rlj 2019.07.13
Motivation Instance segmentation Proposal based Proposal free 慢 回归算法,每个pixel回归到center的距离
基于回归的方法 缺点: Inference的时候需要先得到中心点(聚类算法),再根据距离划 分instance
Framework
Learnable margin e: 像素 C:中心点 缺点: Margin fixed
Learnable margin Since for each instance k the gaussian outputs a foreground background probability map, this can be optimized by using a binary classification loss with the binary foreground background map of each instance as ground-truth. As opposed to using the standard cross-entropy loss function, we opt for using the Lovasz-hinge loss instead.
Seed map
Post processing Seed map取最大值,得到中心点 在同一个位置,取σ,每个像素预测到中心点距离 根据公式 ,>0.5则属于该中心点所属 instance 从seed map上把整个instance mask掉,取下一个最大值
Other details
Experiments
Experiments