Skipped-Hierarchical Feature Pyramid Networks for Nuclei Instance Segmentation

Dealing with multiple scale of object is main problem in computer vision. Feature Pyramid Networks (FPN) has widely used in instance segmentation area to utilize multiple scales of features. Using different scale of feature maps, the method enables to capture a various sizes of objects in a scene. However, FPN still cannot propagate semantic information of deeper layer into the shallow layer which contains spatial information strongly. In this paper, we propose a novel network which consists of stage residual connection and aggregation between $\boldsymbol{C_{i}}$ and $\boldsymbol{P}_{\boldsymbol{i}-1}$ above the FPN to improve the imperfectness of original FPNs for the instance segmentation. Our proposed network is called Skipped-Hierarchical Feature Pyramid Networks (SH-FPN), integrated on Mask R-CNN. Experimental results of SH-FPN show that it has significant improvement on Data Science Bowl 2018 benchmark dataset on nuclei segmentation, compared to FPN.