polarMask

PolarMask: Single Shot Instance Segmentation with Polar Representation

动机
- instance segmentation
- anchor-free
- single-shot
- modified on FCOS
论点
- two-stage methods
  - FCIS, Mask R-CNN
  - bounding box detection then semantic segmentation within each box
- single-shot method
  - formulate the task as instance center classification and dense distance regression in a polar coordinate
  - FCOS can be regarded as a special case that the contours has only 4 directions
- this paper
  - two parallel task：
    - instance center classification
    - dense distance regression
  - Polar IoU Loss can largely ease the optimization and considerably improve the accuary
  - Polar Centerness improves the original idea of “Centreness” in FCOS, leading to further performance boost
方法
- architecture
  - back & fpn are the same as FCOS
  - model the instance mask as one center and n rays
    - conclude that mass-center is more advantageous than box center
    - the angle interval is pre-fixed, thus only the length of the rays is to be regressed
    - positive samples：falls into 1.5xstrides of the area around the gt mass-center，that is 9-16 pixels around gt grid
    - distance regression
      - 如果一条射线上存在多个交点，取最长的
      - 如果一条射线上没有交点，取最小值$\epsilon=10^{-6}$
- potential issuse of the mask regression branch
  - dense regression task with such as 36 rays, may cause imbalance between regression loss and classification loss
  - n rays are relevant and should be trained as a whole rather than a set of independent values—->iou loss
- inference
  - multiply center-ness with classification to obtain final confidence scores, conf thresh=0.05
  - take top-1k predictions per fpn level
  - use the smallest bounding boxes to run NMS, nms thresh=0.5
- polar centerness
  - to suppress low quality detected centers
  - $polar\ centerness=\sqrt{\frac{min(\{d_1,d_2, …, d_n\})}{max(\{d_1,d_2, …, d_n\})}}$
  - $d_{min}$和$d_{max}$越接近，说明中心点质量越好
  - Experiments show that Polar Centerness improves accuracy especially under stricter localization metrics, such as $AP_{75}$
- polar IoU loss
  - polar IoU：$IoU=lim_{N\to\inf}\frac{\sum_{i=1}^N\frac{1}{2} d_{min}^2 \Delta \theta}{\sum_{i=1}^N\frac{1}{2} d_{max}^2 \Delta \theta}$
  - empirically observe that 去掉平方项效果更好：$polar\ IoU=\frac{\sum_{i=1}^n d_{min}}{\sum_{i=1}^n d_{max}}$
  - polar iou loss：bce of polar IoU，$-log(\frac{\sum_{i=1}^n d_{min}}{\sum_{i=1}^n d_{max}})$
  - advantage
    - differentiable, enable bp
    - regards the regression targets as a whole
    - keep balance with classification loss