Content-Aware Unsupervised Deep Homography Estimation

network structure:

1. 网络提取出Ia和Ib的feature map，并预测出两张feature map的mask(和feature map同尺寸，mask的值代表该pixel为内点的概率),将feature map 与 mask相乘得到weighted feature map。
2. 将两张weighted feature map拼接起来得到H × W × 2C的concat_feature_map送入一个resnet预测得到4个 2D offset vectors，即可求出对应的homography

training：

1. 将Ia用预测得到的Hab warp（用STN）得到Ia’,将Ia’送入网络得到Ia’的feature map a’和Mask Ma’。
2. loss设计：
   loss 1:minimize Ia’的feature map 和Ib的feature map经过weighted后的L1距离
   
   loss2:
   
   This strategy avoids the trivial
   all-zero solutions, and enables the network to learn a discriminative feature map.

loss3:

total loss: