PolarMask: Single Shot Instance Segmentation with Polar Representation

April 2020

tl;dr: Extend FCOS for instance segmentation.

Overall impression

One main draw back of this paper is that the paper cannot handle scenarios where the center is not inside the mask, or there is no unique “center” inside the mask, such as a donut.

One question is that given polar centerness metric, the center should always be inside the contour (other than donut shaped mask). This is an alternative to the distance transformation used by ESESeg.

Overall the idea merits promotion, but the presentation of the idea is not impressive. I am still not convinced that the parameterization of the mask contour is optimal.

Key ideas

Technical details