While analyzing an image, our interest will only be drawn to certain instances in the image. So, it was compelled to segment these instances from the remainder of the image. This process of separating the required information from the rest is widely known as segmenting instances.  During this process, the input image is first taken, then the bounding box will be localized with the objects and at last, a pixel-wise mask will be predicted for each of the class. For each of the objects, pixel-level accuracy is calculated. There are several algorithms for segmenting instances. One of the recent algorithms is the Mask RCNN algorithm proposed by He at al. (https://arxiv.org/pdf/1703.06870.pdf). The following figure portrays the architecture of Mask R-CNN:

Reproduced with permission from He et al.

The architecture looks similar to the R-CNN with an addition of segmentation. It is a multi-stage network with end-to-end training. The region proposals are learned. The network is...