场景

we propose a method to segment clothes in settings where there is no restriction on number and type of clothes, pose of the person, viewing angle, occlusion and number of people.

总之一句话：

to segment clothes in unconstrained settings

方法

• 用poselet得到姿态估计。多种人体概率，个数不限。
• 超像素分割

• 使用CRF建模
  相邻的超像素有边连接。目标函数如下:
  
  对图的假设：

  at most two vertices in the graph form a clique thus limiting the order of a potential to two
  

  • 点势能：色彩直方图RGB，色彩直方图LAB，Gabor，坐标分布直方图，对关节点的坐标直方图分布。 跟姿态box没交集的超像素，距离被设置为无穷。
  • 边势能：与yam12论文一样

CRFparsing