人脸检测“Real-Time Rotation-Invariant Face Detection with Progressive Calibration Networks”

源码:https://github.com/Jack-CV/PCN-FaceDetection

提出了PCN网络检测任意角度的人脸,网络分为三级,首先将任意角度人脸从[-180°,180°]转到[-90°,90°],第二级将人脸转到[-45°,45°]范围,第三步将人脸转为正脸,示意图为:

人脸检测“Real-Time Rotation-Invariant Face Detection with Progressive Calibration Networks”

PCN网络进行具有三个任务:人脸/非人脸分类,bbox回归,以及角度标定,损失函数为:

人脸检测“Real-Time Rotation-Invariant Face Detection with Progressive Calibration Networks”

三步PCN网络的结构为:

人脸检测“Real-Time Rotation-Invariant Face Detection with Progressive Calibration Networks”

实验效果

在CPU上具有较快的速度

人脸检测“Real-Time Rotation-Invariant Face Detection with Progressive Calibration Networks”