Mask_R-CNN

Overview (What is Mask RCNN)

• Deep residual network which can solve instance segmentation

  • Instance segmentation referes to seperating different objects from a image or a video.

• Two steps of Mask RCNN

  1. Generates proposals regarding to regions where the object might occupy.
  2. Predicts the class, bounding box, and mask in pixel level of the object.

• Backbone structure

  • Feature Pyramid Netwrok style deep nerual network.
  • Consists of bottom-up path way, top-bottom pathway, and lateral connections.
    
    
  

Step 1

• RPN scans all FPN path and prposes regions where it may contain objects
• Anchors are set of boxes sharing the same center
• RPN uses the anchors to predict and output the coordinates and thes size of the object.

Step 2

• ROI align locates the relevent areas of fetaure map
• One branch generates masks for each object in pixel level

Architecture of Mask RCNN