问题:most approaches rarely distinguish different types ofcontextual dependencies, which may pollute the scene un- derstanding.
以往的上下文关系没有很好的区分类别,对语义造成了污染
解决:创建了Affinity Loss和Context Prior with来区分intra-class and inter- class context
模型:
Affinity Loss:
建立类别之间的关系
To explicitly regularize the network to model the relationship between categories,
L为ground truth L,下采样为与特征图同尺寸的smaller ground truth ~L ,经过one hot encoding变成^L,C为类别数目
损失函数:
P为预测的概率图,A为Ideal Affinity Map,Ls为segment loss,λa = 0.4,λu = 1 and λg = 1.