---
tags: 生物辨識
---
# Consistent Instance False Positive Improves Fairness in Face Recognition
> 改善人臉識別的公平性 -> 解決不同人種間準確率不同的問題

## Motivation

## Contribution
1. 提出false positive rate penalty loss來解決這個問題
2. 資料集不用標註每張照片的種族

## Problem

## Loss
- Arcface
$$
L_{arc} = -log\frac{e^{s \cdot G(cos(\theta_{y_i}))}}{e^{s \cdot G(cos(\theta_{y_i}))} + \Sigma^n_{j=1,j\neq y_i}e^{s\ cos \theta_{y_i}} }
$$
- Instance FPR
$$
\gamma^{+}_{i}=\frac{\sum_{j=1, j\neq y_i}^{n}I(cos \theta_j>T_u)}{n-1}
$$
- weighted FPR
$$
\bar{\gamma}^{+}_{i}=\frac{\sum_{j=1, j\neq y_i}^{n}I(cos \theta_j>T_u)cos\theta_j}{n-1}
$$
- Arcface + false positive rate penalty loss
$$
L = -log\frac{e^{s \cdot G(cos(\theta_{y_i}))}}{e^{s \cdot G(cos(\theta_{y_i}))} + \Sigma^n_{j=1,j\neq y_i}e^{s\ (cos \theta_{y_i}+\alpha\frac{\bar{\gamma}^{+}_{i}}{\gamma^{+}_{u}})} }
$$

## Algorithm

## Result