Coord-FCN for same-class objects segmentation

The segmentation of same-class objects is a challenge in instance segmentation. In this paper, we propose a simple and universal method for same-class objects segmentation. Our method effectively segments the same-class objects in the image. This method called Coord-FCN extends FCN by adding the coordinate convolution(CoordConv) layer. Coord-FCN training is simple and only adds a small overhead compared to FCN. In addition, CoordConv is easy to generalize to other tasks. The common point of these tasks is that the objects have similar characteristics but are in different positions.