Fig. 1

Mask-rcnn structure consists of two stages. The object of interest in the input image is artificially wrapped into boxes, binary-classified and fine-tuned. These boxes are then fed into the second stage of the network to be further fine-tuned to better fit the area where the object is located and multi-classified. Pixels inside the best box are then binary-classified to generate the mask. In this image, RPN stands for regional proposal network, FPN stands for feature pyramid network, RoI for region of interest and ALIGN is the RoI-Align mechanism. The head section is where 3 separate networks (two FCs, i.e. fully connected neural network and one CNN) generate the output. The rectangular boxes connected to the RPN box and Heads box indicate the type of loss functions. C and P represent the CNN layers used to construct the bottom-up and top-down architecture of the FPN respectively [26]