![image-20210615101216497](https://i-blog.csdnimg.cn/blog_migrate/b508b98d5943275cc9090aae36b3a83c.png)
1. Contribution
The main contributions of this work are three-fold:
- We introduce a new perspective to design a single-shot instance segmentation framework, PolarMask, which predicts instance masks and rotated objects in the polar coordinate in an effective and efficient manner.
- With the polar representation, we propose the polar IoU loss and the soft polar centerness for instance center classification and dense coordinate regression.
- Rich experiments show that state-of-the-art performances of object instance segmentation and rotated object detection can be achieved with low computational overhead in multiple challenging benchmarks.
2. Method
The heads in PolarMask++ contain three branches, including a classification branch, a polar centerness branch, and a mask regression branch, which predict the class label, the polar centerness score and the length of each polar ray of each pixel respectively, where k and n indicate the number of categories and the number of rays.
![image-20210615101717519](https://i-blog.csdnimg.cn/blog_migrate/1a015c681dd401273a853f97d4988698.png)
2.1 Overview of Mask Segmentation in Polar Coordinate
-
Polar Representation: n = 36, θ = 10 ° \theta=10° θ=10°
-
Mass Center
- We find that the mass center is more advantageous than the box center because the mass center has a larger probability of falling inside an instance compared to its box center. Although
-
Distance Regression of Ray in Training
- {