---
tags: Human Face
---

# Retinaface (2019)
RetinaFace: Single-stage Dense Face Localisation in the Wild

## Contribution
1. One stage anchor-based face detector
2. multi-task (face bbox, 5-point landmark, mesh decoder)

## Network architecture
- Backbone: ResNet152 or Mobilenet-0.25
- Neck: FPN
- week 10 有講過
- head: Context Module
- ssh的Context Module
- 使用dcn(deformable convolution network)代替conv
- Anchor setting (Input Image 640x640x3, scale step $2^{\frac{1}{3}}$, based pixel 16, aspect ratio 1:1)

| Feature Pyamid | Stride | Anchor |
|:-------------------------------:|:------:|:-------------------:|
| $P_{2} (160\times160\times256)$ | 4 | 16, 20.16, 25.40 |
| $P_{3} (80\times80\times256)$ | 8 | 32, 40.32, 50.80 |
| $P_{4} (40\times40\times256)$ | 16 | 64, 80.63, 101.59 |
| $P_{5} (20\times20\times256)$ | 32 | 128, 161.26, 203.19 |
| $P_{6} (10\times10\times256)$ | 64 | 256, 322.54, 406.37 |

- output & loss:
- face classification (softmax loss for binary classes)
- face bbox (smooth-l1-loss)
- 5-point face landmark (smooth-l1-loss)
- dense regression (mesh decoder and Differentiable Renderer)
- Mesh Decoder:
- 目的：將向量decode至3D的人臉
- 將$P_{ST}\in\mathbb{R}^{128}$利用4層的GCN decode至每個人臉像素的3D位置($D_{P_{ST}}\in\mathbb{R}^{n\times6}$, 利用n個6維的vector畫制3D的人臉, 6個value分別代表x,y,z,r,g,b)
- Differentiable Renderer:
- 目的：將向量3D的人臉投影成2D以計算算loss
- 將3D的人臉($D_{P_{ST}}$)投影至2D的人臉與GT去計算loss,其中此Renderer需要用到$P_{ill}\in\mathbb{R}^{7}$(光照參數)和$P_{cam}\in\mathbb{R}^{9}$(相機參數)
- [github code](https://github.com/google/tf_mesh_renderer)
- loss:
$$
L_{pixel}=\frac{1}{W*H}\sum_{i}^{W}\sum_{j}^{H}||R(D_{P_{ST}},P_{ill},P_{cam})_{i,j}-I(i,j)||
$$
- Total Loss
$$
L_{total}=L_{cls}+0.25L_{box}+0.1L_{pts}+0.01L_{pixel}
$$

## Training Strategy
- OHEM
- Data Augmentation
- random crop
- flip
- positive and negative anchor setting
- IOU(gt, anchor) > 0.5 -> positive anchor
- IOU(gt, anchor) < 0.3 -> negative anchor
- others -> ignores

## Experiment
- Ablation experiments of the proposed methods
- Influence of face detection and alignment on deep face recognition