REGIONAL ATTENTION BASED DEEP FEATURE FOR IMAGE RETRIEVAL

1. idea ¹

对于region的注意力机制。修改R-MAC方法，对每个region计算一个权重，使得对召回贡献最大的region拥有大的权重。所有region用于计算权重的fc层是共享的。

2. ablation
   
   (1) finding an optimal scale of our method
   (2) how much each component of our method improves the performance
   (3) additionally experiment our regional attention network on off-the-shelf RPN instead of R-MAC