C++实现和解读Face Alignment at 3000fps via Local Binary Feature

C++实现代码的error和论文一致(最近又新添加了一个example,重写了一些代码),详见我的github链接以下是论文解读:效果见最后 1.Framework 整个流程基于Cascade Pose Regression(CVPR 2010),分为T个stage,在训练时步骤如下(testing也类似) 每个stage先抽取local binary features, 然后根据真实的$\varDelta {\hat{S}}_i$ 用linear regression训练一个regressor, 最后用训练出来的regressor得到$\varDelta S_i$(是$\varDelta {\hat{S}}_i$的近似)去更新前一个stage的shape,得到更加精确的shape 1.1 Training Phase: Input: Image set {$I$}(N samples), ground truth shapes {$\hat{S}$}, initial shapes set {$S^0$}For t=1:T do        $ features_i = \phi^t (I_i,S_i^{t-1}) $        $\varDelta \hat{S_i}=\hat{S_i} - S^{i-1} $        $E=\sum {\lVert \hat{S_i} - R^t(features_i)\rVert}^2$        $\varDelta S_i=R^t(features_i)$        $S^t_i=S^{t-1}_i+\varDelta S_i$End For     阅读全文
Binbin Xu's avatar
Binbin Xu 6月 07, 2015