# Attentional Feature-Pair Relation Networks for Accurate Face Recognition

Aug 17, 2019
difficulty: 3
rating: 3

[paper](https://arxiv.org/abs/1908.06255v1)

New architecture block to cope with non-perfect localization.

Important: this is not self attention! Thisis Bilinear attention.

As I understood the intuition, the proposed bilinear attention aggregates infromation about pairs of spatial locations. So the tesult should be treated as co-occurence as of spatial features in the picute what makes sence.

They also complain about non-perfect localization network and say that the proposed method helps a bit.

## Architecture
The whole pipeline picture looks complicated so we better look closer at details.

### Backbone

For backbone it is OK to use any convolutional network. In the paper they use resnet-101. I see no preference in used backbone, however, no alternatives were considered in the study.

### Reshaping Feature Maps
Adter backbone we have `9x9` feature maps. Rehsape them


### Computing "Similarities"
This is not a similarity in sence that we compute inner product. THe formula for Biilinear Atteniton Map is

$$
\mathcal{A}_{ij} = p^\top \left(
\operatorname{ReLU}\left(
U^\top F_i
\right) \circ
\operatorname{ReLU}\left(
V^\top F_j
\right)
\right)
$$

There are some parameters:

$U\in \mathbb{R}^{D\times L}$, $V\in \mathbb{R}^{D\times L}$, $r\in \mathbb{R}^{L}$.

$U$, $V$ make this operation non symmetric and $\mathcal{A}_{ij}\ne\mathcal{A}_{ji}$

After all $U^\top F_i$ or $V^\top F_j$ are features on a lower dimensional space.

### Pooling and softmax

$p$ is learnable pooling of low dimensional features on dim $L$.

TODO: continue