Face Alignment by Coarse-to-Fine Shape Searching

CVPR 2015
http://mmlab.ie.cuhk.edu.hk/projects/CFSS.html

本文主要解决的是人脸配准问题,以前的方法 cascaded regression 主要问题在于人脸形状初始化问题。因为初始化可能不好,导致人脸形状优化陷入局部解。本文使用 coarse-to-fine framework 来解决初始化问题,同时对较大的人脸姿态也能很好的解决。

整个框架如下图所示:
Face Alignment by Coarse-to-Fine Shape Searching_第1张图片
图(b) 比较形象的说明了搜索过程。

Face Alignment by Coarse-to-Fine Shape Searching_第2张图片

3.4. Shape searching with hybrid features
In our 3-stage implementation, we use the BRIEF feature in the first two stages, and SIFT in the last stage.

our MATLAB implementation achieves 40 ms per-frame on a single core i5-4590
CPU, compared to 28 ms for cascaded regression without random voting, and 124 ms with 5 random votes

Face Alignment by Coarse-to-Fine Shape Searching_第3张图片

你可能感兴趣的:(人脸识别,ZJ)