Table 1.

Detailed introduction on the state-of-the-art instance segmentation methods, including FCIS [35] (based on Instance-FCN [36]), Mask R-CNN, PAN and MS R-CNN.

Method	Year	Introduction	Shortcoming
FCIS	2017	The first end-to-end fully convolutional network instance segmentation framework. Added outside position-sensitive score maps (which implement classification and segmentation together) and RPN on Instance-FCN. Proposed instance-sensitive score maps to generate instance segmentation results.	Poor discrimination ability for overlapping objects.
Mask R-CNN	2017	The most influential method in instance segmentation. Added additional head for segmentation based on Faster R-CNN (extra segmentation head and original detection head do not share parameters). Changed the previous region of interest pooling (ROIPooling) to ROIAlign (region of interest align) using bilinear interpolation.	Weak ability to predict instance details.
PAN	2018	Proposed an additional feature pyramid network on Mask R-CNN. Improved the previous pooling strategy using adaptive feature pooling. Added a fully connected branch to the mask head, which greatly improves the prediction result.	High time cost.
MS R-CNN	2019	Added a scoring path prediction mask to the segmentation branch of Mask R-CNN. Added the gap between the prediction mask and the ground truth to the loss function, and obtained higher prediction accuracy.	Low accuracy for large instance.