UNET and mask RCNN

encoder It's the lower sampling process on the left ,decoder It is the sampling process on the right . The final output 388x388x2 A partition graph of , Can be classified 2 Categories , A prospect , A background .

mask rcnn

The circle on the left is faster rcnn Structure , The circle on the right is mask Branch


Above this 6x6 Is the original picture zoom 32 Times the result , Then the thick black box is the area where the object is located after scaling , Then if you finally ask for 2x2 Result , Then put this 5x5 It's divided into 4 Share , But there is no equal here .

roi aligh, The blue frame is the area where the object is located after shrinking ,

Every 1/4 The value of the blue box is the center point, such as the orange point , Finally, the value of orange dot is obtained by sampling bilinear interpolation .