[ultra detailed diagram] FPN + mask RCNN

 Mask R-CNN summary . The dimension of the original image is 512×512, adopt Backbone Extracting feature maps , then
stay FPN、RPN and RoI Align after , They are sent to three functional branches .

The two branches in the lower right corner are obtained Roi features maps Yes box Classification and classification of reg

Classification Part of the use has been made of proposal feature maps, adopt full connect Layer and softmax Calculate each proposal Which category does it belong to （ As a man , vehicle , TV, etc ）, Output cls_prob Probability vector ; At the same time make use of bounding box regression Get each proposal Position offset of bbox_pred, A more accurate target detection frame for regression .

The middle branch on the right passes through a simple full convolution network , Directly predict one 0-1mask chart