人脸对齐--Face Alignment at 3000 FPS via Regressing Local Binary Features

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://cloud.tencent.com/developer/article/1437706

Face Alignment at 3000 FPS via Regressing Local Binary Features

CVPR2014

https://github.com/yulequan/face-alignment-in-3000fps

https://github.com/luoyetx/face-alignment-at-3000fps

https://github.com/freesouls/face-alignment-at-3000fps

本文针对人脸对齐问题，提出基于 LBP 特征 的二级回归方法，先局部后整体的思路。

基于 shape regression 的人脸对齐 predicts facial shape S in a cascaded manner，每次的迭代量如下式所示：

Φ 是特征提取器， W 表示回归函数

3 Regressing Local Binary Features

这里我们对每个特征点训练一个回归器来提取一个 LBP 特征，

3.1. Learning local binary features Φ

这里使用 regression random forest 学习 each local mapping function

3.2. Learning global linear regression W

学习整体的回归函数 W

3.3. Locality principle

这里应用了 two important regularization methods in feature learning，as guided by a locality principle:

1) we learn a forest for each landmark independently;

2) we only consider the pixel features in the local region of a landmark

下面是解释为什么做出上面的选择

Why the local region?

Intuitively, the optimal radius r should dependon the distributionof ∆s. If ∆s of all trainingsamples are scattered widely, we should use a large r; otherwise we use a small one

As expected, the radius gradually shrinks from early stage to later stage, because the variation of regressed face shapes decreases during the cascade

Why a single landmark regression?

先局部后整体具有一些优势，文中指出了三点：

1） 局部 feature pool 噪声要少点

2）独立的局部更有利于 global learning

3） the local learning is adaptive in different stages，Local learning is actually more appropriate in the late stage

4 Experiments