Faster RCNN: /zy1034092330/article/detailsYOLO:/zy1034092330/article/detailsSSD:/zy1034092330/article/detailsRFCN:/baiduarticle/details综合:/hx921123/article/detailslocationNum=3&fps=1 * 1、Conv layers。作为一种CNN网络目标检测方法,Faster RCNN首先使用一组基础的conv+relu+pooling层提取image的feature maps。该feature maps被共享用于后续RPN层和全连接层; 2、Region Proposal Networks。RPN网络用于生成region proposals。该层通过softmax判断anchors属于foreground或者background,再利用bounding box regression修正anchors获得精确的proposals; 3、Roi Pooling。该层收集输入的feature maps和proposals,综合这些信息后提取proposal?feature maps,送入后续全连接层判定目标类别; 4、regression and classification。利用proposal?feature maps计算proposal的类别,同时再次bounding box regression获得检测框最终的精确位置。 * 经典的检测方法生成检测框都非常耗时,如OpenCV adaboost使用滑动窗口+图像金字塔生成检测框;或如RCNN使用SS(Selective Search)方法生成检测框。而Faster RCNN则抛弃了传统的滑动窗口和SS方法,直接使用RPN生成检测框,这也是Faster RCNN的巨大优势,能极大提升检测框的生成速度。 可以看到RPN网络实际分为2条线,上面一条通过softmax分类anchors获得foreground和background(检测目标是foreground),下面一条用于计算对于anchors的bounding box regression偏移量,以获得精确的proposal。而最后的Proposal层则负责综合foreground anchors和bounding box regression偏移量获取proposals,同时剔除太小和超出边界的proposals。其实整个网络到了Pro